Files
foxhunt/ml/examples/benchmark_streaming_vs_batch.rs
jgrusewski 35feadf55e 🚀 Wave 160 Phase 6: CUDA Mandatory + TDD Testing + TFT Complete (21 Agents)
## Major Achievements

### 1. CUDA Made Default & Mandatory (Agent 143)
- CUDA now default feature in ml/Cargo.toml
- All training requires GPU (no silent CPU fallback)
- Added get_training_device() helper with fail-fast errors
- Removed --use-gpu flags (GPU mandatory)
- **Impact**: No more wasting time on accidental CPU training

### 2. TFT Training COMPLETE (Agent 144)
-  Training completed successfully in 7.6 minutes
-  Early stopping at epoch 100/200 (best val loss: 0.097318)
-  11 checkpoints saved to ml/trained_models/production/tft/
-  GPU Performance: 99% utilization, 367MB VRAM, 4.4s/epoch
-  10x speedup vs CPU (4.4s vs 43-55s per epoch)
- **Status**: PRODUCTION READY

### 3. TFT CUDA Tensor Contiguity Fix (Agent 142)
- Fixed "matmul not supported for non-contiguous tensors" error
- Added .contiguous() call after narrow() operation in QuantileLayer
- Enabled CUDA-accelerated TFT training
- **Files**: ml/src/tft/quantile_outputs.rs

### 4. MAMBA-2 CUDA Layer Normalization (Agent 145)
- Created CudaLayerNorm wrapper for missing CUDA kernel
- Implemented manual layer norm: γ * (x - μ) / sqrt(σ² + ε) + β
- MAMBA-2 now runs on CUDA (no more "no cuda implementation" error)
- **Files**: ml/src/mamba/mod.rs

### 5. TDD E2E Test Suite (Agent 146) 
- Created comprehensive MAMBA-2 test suite (297 lines)
- 7 tests: shapes, batches, CUDA, gradients, configs
- **16x faster debugging**: 5s per iteration vs 80s
- Already caught dtype mismatch bug (F32 vs F64)
- **Files**: ml/tests/e2e_mamba2_training.rs

## Agent Summary (Agents 126-146)

### Code Fixes (Parallel - Agents 137-141)
- **Agent 137**: MAMBA-2 batch dimension fix (streaming + batch loaders)
- **Agent 138**: Liquid NN API fix (mutable loader, iterator fix)
- **Agent 139**: PPO CheckpointMetadata fix (signature fields)
- **Agent 140**: Paper trading executor (498 lines, 100ms polling)
- **Agent 141**: Real model loading (RealDQNModel, RealPPOModel)

### Infrastructure (Agents 143-146)
- **Agent 143**: CUDA mandatory (Cargo.toml, device helpers)
- **Agent 144**: TFT verification (completion monitoring)
- **Agent 145**: MAMBA-2 CUDA layer norm wrapper
- **Agent 146**: TDD E2E test suite (16x faster debugging)

## Files Modified

### Core ML Infrastructure
- ml/Cargo.toml: Added default = ["minimal-inference", "cuda"]
- ml/src/lib.rs: Added get_training_device() helper (+109 lines)
- ml/src/tft/quantile_outputs.rs: Fixed tensor contiguity
- ml/src/mamba/mod.rs: Added CudaLayerNorm wrapper (+41 lines)

### Training Scripts
- ml/examples/train_tft_dbn.rs: Removed --use-gpu flag
- ml/examples/train_ppo.rs: Removed --use-gpu flag
- ml/examples/train_mamba2_dbn.rs: Forced CUDA-only mode
- ml/examples/train_liquid_dbn.rs: Fixed API usage

### Data Loaders
- ml/src/data_loaders/dbn_sequence_loader.rs: Fixed batch dimensions
- ml/src/data_loaders/streaming_dbn_loader.rs: Fixed batch dimensions

### Trading Service
- services/trading_service/src/paper_trading_executor.rs: New executor (+498 lines)
- services/trading_service/src/services/enhanced_ml.rs: Real model loading
- services/trading_service/src/ensemble_coordinator.rs: Integration

### Tests
- ml/tests/e2e_mamba2_training.rs: New TDD test suite (+297 lines)

### Trainers
- ml/src/trainers/tft.rs: Fixed CheckpointMetadata signature fields

## Performance Metrics

### TFT Training
- Duration: 7.6 minutes (100 epochs with early stopping)
- GPU Utilization: 99%
- GPU Memory: 367MB / 4GB (9%)
- Epoch Time: 4.4 seconds (vs 43-55s on CPU)
- Speedup: 10x vs CPU
- Status:  PRODUCTION READY

### TDD Testing
- Test Execution: 5-10 seconds per test
- Debugging Iteration: 5 seconds (vs 80 seconds before)
- Speedup: 16x faster debugging
- First Bug Found: <1 minute (dtype mismatch)

## Documentation
- 21 comprehensive agent reports
- TDD quick start guide
- CUDA troubleshooting guide
- Training verification procedures

## Next Steps
1. Fix MAMBA-2 dtype mismatch (F32→F64) - 2 minutes
2. Run MAMBA-2 tests until passing - 5-10 minutes
3. Launch full MAMBA-2 training - 200 epochs
4. Launch Liquid NN training

## System Status
- TFT:  COMPLETE (production ready)
- MAMBA-2: 🧪 IN TESTING (TDD suite ready)
- CUDA:  DEFAULT (mandatory for training)
- Tests:  16x faster debugging

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
2025-10-14 23:13:34 +02:00

372 lines
11 KiB
Rust

//! Benchmark: Streaming vs Batch Data Loading
//!
//! Compares memory usage and performance between batch and streaming loaders.
//!
//! ## Metrics Compared
//!
//! - Peak memory usage (RSS)
//! - Loading time
//! - Sequences per second
//! - Memory efficiency ratio
//!
//! ## Usage
//!
//! ```bash
//! # Small dataset (4 files, ~400KB)
//! cargo run -p ml --example benchmark_streaming_vs_batch --release -- \
//! --data-dir test_data/real/databento/ml_training_small
//!
//! # Large dataset (360 files, ~15MB)
//! cargo run -p ml --example benchmark_streaming_vs_batch --release -- \
//! --data-dir test_data/real/databento/ml_training
//! ```
use anyhow::Result;
use clap::Parser;
use ml::data_loaders::{DbnSequenceLoader, StreamingDbnLoader};
use std::path::PathBuf;
use std::time::Instant;
#[derive(Parser, Debug)]
#[command(author, version, about)]
struct Args {
/// Directory containing DBN files
#[arg(long, default_value = "test_data/real/databento/ml_training_small")]
data_dir: PathBuf,
/// Sequence length
#[arg(long, default_value = "60")]
seq_len: usize,
/// Model dimension
#[arg(long, default_value = "256")]
d_model: usize,
/// Train/validation split
#[arg(long, default_value = "0.9")]
train_split: f64,
/// Streaming batch size (bars)
#[arg(long, default_value = "10000")]
batch_size: usize,
/// Stride for sliding window
#[arg(long, default_value = "100")]
stride: usize,
}
/// Memory statistics
#[derive(Debug, Clone)]
struct MemoryStats {
rss_kb: usize,
vms_kb: usize,
}
impl MemoryStats {
/// Get current memory usage
fn current() -> Result<Self> {
let status = std::fs::read_to_string("/proc/self/status")?;
let mut rss_kb = 0;
let mut vms_kb = 0;
for line in status.lines() {
if line.starts_with("VmRSS:") {
rss_kb = line
.split_whitespace()
.nth(1)
.and_then(|s| s.parse().ok())
.unwrap_or(0);
} else if line.starts_with("VmSize:") {
vms_kb = line
.split_whitespace()
.nth(1)
.and_then(|s| s.parse().ok())
.unwrap_or(0);
}
}
Ok(Self { rss_kb, vms_kb })
}
fn rss_mb(&self) -> f64 {
self.rss_kb as f64 / 1024.0
}
fn vms_mb(&self) -> f64 {
self.vms_kb as f64 / 1024.0
}
}
/// Benchmark results
#[derive(Debug)]
struct BenchmarkResult {
name: String,
total_sequences: usize,
duration_secs: f64,
sequences_per_sec: f64,
peak_memory_mb: f64,
memory_efficiency_ratio: f64,
}
impl BenchmarkResult {
fn print_report(&self) {
println!("\n{:=<60}", "");
println!(" {} BENCHMARK RESULTS", self.name.to_uppercase());
println!("{:=<60}", "");
println!(" Total Sequences: {}", self.total_sequences);
println!(" Duration: {:.2}s", self.duration_secs);
println!(" Throughput: {:.0} sequences/sec", self.sequences_per_sec);
println!(" Peak Memory (RSS): {:.1} MB", self.peak_memory_mb);
println!(" Memory Efficiency: {:.2}x", self.memory_efficiency_ratio);
println!("{:=<60}\n", "");
}
}
/// Benchmark batch loading
async fn benchmark_batch(args: &Args) -> Result<BenchmarkResult> {
println!("\n🔄 BATCH LOADING BENCHMARK");
println!(" Loading all data into memory...\n");
// Measure baseline memory
let baseline_memory = MemoryStats::current()?;
println!(" Baseline memory: {:.1} MB RSS", baseline_memory.rss_mb());
let start = Instant::now();
let mut peak_memory = baseline_memory.clone();
// Create batch loader
let mut loader = DbnSequenceLoader::with_limits(
args.seq_len,
args.d_model,
Some(1_000), // Limit sequences to prevent OOM
args.stride,
)
.await?;
// Load all sequences at once
let (train_data, val_data) = loader.load_sequences(&args.data_dir, args.train_split).await?;
// Measure peak memory
let current_memory = MemoryStats::current()?;
if current_memory.rss_kb > peak_memory.rss_kb {
peak_memory = current_memory;
}
let duration = start.elapsed();
let total_sequences = train_data.len() + val_data.len();
println!(" ✅ Loaded {} sequences", total_sequences);
println!(" Peak memory: {:.1} MB RSS", peak_memory.rss_mb());
println!(" Duration: {:.2}s", duration.as_secs_f64());
Ok(BenchmarkResult {
name: "Batch Loading".to_string(),
total_sequences,
duration_secs: duration.as_secs_f64(),
sequences_per_sec: total_sequences as f64 / duration.as_secs_f64(),
peak_memory_mb: peak_memory.rss_mb() - baseline_memory.rss_mb(),
memory_efficiency_ratio: 1.0, // Baseline
})
}
/// Benchmark streaming loading
async fn benchmark_streaming(args: &Args, batch_baseline: &BenchmarkResult) -> Result<BenchmarkResult> {
println!("\n🌊 STREAMING LOADING BENCHMARK");
println!(" Loading data in batches of {} bars...\n", args.batch_size);
// Measure baseline memory
let baseline_memory = MemoryStats::current()?;
println!(" Baseline memory: {:.1} MB RSS", baseline_memory.rss_mb());
let start = Instant::now();
let mut peak_memory = baseline_memory.clone();
// Create streaming loader
let loader = StreamingDbnLoader::with_config(
args.seq_len,
args.d_model,
args.batch_size,
args.stride,
)
.await?;
// Stream sequences
let mut stream = loader.stream_sequences(&args.data_dir, args.train_split).await?;
let mut total_sequences = 0;
let mut batch_count = 0;
// Process training data
loop {
// Measure memory before batch
let current_memory = MemoryStats::current()?;
if current_memory.rss_kb > peak_memory.rss_kb {
peak_memory = current_memory;
}
match stream.next_batch().await? {
Some(batch) => {
batch_count += 1;
total_sequences += batch.len();
if batch_count % 10 == 0 {
let current_mem = MemoryStats::current()?;
println!(
" Batch {}: {} sequences (memory: {:.1} MB)",
batch_count,
batch.len(),
current_mem.rss_mb() - baseline_memory.rss_mb()
);
}
}
None => break,
}
}
let duration = start.elapsed();
println!(" ✅ Processed {} sequences in {} batches", total_sequences, batch_count);
println!(" Peak memory: {:.1} MB RSS", peak_memory.rss_mb());
println!(" Duration: {:.2}s", duration.as_secs_f64());
// Calculate memory efficiency ratio
let memory_used = peak_memory.rss_mb() - baseline_memory.rss_mb();
let memory_efficiency = batch_baseline.peak_memory_mb / memory_used.max(1.0);
Ok(BenchmarkResult {
name: "Streaming Loading".to_string(),
total_sequences,
duration_secs: duration.as_secs_f64(),
sequences_per_sec: total_sequences as f64 / duration.as_secs_f64(),
peak_memory_mb: memory_used,
memory_efficiency_ratio: memory_efficiency,
})
}
/// Print comparison table
fn print_comparison(batch: &BenchmarkResult, streaming: &BenchmarkResult) {
println!("\n{:=<80}", "");
println!(" COMPREHENSIVE COMPARISON");
println!("{:=<80}", "");
println!();
println!(" {:<30} {:>20} {:>20}", "Metric", "Batch", "Streaming");
println!(" {:-<30} {:-<20} {:-<20}", "", "", "");
println!(
" {:<30} {:>20} {:>20}",
"Sequences Loaded",
batch.total_sequences,
streaming.total_sequences
);
println!(
" {:<30} {:>18.2}s {:>18.2}s",
"Duration",
batch.duration_secs,
streaming.duration_secs
);
let speed_ratio = streaming.duration_secs / batch.duration_secs;
let speed_pct = (speed_ratio - 1.0) * 100.0;
println!(
" {:<30} {:>20.0} {:>20.0} ({:+.1}%)",
"Throughput (seq/s)",
batch.sequences_per_sec,
streaming.sequences_per_sec,
-speed_pct
);
println!(
" {:<30} {:>18.1} MB {:>18.1} MB",
"Peak Memory",
batch.peak_memory_mb,
streaming.peak_memory_mb
);
let memory_reduction = (1.0 - streaming.peak_memory_mb / batch.peak_memory_mb) * 100.0;
println!(
" {:<30} {:>20} {:>18.2}x ({:.0}% reduction)",
"Memory Efficiency",
"1.0x",
streaming.memory_efficiency_ratio,
memory_reduction
);
println!("\n{:=<80}", "");
// Success criteria check
println!("\n SUCCESS CRITERIA:");
println!(" {:-<80}", "");
let memory_ok = streaming.peak_memory_mb < 512.0;
let speed_ok = speed_pct.abs() < 10.0;
println!(
" ✓ Memory < 512MB: {} ({:.1} MB)",
if memory_ok { "✅ PASS" } else { "❌ FAIL" },
streaming.peak_memory_mb
);
println!(
" ✓ Speed penalty < 10%: {} ({:+.1}%)",
if speed_ok { "✅ PASS" } else { "❌ FAIL" },
speed_pct
);
if memory_ok && speed_ok {
println!("\n 🎉 ALL SUCCESS CRITERIA MET!");
} else {
println!("\n ⚠️ Some criteria not met - may need tuning");
}
println!("{:=<80}\n", "");
}
#[tokio::main]
async fn main() -> Result<()> {
// Initialize logging
tracing_subscriber::fmt()
.with_max_level(tracing::Level::INFO)
.init();
let args = Args::parse();
println!("\n{:=<80}", "");
println!(" STREAMING VS BATCH DATA LOADING BENCHMARK");
println!("{:=<80}", "");
println!(" Data Directory: {:?}", args.data_dir);
println!(" Sequence Length: {}", args.seq_len);
println!(" Model Dimension: {}", args.d_model);
println!(" Batch Size: {} bars", args.batch_size);
println!(" Stride: {}", args.stride);
println!(" Train Split: {:.0}%", args.train_split * 100.0);
println!("{:=<80}\n", "");
// Verify data directory exists
if !args.data_dir.exists() {
eprintln!("❌ Error: Data directory not found: {:?}", args.data_dir);
eprintln!("\nAvailable test directories:");
eprintln!(" - test_data/real/databento/ml_training_small (4 files, ~400KB)");
eprintln!(" - test_data/real/databento/ml_training (360 files, ~15MB)");
std::process::exit(1);
}
// Run benchmarks
println!("🚀 Starting benchmarks...\n");
let batch_result = benchmark_batch(&args).await?;
batch_result.print_report();
// Force garbage collection between benchmarks
println!("🧹 Cleaning up memory...");
tokio::time::sleep(tokio::time::Duration::from_secs(2)).await;
let streaming_result = benchmark_streaming(&args, &batch_result).await?;
streaming_result.print_report();
// Print comparison
print_comparison(&batch_result, &streaming_result);
Ok(())
}