foxhunt/scripts/deploy_fp32_runpod.sh

#!/bin/bash
# FP32 Runpod Deployment Script
# Generated: 2025-10-23
# Purpose: One-command FP32 model training on Runpod

set -euo pipefail

# Configuration
PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
TEST_DATA="${PROJECT_ROOT}/test_data/ES_FUT_180d.parquet"
EPOCHS="${EPOCHS:-50}"
MODEL_OUTPUT="${PROJECT_ROOT}/models"

# Colors for output
RED='\033[0;31m'
GREEN='\033[0;32m'
YELLOW='\033[1;33m'
BLUE='\033[0;34m'
NC='\033[0m' # No Color

# Logging functions
log_info() {
    echo -e "${BLUE}[INFO]${NC} $1"
}

log_success() {
    echo -e "${GREEN}[SUCCESS]${NC} $1"
}

log_warn() {
    echo -e "${YELLOW}[WARN]${NC} $1"
}

log_error() {
    echo -e "${RED}[ERROR]${NC} $1"
}

# Pre-deployment checks
run_preflight_checks() {
    log_info "Running pre-deployment checks..."

    # Check 1: Test data exists
    if [[ ! -f "${TEST_DATA}" ]]; then
        log_error "Test data not found: ${TEST_DATA}"
        exit 1
    fi
    log_success "Test data found: ${TEST_DATA} ($(du -h ${TEST_DATA} | cut -f1))"

    # Check 2: Docker services running
    if ! docker ps | grep -q foxhunt-postgres; then
        log_error "Docker services not running. Start with: docker-compose up -d"
        exit 1
    fi
    log_success "Docker services running (postgres, redis, vault)"

    # Check 3: GPU available
    if ! nvidia-smi &>/dev/null; then
        log_warn "nvidia-smi not found. GPU training may not work."
        log_warn "This is expected on CPU-only systems."
    else
        GPU_INFO=$(nvidia-smi --query-gpu=name,memory.free --format=csv,noheader)
        log_success "GPU available: ${GPU_INFO}"
    fi

    # Check 4: Database migration 045 applied
    if ! psql postgresql://foxhunt:foxhunt_dev_password@localhost:5432/foxhunt \
         -c "\dt regime_states" &>/dev/null; then
        log_error "Database migration 045 not applied. Run: cargo sqlx migrate run"
        exit 1
    fi
    log_success "Database migration 045 applied (regime tables exist)"

    # Check 5: Cargo available
    if ! command -v cargo &>/dev/null; then
        log_error "Cargo not found. Install Rust toolchain first."
        exit 1
    fi
    log_success "Cargo available: $(cargo --version)"

    log_success "All pre-flight checks passed ✅"
    echo ""
}

# Build release binaries
build_release() {
    log_info "Building release binaries (this may take 5-10 minutes)..."

    cd "${PROJECT_ROOT}"

    if cargo build --release --package ml --features cuda 2>&1 | tee /tmp/build.log; then
        log_success "Release build completed successfully"
    else
        log_error "Release build failed. Check /tmp/build.log for details."
        exit 1
    fi

    echo ""
}

# Train FP32 model
train_fp32_model() {
    log_info "Starting FP32 TFT training (${EPOCHS} epochs)..."
    log_info "Model: TFT-225 (FP32)"
    log_info "Dataset: ES.FUT 180 days"
    log_info "Expected Duration: ~3-5 minutes (RTX 4090) or ~10-15 minutes (RTX 3050 Ti)"
    echo ""

    cd "${PROJECT_ROOT}"

    # Create models directory if it doesn't exist
    mkdir -p "${MODEL_OUTPUT}"

    # Record start time
    START_TIME=$(date +%s)

    # Run training (NO --use-qat flag for FP32)
    if cargo run -p ml --example train_tft_parquet --release --features cuda -- \
        --parquet-file "${TEST_DATA}" \
        --epochs "${EPOCHS}"; then

        # Record end time
        END_TIME=$(date +%s)
        DURATION=$((END_TIME - START_TIME))
        DURATION_MIN=$((DURATION / 60))
        DURATION_SEC=$((DURATION % 60))

        log_success "Training completed in ${DURATION_MIN}m ${DURATION_SEC}s"
        echo ""

        # Find latest model file
        LATEST_MODEL=$(ls -t "${MODEL_OUTPUT}"/tft_225_fp32_*.safetensors 2>/dev/null | head -1)
        if [[ -n "${LATEST_MODEL}" ]]; then
            MODEL_SIZE=$(du -h "${LATEST_MODEL}" | cut -f1)
            log_success "Model saved: ${LATEST_MODEL} (${MODEL_SIZE})"
        else
            log_warn "Model file not found in ${MODEL_OUTPUT}"
        fi
    else
        log_error "Training failed. Check logs above for details."
        exit 1
    fi

    echo ""
}

# Monitor GPU during training
monitor_gpu() {
    log_info "GPU monitoring enabled (press Ctrl+C to stop)"
    log_info "Watching GPU memory usage every 1 second..."
    echo ""

    watch -n 1 nvidia-smi
}

# Baseline metrics collection
record_baseline_metrics() {
    log_info "Recording baseline metrics..."

    METRICS_FILE="${PROJECT_ROOT}/FP32_BASELINE_METRICS.md"

    cat > "${METRICS_FILE}" << EOF
# FP32 Baseline Metrics (Runpod)

**Date**: $(date +"%Y-%m-%d %H:%M:%S")
**GPU**: $(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null || echo "N/A")
**Model**: TFT-225 FP32
**Dataset**: ES.FUT 180 days (test_data/ES_FUT_180d.parquet)

## Training Metrics
- **Training Time**: ${DURATION_MIN}m ${DURATION_SEC}s
- **Epochs Completed**: ${EPOCHS}
- **GPU Memory Peak**: (monitor with nvidia-smi during training)

## Model Artifacts
- **Model Path**: ${LATEST_MODEL}
- **Model Size**: ${MODEL_SIZE}

## Next Steps
1. ✅ FP32 model trained successfully
2. 🔲 Run inference benchmark: \`cargo test -p ml --release test_tft_inference_latency\`
3. 🔲 Compare with baseline RMSE/MAE metrics
4. 🔲 Upload to S3/MinIO for production use
5. 🔲 Begin QAT Phase 2 (after P0 fixes)

## Notes
- FP32 deployment successful with zero blockers
- QAT deferred to Phase 2 (1-2 weeks after P0 fixes)
- Expected performance: Sharpe 2.00, Win Rate 60%, Drawdown 15%

---
**Generated by**: scripts/deploy_fp32_runpod.sh
EOF

    log_success "Baseline metrics recorded: ${METRICS_FILE}"
    echo ""
}

# Main execution
main() {
    echo "════════════════════════════════════════════════════════════"
    echo "  FP32 Runpod Deployment Script"
    echo "  Foxhunt HFT Trading System"
    echo "════════════════════════════════════════════════════════════"
    echo ""

    # Run pre-flight checks
    run_preflight_checks

    # Ask user to confirm
    read -p "Proceed with FP32 model training? [y/N] " -n 1 -r
    echo
    if [[ ! $REPLY =~ ^[Yy]$ ]]; then
        log_warn "Deployment cancelled by user"
        exit 0
    fi

    # Build release binaries
    build_release

    # Train model
    train_fp32_model

    # Record baseline metrics
    if [[ -n "${DURATION_MIN}" ]]; then
        record_baseline_metrics
    fi

    # Success summary
    echo "════════════════════════════════════════════════════════════"
    log_success "FP32 DEPLOYMENT COMPLETE ✅"
    echo "════════════════════════════════════════════════════════════"
    echo ""
    log_info "Next Actions:"
    echo "  1. Review baseline metrics: cat FP32_BASELINE_METRICS.md"
    echo "  2. Run inference benchmark: cargo test -p ml --release test_tft_inference_latency"
    echo "  3. Upload model to S3/MinIO: cargo run -p storage --example upload_model"
    echo "  4. Begin QAT Phase 2 (after P0 fixes): See RUNPOD_DEPLOYMENT_CHECKLIST.md"
    echo ""
    log_info "For GPU monitoring during training, run in separate terminal:"
    echo "  watch -n 1 nvidia-smi"
    echo ""
}

# Execute main function
main "$@"