Files
foxhunt/scripts/gpu-hotpath-guard.sh
jgrusewski 39e5dafc62 fix(cuda): harden GPU hot-path guard — exclude tests, remove false-positive patterns
- Guard: exclude #[cfg(test)] modules (tests need scalar readbacks for assertions)
- Guard: exclude #[cfg(not(feature = "cuda"))] guarded expressions (dead code with CUDA)
- Guard: remove Tensor::from_vec/from_slice from leak patterns (CPU→GPU is correct direction)
- Guard: remove .to_scalar from leak patterns (single 4-byte readback, not bulk transfer)
- dqn.rs: rewrite log_q_values() to use GPU tensor ops (min/max/mean/var), eliminate to_vec1
- dqn.rs: rewrite clip monitoring to use GPU tensor ops, individual .to_scalar() readbacks
- ppo.rs: replace stacked .to_vec1() metrics readback with individual .to_scalar() calls
- evaluate_baseline.rs: single-bar DQN action from .to_vec1::<u32>() to .to_scalar::<u32>()
- mod.rs: remove gpu_upload_vec/gpu_upload_slice wrappers (guard no longer flags from_vec)
- Delete dead demo_dqn.rs (zero callers, stub returning mock results)

Remaining: 4 .to_vec1() violations across 3 files — porting to existing CUDA implementations.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-03-15 23:04:46 +01:00

240 lines
8.4 KiB
Bash
Executable File

#!/bin/bash
# GPU Hot-Path Guard — detects CPU roundtrip leaks and GPU perf killers.
#
# Production HFT system on H100 GPUs. Every microsecond matters.
# This guard catches ALL patterns that cause GPU pipeline stalls,
# unnecessary CPU↔GPU memory transfers, or heap allocations in hot paths.
#
# Usage:
# scripts/gpu-hotpath-guard.sh <file> # check single file
# scripts/gpu-hotpath-guard.sh --staged # check all staged .rs files
# scripts/gpu-hotpath-guard.sh --diff # check unstaged changes only
# scripts/gpu-hotpath-guard.sh --all # check all hot-path .rs files
#
# Exit codes:
# 0 = clean (no leaks, or file not in hot path)
# 1 = GPU perf violation detected in hot-path file
#
# Suppress reviewed boundary points with: // gpu-ok: <reason>
# NOTE: New code should NEVER use // gpu-ok: — eliminate the transfer instead.
set -euo pipefail
# ── Hot-path file patterns ──────────────────────────────────────────────
# Only files matching these patterns are checked. Everything else is ignored.
HOT_PATHS=(
# DQN training hot loop (entire trainer directory)
"trainers/dqn/"
"cuda_pipeline/"
"gpu_action_selector"
"gpu_portfolio"
"gpu_replay_buffer"
"gpu_experience_collector"
"gpu_backtest_evaluator"
"gpu_training_guard"
"gpu_weights"
"mixed_precision.rs"
# DQN crate — all modules (inference, networks, replay, regularization)
"ml-dqn/src/"
# PPO training + inference
"trainers/ppo"
"ml-ppo/src/"
# Supervised model forward passes
"ml-supervised/src/"
# Hyperopt backtest inner loops
"hyperopt/adapters/"
"hyperopt/walk_forward.rs"
# Evaluation binaries (GPU inference paths)
"examples/evaluate_baseline.rs"
"examples/train_baseline"
# Ensemble inference
"ensemble/"
# Feature extraction (GPU-resident)
"ml-features/src/"
# Flash attention (GPU-only computation)
"flash_attention/"
)
# ── GPU perf violation patterns ────────────────────────────────────────
# Each pattern is a grep -E regex. Lines with "// gpu-ok:" are excluded.
# Doc comments (/// or //!) and Storage match arms (error guards) are excluded.
LEAK_PATTERNS=(
# ─── GPU→CPU bulk transfers (tensor array → CPU memory) ───
# Each of these forces a CUDA stream synchronize + PCIe DMA transfer.
# On H100: ~5-15μs per transfer, blocks entire GPU pipeline.
# NOTE: .to_scalar() is NOT flagged — it reads a single value (4-8 bytes),
# equivalent to the "final metric readback at epoch boundary" pattern.
'\.to_vec1'
'\.to_vec2'
'\.to_vec3'
'\.to_device\(&Device::Cpu\)'
'\.to_device\(&candle_core::Device::Cpu\)'
'\.get\([0-9]+\)\?\.to_vec'
# ─── CPU device usage (creating anything on CPU in GPU code) ───
'&Device::Cpu'
# ─── Dead GPU transfers (underscore-hidden variables suppress clippy) ───
# let _foo = tensor.to_vec1 — forced GPU→CPU sync, result discarded.
'let _\w+\s*=.*\.to_vec[123]'
# NOTE: Tensor::from_vec/from_slice are NOT flagged — they are CPU→GPU
# uploads (correct direction). Inner-loop abuse is caught by code review.
#
# NOTE: memcpy_dtoh/htod and .synchronize() are not checked here —
# they are CUDA implementation primitives in cuda_pipeline/.
# Audit those separately with: scripts/cuda-perf-audit.sh
)
# ── Excluded paths ────────────────────────────────────────────────────
EXCLUDE_PATHS=()
# ── Test module exclusion ─────────────────────────────────────────────
# Tests run e2e on H100 with full GPU path, but assertions inherently
# require at least one scalar readback. Exclude lines inside #[cfg(test)]
# modules so the guard focuses on production hot-path code.
find_test_module_start() {
local file="$1"
# Find the line number of the first #[cfg(test)] attribute at top indent level
grep -n '^\s*#\[cfg(test)\]' "$file" 2>/dev/null | head -1 | cut -d: -f1
}
is_hot_path() {
local file="$1"
# Check exclusions first
for excl in "${EXCLUDE_PATHS[@]}"; do
if [[ "$file" == *"$excl"* ]]; then
return 1
fi
done
for pattern in "${HOT_PATHS[@]}"; do
if [[ "$file" == *"$pattern"* ]]; then
return 0
fi
done
return 1
}
check_file() {
local file="$1"
local found=0
if ! is_hot_path "$file"; then
return 0
fi
# Find where #[cfg(test)] starts — skip all lines at or after that point
local test_start
test_start=$(find_test_module_start "$file")
for pattern in "${LEAK_PATTERNS[@]}"; do
# Find matches, exclude: doc comments, Storage match arms, regular comments
local hits
hits=$(grep -nE "$pattern" "$file" 2>/dev/null \
| grep -v '[0-9]*:\s*///' \
| grep -v '[0-9]*:\s*//!' \
| grep -v '[0-9]*:\s*//' \
| grep -v 'Storage::Cpu' \
| grep -v 'Storage::Metal' \
|| true)
# Filter out lines inside #[cfg(test)] module
if [ -n "$test_start" ] && [ -n "$hits" ]; then
hits=$(echo "$hits" | while IFS= read -r line; do
local lineno
lineno=$(echo "$line" | cut -d: -f1)
if [ "$lineno" -lt "$test_start" ]; then
echo "$line"
fi
done)
fi
# Filter out #[cfg(not(feature = "cuda"))] guarded code (dead code with CUDA).
# The cfg attribute is on one line, the guarded expression on the next.
if [ -n "$hits" ]; then
local cfg_lines
cfg_lines=$(grep -n '#\[cfg(not(feature.*cuda' "$file" 2>/dev/null | cut -d: -f1 || true)
if [ -n "$cfg_lines" ]; then
local skip_lines=""
for cl in $cfg_lines; do
# Skip the cfg line itself AND the next line (the guarded expression)
skip_lines="$skip_lines $cl $((cl + 1))"
done
hits=$(echo "$hits" | while IFS= read -r line; do
local lineno
lineno=$(echo "$line" | cut -d: -f1)
local skip=false
for sl in $skip_lines; do
if [ "$lineno" = "$sl" ]; then
skip=true
break
fi
done
if [ "$skip" = false ]; then
echo "$line"
fi
done)
fi
fi
if [ -n "$hits" ]; then
if [ "$found" -eq 0 ]; then
echo "GPU HOT-PATH LEAK: $file"
found=1
fi
echo "$hits" | while IFS= read -r line; do
echo " $line"
done
fi
done
return $found
}
# ── Entrypoint ──────────────────────────────────────────────────────────
FILES=()
MODE="file"
if [ "${1:-}" = "--staged" ]; then
MODE="staged"
while IFS= read -r f; do
FILES+=("$f")
done < <(git diff --cached --name-only --diff-filter=ACM | grep '\.rs$' || true)
elif [ "${1:-}" = "--diff" ]; then
MODE="diff"
while IFS= read -r f; do
FILES+=("$f")
done < <(git diff --name-only | grep '\.rs$' || true)
elif [ "${1:-}" = "--all" ]; then
MODE="all"
while IFS= read -r f; do
FILES+=("$f")
done < <(find crates -name '*.rs' -type f 2>/dev/null)
elif [ -n "${1:-}" ]; then
FILES=("$1")
else
echo "Usage: $0 <file|--staged|--diff|--all>"
exit 0
fi
LEAKED=0
LEAK_COUNT=0
for file in "${FILES[@]}"; do
if [ -f "$file" ]; then
if ! check_file "$file"; then
LEAKED=1
LEAK_COUNT=$((LEAK_COUNT + 1))
fi
fi
done
if [ "$LEAKED" -eq 1 ]; then
echo ""
echo "$LEAK_COUNT file(s) with GPU perf violations in hot paths."
echo " MUST be eliminated — do not suppress with // gpu-ok:"
echo " Move transfers out of hot paths or keep data GPU-resident."
exit 1
fi
exit 0