jgrusewski 29d58efca2 perf: pre-allocate insert + flush scratch buffers (zero cuMemAlloc per step) ...

Eliminated 5 per-step GPU allocations:
- insert_prio_buf, insert_idx_buf, insert_ep_buf: 3x cuMemAlloc per
  insert_batch/insert_batch_bf16 call (up to 4M elements each)
- scratch_f32: reusable 1-element temp for flush_max_priority and
  apply_max_priority_scalar (was 2x a32f per call)
- update_batch_spare: pre-allocated spare for None→Some priority swap,
  replenished by flush_max_priority (was 1x a32f per epoch start)

cuMemAlloc is a synchronizing driver call that stalls the GPU pipeline.
With 8192 batch_size and multiple inserts per epoch, this was measurable
overhead on H100.

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

2026-04-06 19:38:50 +02:00

..

evaluation

fix(metrics): unify return basis — total_return from same series as Omega

2026-03-29 20:03:48 +02:00

agent.rs

feat: gradient stability + 9-agent audit bug fixes

2026-03-27 19:52:44 +01:00

attention.rs

refactor(ml-dqn): eliminate candle — pure cudarc + GpuTensor/GpuLinear

2026-03-17 23:12:18 +01:00

branching_composition_tests.rs

feat(cuda): add branching_action_select kernel + composition tests

2026-03-10 12:45:09 +01:00

branching.rs

feat(bf16): ALL tests pass — ml-core 300/300, ml-dqn 359/359, ml 890/895

2026-03-28 12:13:45 +01:00

cast_kernels.cu

feat(bf16): precompile ml-dqn + ml-ppo inline kernels, fix cubin mappings

2026-03-28 11:39:07 +01:00

checkpoint.rs

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

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

circuit_breaker.rs

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

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

count_bonus.rs

feat(dqn): H100 curiosity module, branching config, regime conditioning + clippy clean

2026-03-13 10:30:51 +01:00

curiosity.rs

feat: ExposureLevel 5→9 variants (Short75/Short25/Long25/Long75)

2026-03-23 09:21:58 +01:00

distributional_dueling.rs

fix(clippy): ml-dqn — 148 errors fixed

2026-03-19 00:49:38 +01:00

distributional.rs

fix(clippy): ml-dqn — 148 errors fixed

2026-03-19 00:49:38 +01:00

dqn.rs

cleanup: remove dead GpuTrainResult, cast helpers, stale annotations

2026-04-06 10:23:39 +02:00

dsr_gpu_tests.rs

fix(clippy): resolve all clippy warnings across entire workspace (--all-targets)

2026-03-13 10:18:35 +01:00

dueling.rs

fix(clippy): ml-dqn — 148 errors fixed

2026-03-19 00:49:38 +01:00

ensemble_network.rs

feat: gradient stability + 9-agent audit bug fixes

2026-03-27 19:52:44 +01:00

entropy_regularization.rs

fix(ml-dqn): resolve final 24 compile errors — ml-dqn builds clean

2026-03-18 08:23:30 +01:00

experience_dataset.rs

fix(clippy): achieve zero warnings across entire workspace

2026-03-10 13:18:57 +01:00

experience.rs

feat: core hyperopt families (30D→14D) + regime distribution logging per fold

2026-03-31 01:05:25 +02:00

gae.rs

fix(clippy): achieve zero warnings across entire workspace

2026-03-10 13:18:57 +01:00

gpu_replay_buffer.rs

perf: pre-allocate insert + flush scratch buffers (zero cuMemAlloc per step)

2026-04-06 19:38:50 +02:00

hindsight_replay.rs

fix(clippy): resolve all clippy warnings across entire workspace (--all-targets)

2026-03-13 10:18:35 +01:00

iql.rs

fix(clippy): ml-dqn — 148 errors fixed

2026-03-19 00:49:38 +01:00

lib.rs

fix(ml-dqn): resolve final 24 compile errors — ml-dqn builds clean

2026-03-18 08:23:30 +01:00

logging.rs

fix(clippy): add doc backticks and const fn across ML crates

2026-03-10 11:51:31 +01:00

logit_clipping.rs

refactor(ml-dqn): eliminate candle — pure cudarc + GpuTensor/GpuLinear

2026-03-17 23:12:18 +01:00

multi_asset.rs

fix(clippy): resolve all clippy warnings across entire workspace (--all-targets)

2026-03-13 10:18:35 +01:00

network.rs

feat: gradient stability + 9-agent audit bug fixes

2026-03-27 19:52:44 +01:00

noisy_exploration.rs

fix(clippy): add doc backticks and const fn across ML crates

2026-03-10 11:51:31 +01:00

noisy_kernels.cu

feat(bf16): mixed-precision kernels, f32 IS-weights, CUTLASS padding, fast_isnan

2026-03-28 21:40:09 +01:00

noisy_layers.rs

feat(bf16): ALL tests pass — ml-core 300/300, ml-dqn 359/359, ml 890/895

2026-03-28 12:13:45 +01:00

noisy_sigma_scheduler.rs

fix(clippy): add doc backticks and const fn across ML crates

2026-03-10 11:51:31 +01:00

nstep_buffer.rs

feat: DQN gems — spectral decoupling, manifold mixup, regime replay, family hyperopt

2026-03-31 00:39:23 +02:00

nstep_gpu_tests.rs

fix(clippy): resolve all clippy warnings across entire workspace (--all-targets)

2026-03-13 10:18:35 +01:00

per_kernels.cu

fix: dynamic tile assignment prevents decoupled lookback deadlock

2026-04-06 15:11:07 +02:00

performance_tests.rs

refactor(ml-dqn): eliminate candle — pure cudarc + GpuTensor/GpuLinear

2026-03-17 23:12:18 +01:00

performance_validation.rs

fix(clippy): add doc backticks and const fn across ML crates

2026-03-10 11:51:31 +01:00

prioritized_replay_staleness.rs

fix(clippy): add doc backticks and const fn across ML crates

2026-03-10 11:51:31 +01:00

prioritized_replay.rs

feat: DQN gems — spectral decoupling, manifold mixup, regime replay, family hyperopt

2026-03-31 00:39:23 +02:00

quantile_regression.rs

fix(spec-b): silent fallbacks, gradient clip defaults, CUTLASS weight padding

2026-03-28 23:40:09 +01:00

rainbow_agent.rs

fix(clippy): ml-dqn — 148 errors fixed

2026-03-19 00:49:38 +01:00

rainbow_config.rs

refactor(cuda): eliminate all CPU fallbacks — CUDA mandatory across ML stack

2026-03-16 21:01:28 +01:00

rainbow_integration.rs

refactor(cuda): eliminate all CPU fallbacks — CUDA mandatory across ML stack

2026-03-16 21:01:28 +01:00

rainbow_network.rs

fix(clippy): ml-dqn — 148 errors fixed

2026-03-19 00:49:38 +01:00

regime_conditional.rs

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

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

replay_buffer_kernels.cu

feat(generalization): #30 native f32 experience pipeline — zero bf16 truncation

2026-03-30 12:08:34 +02:00

replay_buffer_type.rs

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

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

replay_buffer.rs

fix(clippy): resolve all clippy warnings across entire workspace (--all-targets)

2026-03-13 10:18:35 +01:00

residual_kernels.cu

feat(bf16): precompile ml-dqn + ml-ppo inline kernels, fix cubin mappings

2026-03-28 11:39:07 +01:00

residual.rs

feat(bf16): precompile ml-dqn + ml-ppo inline kernels, fix cubin mappings

2026-03-28 11:39:07 +01:00

reward.rs

refactor: remove dead CPU reward code — RewardNormalizer, use_dsr toggle, hold_reward

2026-03-22 21:27:00 +01:00

rmsnorm_kernels.cu

feat(bf16): precompile ml-dqn + ml-ppo inline kernels, fix cubin mappings

2026-03-28 11:39:07 +01:00

rmsnorm.rs

feat(bf16): precompile ml-dqn + ml-ppo inline kernels, fix cubin mappings

2026-03-28 11:39:07 +01:00

softmax.rs

refactor(ml-dqn): eliminate candle — pure cudarc + GpuTensor/GpuLinear

2026-03-17 23:12:18 +01:00

target_update.rs

feat(bf16): ALL tests pass — ml-core 300/300, ml-dqn 359/359, ml 890/895

2026-03-28 12:13:45 +01:00

trade_executor.rs

fix(clippy): resolve all clippy warnings across entire workspace (--all-targets)

2026-03-13 10:18:35 +01:00