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>
15-task plan covering Phase 1 (seal GPU→CPU roundtrips in training loop),
Phase 2 (vectorized GPU backtest environment for hyperopt), and Phase 3
(general GPU backtester integration).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Three-phase plan to eliminate all GPU→CPU roundtrips from training:
- Phase 1: Seal training loop (persistent GPU epoch state, async monitoring)
- Phase 2: Vectorized CUDA backtest kernel for hyperopt evaluation
- Phase 3: General-purpose GPU backtester replacing CPU SIMD path
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Root cause: `option_env!("FOXHUNT_BUILD_VERSION")` in common/build_info.rs
was a compile-time macro tracked by cargo fingerprints (Rust 1.80+). Every
pipeline run with a new tag invalidated `common` → cascading rebuild of all
38 dependent workspace crates, even when zero source files changed.
Fix: replace `option_env!()` with `std::env::var()` (runtime LazyLock). Cargo
no longer tracks the version env var, so `common` only recompiles when its
source actually changes.
Also: skip git checkout when HEAD already matches target SHA (zero mtime
changes), and drop `-x` from git clean to preserve gitignored files.
Expected: ~3.7min → <1min for unchanged-crate rebuilds on warm PVC.
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>
Keep gpu-warmup only in compile-and-train-template where it runs
parallel with the CPU compile step. The other templates don't need
it — CI pipeline doesn't always train, and training-workflow already
has the GPU node available from the triggering event.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
git clone gives all files the current timestamp, causing cargo to
recompile every workspace crate even when source is unchanged. Switch
to persistent checkout on the PVC: git fetch + git checkout --force
only updates files that actually changed between commits, preserving
mtimes on unchanged files so cargo correctly skips them.
- compile-services: /cargo-target/src persistent checkout
- compile-training: /cargo-target/src persistent checkout
- compile-and-train: /cargo-target/src persistent checkout + CARGO_HOME
- First run: full clone (fallback), subsequent: fetch + checkout
Expected: ~40 crate recompile → only changed crates (~1-2 min)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
With 37+ workspace crates, cargo's incremental fingerprinting already
skips unchanged crates — sccache adds per-rustc wrapping overhead for
no benefit. The real speedup is CARGO_HOME on PVC (eliminates crate
re-downloads every build).
Kept: CARGO_HOME=/cargo-target/cargo-home on PVC
Removed: RUSTC_WRAPPER=sccache, SCCACHE_DIR, CARGO_INCREMENTAL=0
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Both compile-services and compile-training were doing full rebuilds
every run because:
1. No RUSTC_WRAPPER=sccache — sccache was installed but never used
2. No persistent CARGO_HOME — crates.io registry re-downloaded each time
3. Incremental compilation doesn't work across git clones (mtime-based)
Fix: enable sccache (content-hash based, clone-agnostic), persist
CARGO_HOME on the PVC, disable incremental (conflicts with sccache).
Expected: first run populates cache, subsequent runs ~2-3min vs ~8min.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
apply-argo-templates failed because ci-deploy role was missing
permissions for eventbus (argoproj.io), services (core), and
roles/rolebindings (rbac) needed to apply events/*.yaml files.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Argo forbids mixing `dependencies` and `depends` in the same DAG.
Converted all tasks to `depends` with explicit state conditions so
infra steps can wait for image rebuilds while still running when
rebuilds are skipped.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Uses Argo `depends` conditions so apply-argo-templates and
terragrunt-apply wait for rebuild-ci-builder-cpu to finish (or skip)
before running, ensuring they always get the latest image.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
apply-argo-templates needs git (clone repo) + kubectl (apply manifests).
ci-builder-cpu has git but lacked kubectl — added it. foxhunt-runtime
has kubectl but no git and runs as non-root, so can't be used here.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Pipeline now auto-applies WorkflowTemplates, EventSources, Sensors,
and RBAC when infra/k8s/argo/ files change — no more manual kubectl.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add terragrunt-apply step to Argo CI pipeline (plan+apply on main push
when infra/live/ or infra/modules/ change)
- Bake OpenTofu 1.9.0 + Terragrunt 0.77.12 into ci-builder-cpu image
with SHA256 checksum verification
- Remove 3 ghost node pools (foxhunt, gitlab, h100-sxm8) from kapsule
module to match Scaleway reality
- Make terragrunt.hcl single source of truth (remove variable defaults)
- Fix GitLab TF state lock methods (POST/DELETE for HTTP backend)
- Harden PAT rotation: more retries, verification step, recovery docs
- Add weekly PAT expiry check CronJob (warns 14 days before expiry)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Replace fragile `git fetch --depth=1 origin "$SHA"` (fails with short SHAs)
with `git clone --no-checkout --filter=blob:none` + `git checkout "$SHA"`
across all 3 Argo templates (5 clone blocks total)
- Upgrade CI compile pool from POP2-32C-128G to POP2-HC-32C-64G
(higher clock EPYC cores, 28% cheaper at €0.851/hr vs €1.18/hr)
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>
