foxhunt/ml/tests/gpu_memory_budget_validation.rs

//! GPU Memory Budget Validation Test
//!
//! Comprehensive test to verify that all 4 trained models (DQN, PPO, MAMBA-2, TFT)
//! fit within the RTX 3050 Ti 4GB VRAM budget with sufficient headroom for inference.
//!
//! ## Test Objectives
//!
//! 1. Measure baseline GPU memory usage
//! 2. Load each model sequentially and measure memory consumption
//! 3. Verify total memory budget <4GB (4096 MB)
//! 4. Verify >500MB headroom for inference buffers
//! 5. Generate detailed memory breakdown table
//!
//! ## Expected Memory Targets
//!
//! - DQN:     <150 MB (validated: 6 MB ✅)
//! - PPO:     <200 MB (validated: 145 MB ✅)
//! - MAMBA-2: <500 MB (validated: 164 MB ✅)
//! - TFT:     <500 MB (needs validation)
//! - Total:   <815 MB target (~20% of 4GB)
//! - Headroom: >500 MB for inference (target: >3200 MB free)
//!
//! ## RTX 3050 Ti Specifications
//!
//! - Total VRAM: 4096 MB (4 GB)
//! - CUDA Cores: 2560
//! - Compute Capability: 8.6
//! - Memory Bandwidth: 192 GB/s

use candle_core::Device;
use ml::benchmark::memory_profiler::MemoryProfiler;
use ml::dqn::{WorkingDQN, WorkingDQNConfig};
use ml::ppo::{PPOConfig, UnifiedPPO};
use ml::mamba::Mamba2SSM;
use ml::tft::{TrainableTFT, TFTConfig};
use ml::MLError;

/// GPU memory budget test configuration
const GPU_TOTAL_MB: f64 = 4096.0;
const MIN_HEADROOM_MB: f64 = 500.0;
const DQN_TARGET_MB: f64 = 150.0;
const PPO_TARGET_MB: f64 = 200.0;
const MAMBA2_TARGET_MB: f64 = 500.0;
const TFT_TARGET_MB: f64 = 500.0;

/// Individual model memory measurement
#[derive(Debug, Clone)]
struct ModelMemory {
    name: String,
    memory_mb: f64,
    target_mb: f64,
    meets_target: bool,
}

impl ModelMemory {
    fn new(name: &str, memory_mb: f64, target_mb: f64) -> Self {
        Self {
            name: name.to_string(),
            memory_mb,
            target_mb,
            meets_target: memory_mb <= target_mb,
        }
    }

    fn percent_of_budget(&self) -> f64 {
        (self.memory_mb / GPU_TOTAL_MB) * 100.0
    }

    fn percent_of_target(&self) -> f64 {
        (self.memory_mb / self.target_mb) * 100.0
    }
}

/// Complete memory budget analysis
#[derive(Debug)]
struct MemoryBudgetReport {
    baseline_mb: f64,
    models: Vec<ModelMemory>,
    total_memory_mb: f64,
    headroom_mb: f64,
    meets_budget: bool,
    meets_headroom: bool,
}

impl MemoryBudgetReport {
    fn print_summary(&self) {
        println!("\n{}", "=".repeat(70));
        println!("GPU MEMORY BUDGET VALIDATION REPORT");
        println!("{}", "=".repeat(70));
        println!();
        println!("GPU: RTX 3050 Ti (4GB VRAM)");
        println!("Total Budget: {:.0} MB", GPU_TOTAL_MB);
        println!("Required Headroom: {:.0} MB", MIN_HEADROOM_MB);
        println!();

        // Baseline
        println!("Baseline GPU Memory: {:.0} MB", self.baseline_mb);
        println!();

        // Individual models
        println!("MODEL MEMORY BREAKDOWN:");
        println!("{}", "-".repeat(70));
        println!("{:<15} {:>10} {:>10} {:>12} {:>10} {:>8}",
                 "Model", "Memory", "Target", "%Budget", "%Target", "Status");
        println!("{}", "-".repeat(70));

        for model in &self.models {
            let status = if model.meets_target { "✅ PASS" } else { "❌ FAIL" };
            println!("{:<15} {:>8.0} MB {:>8.0} MB {:>11.2}% {:>9.1}% {:>8}",
                     model.name,
                     model.memory_mb,
                     model.target_mb,
                     model.percent_of_budget(),
                     model.percent_of_target(),
                     status);
        }

        println!("{}", "-".repeat(70));
        println!("{:<15} {:>8.0} MB {:>10} {:>11.2}% {:>9} {:>8}",
                 "TOTAL",
                 self.total_memory_mb,
                 "",
                 (self.total_memory_mb / GPU_TOTAL_MB) * 100.0,
                 "",
                 if self.meets_budget { "✅ PASS" } else { "❌ FAIL" });
        println!("{}", "=".repeat(70));
        println!();

        // Headroom analysis
        println!("HEADROOM ANALYSIS:");
        println!("{}", "-".repeat(70));
        println!("Total Model Memory:  {:.0} MB ({:.1}% of budget)",
                 self.total_memory_mb,
                 (self.total_memory_mb / GPU_TOTAL_MB) * 100.0);
        println!("Available Headroom:  {:.0} MB ({:.1}% of budget)",
                 self.headroom_mb,
                 (self.headroom_mb / GPU_TOTAL_MB) * 100.0);
        println!("Required Headroom:   {:.0} MB", MIN_HEADROOM_MB);
        println!("Status:              {}",
                 if self.meets_headroom { "✅ PASS" } else { "❌ FAIL" });
        println!("{}", "=".repeat(70));
        println!();

        // Overall verdict
        let all_pass = self.meets_budget && self.meets_headroom &&
                       self.models.iter().all(|m| m.meets_target);

        if all_pass {
            println!("🎉 OVERALL: ✅ ALL TESTS PASSED");
            println!();
            println!("All 4 models fit within RTX 3050 Ti 4GB VRAM budget with");
            println!("sufficient headroom ({:.0} MB) for inference operations.", self.headroom_mb);
        } else {
            println!("❌ OVERALL: TESTS FAILED");
            println!();
            if !self.meets_budget {
                println!("⚠️  Total memory exceeds 4GB budget!");
            }
            if !self.meets_headroom {
                println!("⚠️  Insufficient headroom for inference!");
            }
            for model in &self.models {
                if !model.meets_target {
                    println!("⚠️  {} exceeds target ({:.0} MB > {:.0} MB)",
                             model.name, model.memory_mb, model.target_mb);
                }
            }
        }
        println!("{}", "=".repeat(70));
    }

    fn print_ascii_bar_chart(&self) {
        println!("\n{}", "=".repeat(70));
        println!("MEMORY USAGE BAR CHART");
        println!("{}", "=".repeat(70));
        println!();

        let max_width = 50;

        for model in &self.models {
            let bar_width = ((model.memory_mb / GPU_TOTAL_MB) * max_width as f64) as usize;
            let bar = "█".repeat(bar_width);
            println!("{:<10} │{:<50}│ {:.0} MB", model.name, bar, model.memory_mb);
        }

        println!("{}", "-".repeat(70));

        let total_bar_width = ((self.total_memory_mb / GPU_TOTAL_MB) * max_width as f64) as usize;
        let total_bar = "█".repeat(total_bar_width);
        println!("{:<10} │{:<50}│ {:.0} MB", "TOTAL", total_bar, self.total_memory_mb);

        let headroom_bar_width = ((self.headroom_mb / GPU_TOTAL_MB) * max_width as f64) as usize;
        let headroom_bar = "░".repeat(headroom_bar_width);
        println!("{:<10} │{:<50}│ {:.0} MB", "HEADROOM", headroom_bar, self.headroom_mb);

        println!();
        println!("Scale: 0 MB{:>61} 4096 MB", "");
        println!("{}", "=".repeat(70));
    }
}

/// Measure GPU memory for a specific model
fn measure_model_memory<F>(
    profiler: &mut MemoryProfiler,
    baseline_mb: f64,
    model_name: &str,
    load_fn: F,
) -> Result<f64, MLError>
where
    F: FnOnce() -> Result<(), MLError>,
{
    println!("Loading {} model...", model_name);

    // Load model
    load_fn()?;

    // Take memory snapshot
    let snapshot = profiler.take_snapshot().map_err(|e| {
        MLError::TrainingError(format!("Failed to take memory snapshot: {}", e))
    })?;

    // Calculate memory delta
    let model_memory_mb = snapshot.vram_used_mb - baseline_mb;

    println!("  {} Memory: {:.0} MB", model_name, model_memory_mb);

    Ok(model_memory_mb)
}

#[test]
#[ignore] // Only run with --ignored flag (requires GPU)
fn test_gpu_memory_budget_all_models() -> Result<(), MLError> {
    println!("\n{}", "=".repeat(70));
    println!("GPU MEMORY BUDGET VALIDATION TEST");
    println!("{}", "=".repeat(70));
    println!();

    // Initialize device
    let device = Device::cuda_if_available(0)?;

    match &device {
        Device::Cpu => {
            println!("⚠️  CPU device detected - skipping GPU memory test");
            println!("This test requires CUDA GPU (RTX 3050 Ti)");
            return Ok(());
        }
        Device::Cuda(_) => {
            println!("✅ CUDA GPU detected: {:?}", device);
        }
        _ => {
            println!("⚠️  Unknown device - skipping test");
            return Ok(());
        }
    }

    // Initialize memory profiler
    let mut profiler = MemoryProfiler::new(0);

    // Measure baseline memory
    let baseline_snapshot = profiler.take_snapshot().map_err(|e| {
        MLError::TrainingError(format!("Failed to measure baseline memory: {}", e))
    })?;
    let baseline_mb = baseline_snapshot.vram_used_mb;

    println!("Baseline GPU Memory: {:.0} MB", baseline_mb);
    println!("Total GPU VRAM: {:.0} MB", baseline_snapshot.vram_total_mb);
    println!();

    // Storage for model measurements
    let mut model_memories = Vec::new();

    // Test 1: DQN Model
    println!("Test 1/4: DQN Model");
    println!("{}", "-".repeat(70));

    let dqn_config = WorkingDQNConfig {
        state_dim: 16,
        num_actions: 3,
        hidden_dims: vec![256, 256],
        learning_rate: 0.001,
        gamma: 0.99,
        epsilon_start: 1.0,
        epsilon_end: 0.01,
        epsilon_decay: 0.995,
        replay_buffer_capacity: 10000,
        batch_size: 32,
        min_replay_size: 100,
        target_update_freq: 100,
        use_double_dqn: true,
    };

    let dqn_memory_mb = measure_model_memory(
        &mut profiler,
        baseline_mb,
        "DQN",
        move || {
            let _dqn = WorkingDQN::new(dqn_config)?;
            Ok(())
        },
    )?;

    model_memories.push(ModelMemory::new("DQN", dqn_memory_mb, DQN_TARGET_MB));
    println!();

    // Test 2: PPO Model
    println!("Test 2/4: PPO Model");
    println!("{}", "-".repeat(70));

    use ml::ppo::GAEConfig;

    let ppo_config = PPOConfig {
        state_dim: 16,
        num_actions: 3,
        policy_hidden_dims: vec![256, 256],
        value_hidden_dims: vec![256, 256],
        policy_learning_rate: 0.0003,
        value_learning_rate: 0.001,
        clip_epsilon: 0.2,
        value_loss_coeff: 0.5,
        entropy_coeff: 0.01,
        gae_config: GAEConfig {
            gamma: 0.99,
            lambda: 0.95,
            normalize_advantages: true,
        },
        batch_size: 64,
        mini_batch_size: 32,
        num_epochs: 10,
        max_grad_norm: 0.5,
    };

    let ppo_baseline = profiler.take_snapshot()
        .map_err(|e| MLError::TrainingError(format!("PPO baseline snapshot failed: {}", e)))?
        .vram_used_mb;

    let device_clone2 = device.clone();
    let ppo_memory_mb = measure_model_memory(
        &mut profiler,
        ppo_baseline,
        "PPO",
        move || {
            let _ppo = UnifiedPPO::new(ppo_config, device_clone2)?;
            Ok(())
        },
    )?;

    model_memories.push(ModelMemory::new("PPO", ppo_memory_mb, PPO_TARGET_MB));
    println!();

    // Test 3: MAMBA-2 Model
    println!("Test 3/4: MAMBA-2 Model");
    println!("{}", "-".repeat(70));

    let mamba2_baseline = profiler.take_snapshot()
        .map_err(|e| MLError::TrainingError(format!("MAMBA-2 baseline snapshot failed: {}", e)))?
        .vram_used_mb;

    let device_clone3 = device.clone();
    let mamba2_memory_mb = measure_model_memory(
        &mut profiler,
        mamba2_baseline,
        "MAMBA-2",
        move || {
            let _mamba2 = Mamba2SSM::default_hft(&device_clone3)?;
            Ok(())
        },
    )?;

    model_memories.push(ModelMemory::new("MAMBA-2", mamba2_memory_mb, MAMBA2_TARGET_MB));
    println!();

    // Test 4: TFT Model
    println!("Test 4/4: TFT Model");
    println!("{}", "-".repeat(70));

    let tft_config = TFTConfig {
        input_dim: 16,
        hidden_dim: 256,
        num_heads: 4,
        num_layers: 3,
        prediction_horizon: 10,
        sequence_length: 50,
        num_quantiles: 9,
        num_static_features: 4,
        num_known_features: 8,
        num_unknown_features: 4,
        learning_rate: 0.001,
        batch_size: 32,
        dropout_rate: 0.1,
        l2_regularization: 0.001,
        use_flash_attention: true,
        mixed_precision: true,
        memory_efficient: true,
        max_inference_latency_us: 50,
        target_throughput_pps: 100_000,
    };

    let tft_baseline = profiler.take_snapshot()
        .map_err(|e| MLError::TrainingError(format!("TFT baseline snapshot failed: {}", e)))?
        .vram_used_mb;

    let tft_memory_mb = measure_model_memory(
        &mut profiler,
        tft_baseline,
        "TFT",
        || {
            let _tft = TrainableTFT::new(tft_config)?;
            Ok(())
        },
    )?;

    model_memories.push(ModelMemory::new("TFT", tft_memory_mb, TFT_TARGET_MB));
    println!();

    // Calculate totals
    let total_memory_mb: f64 = model_memories.iter().map(|m| m.memory_mb).sum();
    let headroom_mb = GPU_TOTAL_MB - total_memory_mb;
    let meets_budget = total_memory_mb < GPU_TOTAL_MB;
    let meets_headroom = headroom_mb > MIN_HEADROOM_MB;

    // Generate report
    let report = MemoryBudgetReport {
        baseline_mb,
        models: model_memories,
        total_memory_mb,
        headroom_mb,
        meets_budget,
        meets_headroom,
    };

    // Print results
    report.print_summary();
    report.print_ascii_bar_chart();

    // Assertions
    assert!(meets_budget,
            "Total memory ({:.0} MB) exceeds 4GB budget ({:.0} MB)",
            total_memory_mb, GPU_TOTAL_MB);

    assert!(meets_headroom,
            "Insufficient headroom ({:.0} MB) for inference buffers (required: {:.0} MB)",
            headroom_mb, MIN_HEADROOM_MB);

    // Verify individual model targets
    for model in &report.models {
        assert!(model.meets_target,
                "{} exceeds target: {:.0} MB > {:.0} MB",
                model.name, model.memory_mb, model.target_mb);
    }

    println!();
    println!("🎉 GPU MEMORY BUDGET VALIDATION: ALL TESTS PASSED ✅");
    println!();

    Ok(())
}

#[test]
#[ignore] // Only run with --ignored flag (requires GPU)
fn test_gpu_memory_budget_conservative_estimate() -> Result<(), MLError> {
    println!("\n{}", "=".repeat(70));
    println!("GPU MEMORY BUDGET CONSERVATIVE ESTIMATE");
    println!("{}", "=".repeat(70));
    println!();
    println!("This test uses validated memory measurements from previous tests:");
    println!("- DQN:     6 MB (validated in Wave 7.17)");
    println!("- PPO:     145 MB (validated in Wave 7.18)");
    println!("- MAMBA-2: 164 MB (validated in Wave 6)");
    println!("- TFT:     Estimated 400-500 MB (needs validation)");
    println!();

    // Conservative estimates based on previous validations
    let dqn_memory = 6.0;
    let ppo_memory = 145.0;
    let mamba2_memory = 164.0;
    let tft_memory_estimate = 500.0; // Conservative upper bound

    let total_memory = dqn_memory + ppo_memory + mamba2_memory + tft_memory_estimate;
    let headroom = GPU_TOTAL_MB - total_memory;

    println!("CONSERVATIVE MEMORY ESTIMATE:");
    println!("{}", "-".repeat(70));
    println!("DQN:              {:>8.0} MB (validated)", dqn_memory);
    println!("PPO:              {:>8.0} MB (validated)", ppo_memory);
    println!("MAMBA-2:          {:>8.0} MB (validated)", mamba2_memory);
    println!("TFT:              {:>8.0} MB (estimated)", tft_memory_estimate);
    println!("{}", "-".repeat(70));
    println!("TOTAL:            {:>8.0} MB ({:.1}% of 4GB)", total_memory, (total_memory / GPU_TOTAL_MB) * 100.0);
    println!("HEADROOM:         {:>8.0} MB ({:.1}% of 4GB)", headroom, (headroom / GPU_TOTAL_MB) * 100.0);
    println!("{}", "=".repeat(70));
    println!();

    // Assertions
    assert!(total_memory < GPU_TOTAL_MB,
            "Conservative estimate ({:.0} MB) exceeds 4GB budget", total_memory);

    assert!(headroom > MIN_HEADROOM_MB,
            "Conservative estimate leaves insufficient headroom: {:.0} MB < {:.0} MB",
            headroom, MIN_HEADROOM_MB);

    println!("✅ Conservative estimate: {:.0} MB total ({:.1}% of budget)",
             total_memory, (total_memory / GPU_TOTAL_MB) * 100.0);
    println!("✅ Headroom available: {:.0} MB ({:.1}% of budget)",
             headroom, (headroom / GPU_TOTAL_MB) * 100.0);
    println!();
    println!("🎉 CONSERVATIVE ESTIMATE: PASS ✅");
    println!();

    Ok(())
}