The GPU PER replay buffer had a hardcoded 4 GB MAX_BYTES limit that
rejected the auto-sizer's 10M-entry proposal on H100 (80 GB VRAM).
Now per_max_buffer_bytes() computes 20% of total VRAM (min 1 GB) and
flows through OptimalReplayConfig → DQNConfig → GpuReplayBufferConfig
so both subsystems agree on the budget.
Also fixes misleading regime detection log (indices 211/219 → 40/41)
and renames dqn_config_2025 → dqn_default_config.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Training data is already on the PVC — the rclone sync step was
wasting 2-3 minutes per job. Also fixes --output-dir → --output
to match the actual CLI flag.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three fixes for GPU-accelerated Branching DQN training:
1. **GPU experience collector**: NoisyLinear creates standalone Vars via
Var::from_tensor(), bypassing VarMap registration. The GPU collector
looks up weights by name ("value_fc.weight") from VarMap and falls
back to CPU (~5x slower) when missing. Fix: register mu vars in
VarMap at construction, keep sigma vars standalone.
2. **Optimizer device mismatch**: Using only vars().all_vars() left
NoisyLinear head params frozen. backward() produces gradients the
optimizer doesn't know about → device mismatch in clip_grad_norm.
Fix: all_trainable_vars() = VarMap (shared+mu) + sigma.
3. **Single-threaded CPU bottleneck**: Runtime::new() creates a
current-thread scheduler → 1 OS thread → all async work serialized.
Fix: multi-thread runtime (4 workers) created once in DQNTrainer::new(),
shared across preload/training/backtest phases. Eliminates 3 fallback
Runtime::new() callsites.
Also: polyak_update_var_pairs with debug_assert_eq, two-phase target
network sync (VarMap Polyak + sigma var_pairs Polyak), copy_weights_from
handles NoisyLinear heads.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
sccache forces CARGO_INCREMENTAL=0, causing all 37 workspace crates to
recompile from scratch every CI run (~20 min). Only upstream deps were
cached (635 hits); 109 workspace rlib crates were non-cacheable.
Changes:
- Add cargo-target-cpu and cargo-target-cuda PVCs (30Gi each)
- Mount persistent target dir at /cargo-target via CARGO_TARGET_DIR
- Drop RUSTC_WRAPPER=sccache and SCCACHE_DIR from both compile steps
- Drop hardcoded CARGO_BUILD_JOBS=14 (let cargo auto-detect from nproc)
- Add 25GB cleanup guard to prevent unbounded PVC growth
- Update binary copy paths to use $CARGO_TARGET_DIR/release/
First build (cold PVC) is same speed. Subsequent builds with typical
3-5 file changes should drop from ~20 min to ~2-3 min via incremental.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three optimizations targeting GPU allocation churn and lock contention:
1. In-place polyak update: Var::set() reuses existing GPU buffer instead
of Var::from_tensor() which allocates a new one per call. Eliminates
~10,000 cudaMalloc/cudaFree per epoch (20 params × 500 steps).
2. Fused affine ops: Replace 13 Tensor::full()/Tensor::ones() constant
tensor allocations per step with tensor.affine(mul, add) — a single
fused kernel. Patterns: 1-x → x.affine(-1,1), γ*x → x.affine(γ,0),
0.5*x² → (x*x).affine(0.5,0). Applied across all 4 loss paths
(branching Bellman/Huber, standard Bellman/Huber, IQN). Eliminates
~6,500 GPU allocs/epoch.
3. Batch pre-sampling (K=8): Sample 8 batches under one READ lock, train
all 8 under one WRITE lock. Reduces async lock acquisitions from
2×N to 2×ceil(N/8). Priority staleness across 8 steps is negligible.
Combined estimated impact: 20-35% H100 throughput improvement.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Candle's Tensor::cumsum(0) internally allocates an [n,n] upper-triangular
matrix (9.3 GB for n=50K) causing OOM on GPUs ≤48 GB. Replace with a
block-parallel Hillis-Steele scan kernel that is O(n) in time and memory.
Additional fixes in this commit:
- Break autograd chain leak in loss/grad accumulation via .detach()
(was leaking ~32 MB/step across entire training run)
- Release features_raw_cuda/targets_raw_cuda after GPU experience
collection (~164 MB VRAM reclaimed)
- Use softmax eval (temp=0.3) in walk-forward backtest to prevent
action collapse causing trades=0 on early-stage models
Validated: 1286 tests pass (408 ml-dqn + 878 ml), 0 failures.
VRAM stable at 754 MB across 45K+ training steps on RTX 3050 Ti.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Prevents experience collector output tensors from starving backward
graph and GPU PER on VRAM-constrained GPUs like RTX 3050 Ti 4GB.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
free_memory_mb is sampled after model/optimizer load, so subtract only
backward-graph + fragmentation + data upload headroom (400 MB), not the
full model/PER/data stack. Increase collector share to 50%.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Staging/GPU replay buffer: truncate batch when it exceeds ring buffer
capacity (CPU fallback can flush 100K+ experiences at once)
- GpuExperienceCollector: compute shmem tile rows dynamically to stay
under 48 KB default limit instead of hardcoded 64-row constant
- Auto batch size: subtract concurrent VRAM consumers (~530 MB model +
optimizer + PER + data) before budgeting experience collector at 40%
- Hyperopt DQN adapter: GPU PER always on, include replay buffer in
VRAM estimate for small GPUs (was excluded assuming CPU-only PER)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move the cfg gate inside the method body instead of gating the entire
function. Without cuda, the GpuPrioritized variant doesn't exist so
matches! returns false — no need for separate cfg blocks at call sites.
Fixes compile-services CI failure (services build without --features cuda).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- DQN hyperopt adapter: remove static VRAM gate for GPU experience
collector and GPU PER — dynamic scaling handles constraints at
runtime with graceful CPU fallback on init failure
- Remove is_parquet_file branching from hyperopt eval path (all data
loads via DBN pipeline now)
- Update training example CLI args for consistency
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace monolithic compile-all with granular per-binary change detection.
detect-changes now outputs space-separated package/example lists based on
a dependency map from source directories to binary targets:
- Shared crates (common, config, Cargo.toml) → all binaries
- Service-specific dirs → only that service binary
- Domain crates (trading_engine, risk) → dependent service subset
- ML crates → ml-training-service + trading-service + all training
- ML subdirs (trainers/, hyperopt/, evaluation/) → specific training binaries
compile-services and compile-training accept package lists and build only
affected binaries, saving ~20-30s link time per skipped binary.
deploy-services restarts only affected deployments (trading-service
excluded from auto-deploy for safety).
Fix: 'latest' package update now replaces individual files instead of
deleting the entire package, preventing corruption during partial builds.
compile-and-train-template: derive needed training binaries from model
parameter (3 instead of 7), drop unused training_uploader build.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Eliminate all per-step GPU→CPU synchronization barriers from the
training guard. Replace device+host buffer pairs and memcpy_dtoh
with cuMemHostAlloc(DEVICEMAP) mapped pinned memory.
Key changes:
- MappedBuffer struct: cuMemHostAlloc + cuMemHostGetDevicePointer_v2
allocates memory visible to both CPU and GPU simultaneously
- Double-buffering: kernel writes to buffer[N%2], CPU reads buffer
[(N-1)%2] — one-step delayed halt detection, zero sync
- __threadfence_system() in CUDA kernels ensures writes visible to
CPU across PCIe without explicit memcpy
- read_volatile on host pointer prevents CPU-side caching
Eliminated:
- check_and_accumulate: 28-byte memcpy_dtoh (every training step)
- qvalue_stats: 16-byte memcpy_dtoh (every 50 steps)
- qvalue_divergence: 20-byte memcpy_dtoh (every 50 steps)
Kept: read_accumulators memcpy_dtoh (12 bytes, epoch boundary only —
accumulator buffer stays in device memory for kernel read-modify-write).
1286 tests pass (878 ml + 408 ml-dqn), 0 failures.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Two remaining GPU→CPU synchronization barriers in the CUDA-active
training loop:
1. detect_dead_neurons() → to_vec0() called every training step from
log_diagnostics() in the guard path. Added check_gradient_collapse()
method that performs the same collapse detection logic but skips the
expensive per-parameter weight scan. Dead neuron detection now only
runs at epoch boundary via log_diagnostics().
2. forward() Q-value clipping monitoring → to_vec2() called every 1000
steps during compute_loss_internal() and Q-value estimation forward
passes. Added training_forward_active flag that gates the monitoring
block; set to true during all training-path forward() calls
(compute_loss_internal + Q-value estimation), false during
inference/evaluation.
All 1577 tests pass (878 ml + 408 ml-dqn + 291 ml-core), 0 failures.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Wire route_exposure_to_factored() into both select_actions_batch_gpu
and select_actions_batch GPU paths, eliminating per-item CPU routing.
Previously, exposure indices (0-4) were downloaded from GPU and routed
to factored indices (0-44) one-by-one on CPU via route_action(). Now
the exposure→factored mapping runs entirely on GPU via the routing
kernel, with a single batch readback of the final factored indices.
Branching DQN path unchanged (already produces factored indices 0-44).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace CPU-bound Q-value monitoring with GPU-resident qvalue_stats and
qvalue_divergence kernels from GpuTrainingGuard. The two readback sites
(estimate_avg_q_value_with_early_stopping's mean_all().to_scalar() and
log_q_values' to_vec2()) are now bypassed when the GPU training guard is
active. CPU fallback path preserved for non-CUDA builds and guard-absent
scenarios.
Changes:
- Add DQN::log_q_values_from_stats() accepting pre-computed GPU stats
- Add DQNAgentType::log_q_values_from_stats() delegate
- Replace Q-estimation in train_step_single_batch with GPU kernel path
- Replace Q-estimation in train_step_with_accumulation with GPU kernel path
- Import IndexOp trait for Tensor::i() in trainer.rs
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds a standalone CUDA kernel that converts DQN exposure indices (0-4)
to factored action indices (0-44) entirely on GPU, reusing the existing
route_order() device function from common_device_functions.cuh. Wired
into GpuActionSelector as route_exposure_to_factored() method, following
the same DtoD copy pattern as the existing select_actions methods. This
eliminates a GPU->CPU->GPU roundtrip when post-hoc routing is needed
after epsilon_greedy_select.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace the GPU->CPU sync barrier (to_vec1 readback) in the DQN training
hot path with the GpuTrainingGuard CUDA kernel that performs NaN detection,
loss clipping, and gradient collapse checks entirely on-device. The guard
is lazy-initialized on first training step and falls back to the original
CPU readback path if CUDA kernel compilation fails.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Seven pure-Rust reference tests covering all four training_guard_kernel.cu
kernels: PTX NVRTC smoke-test (skips on CPU-only), NaN/Inf detection,
loss clip, gradient collapse, accumulator averaging, GuardResult boolean
threshold construction, and qvalue_stats_reduce per-sample max-Q reduction.
Zero clippy warnings in the new file.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Wraps 4 CUDA kernels (training_guard_check, training_guard_accumulate,
qvalue_stats_reduce, qvalue_divergence_check) with a Rust struct that
uses OnceLock PTX caching, pre-allocated device buffers, and host-side
Vec mirrors for zero-allocation readbacks per training step.
Accumulator (3-float acc_buf) stays on-device for epoch-boundary
averaging without CPU roundtrips.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
11-task plan across 4 chunks: Training Guard kernel + wrapper (Tasks 1-3),
wire into trainer (Tasks 4-6), Q-value monitor + action routing (Tasks 7-9),
experience collector audit + final verification (Tasks 10-11).
Eliminates all 10 GPU→CPU sync barriers from the DQN training hot path.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Defines architecture for eliminating all 10 GPU→CPU sync barriers from
the training loop via 4 new GPU components: Training Guard (pinned
memory predicates), Q-Value Monitor (on-device accumulator), GPU-resident
action selection, and async experience collector readback.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add use_dsr, dsr_eta, and n_steps fields to ExperienceCollectorConfig
with safe defaults (DSR off, eta=0.01, n_steps=1). Pass them as kernel
args after fill_simulation_enabled, matching the parameter order added
in Tasks 4 and 5. Wire from DqnHyperparams in the trainer hot path.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Mirror the DSR (Differential Sharpe Ratio) and n-step return
accumulation added by Task 4 to the scalar kernel into the warp
kernel (dqn_full_experience_kernel_warp). DSR/n-step state runs
on lane 0 only, matching the existing reward computation pattern.
All accumulators are reset on episode boundaries.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds Differential Sharpe Ratio reward shaping and n-step return
accumulation to dqn_full_experience_kernel (scalar path). DSR replaces
EMA normalization when use_dsr=1; n-step ring buffer accumulates
discounted returns for effective_n>1. Both are reset on episode
boundaries.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Append third CUDA kernel to epsilon_greedy_kernel.cu for Branching DQN:
3 independent epsilon-greedy heads (exposure/5, order/3, urgency/3)
compose to factored action index 0-44 (exposure*9 + order*3 + urgency).
Add Rust composition tests verifying full 45-action coverage and
round-trip decomposition correctness.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add backticks to type names in doc comments (doc_markdown)
- Mark eligible functions as const fn (missing_const_for_fn)
No behavior changes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
module_name_repetitions: PpoConfig in ppo module is conventional ML naming.
Renaming breaks every import across the workspace.
integer_division: Basis point calculations, batch size math, combinatorial
formulas — truncation is intentional. Float conversion would introduce bugs.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
is_gpu_prioritized() and grad_norm squeeze are only available with CUDA.
Without the gate, compile-services (CPU-only) fails with E0599.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace per-step to_scalar()/to_vec1() readbacks with GPU-resident
tensor accumulation and single epoch-boundary sync. On H100 this
removes ~12-15 pipeline flushes per training step (~5μs each),
enabling full GPU saturation with zero CPU sync in the inner loop.
Key changes:
- GpuTrainResult: train_step() returns GPU scalar tensors (loss_gpu,
grad_norm_gpu) instead of f32 — zero readback per step
- Regime conditional: train all 3 heads unconditionally with
zero-masked weights (mathematical no-op) instead of 3-6
to_scalar() mask count checks per step
- GPU PER: max_priority as GPU tensor with flush_max_priority(),
delta-based priority update via index_add (no CPU dedup)
- Deferred diagnostics: NaN detection, CQL logging, Q-value
estimation all moved to epoch boundary where pipeline is
already synced
- Legacy CPU-readback wrappers removed entirely
9 files changed, +393/-295 lines. 520 DQN tests passing.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
