Chapter 1: Introduction

The Problem Space

Modern ML training at scale faces critical challenges:

GPU Failure Rates

• 0.1% GPU failure rate → 10% throughput loss with TP64
• Mean Time Between Failures (MTBF): 26-56 hours for large clusters
• Recovery requires full job restart, losing hours of progress

Current Solutions Fall Short

• Checkpointing: 30-40 minute overhead, still loses recent work
• Redundancy: Expensive, doesn't handle software failures
• Static allocation: Can't adapt to changing resources

The BEAM Inspiration

The Erlang/BEAM runtime achieves 99.999% uptime through:

Actor Model

• Lightweight processes with isolated state
• Message passing only communication
• No shared memory, no locks

Preemptive Scheduling

• Reduction-based fair scheduling
• No process monopolizes the system
• Predictable latency

Supervision Trees

• "Let it crash" philosophy
• Automatic restart strategies
• Hierarchical failure isolation

Hot Code Swapping

• Update running systems
• Zero downtime deployments
• Gradual rollout

zbmd Vision

Apply BEAM's proven principles to GPU/ML workloads:

Actors for Everything

pub const MLActor = union(enum) {
    tensor: TensorActor,        // Data containers
    operator: OperatorActor,    // Computations
    layer: LayerActor,          // Model components
    optimizer: OptimizerActor,  // Training logic
    supervisor: SupervisorActor,// Fault handling
};

GPU Kernels as Actors

• Each CUDA kernel is a supervised actor
• Automatic retry on failure
• Migration to healthy GPUs

Distributed by Default

• Transparent multi-node operation
• Automatic work distribution
• Elastic scaling

Design Principles

1. Safety First (Tiger Style)

• No undefined behavior
• Fixed memory limits
• Bounded operations
• Fail fast with recovery

2. Zero-Cost Abstractions

• Comptime dispatch for backends
• No runtime overhead
• Direct GPU memory access

3. Test-Driven Development

• Tests alongside implementation
• Property-based testing
• Fault injection testing

4. Domain-Driven Design

• Clear bounded contexts
• Actor boundaries match domains
• Message contracts as APIs

Non-Goals

zbmd is NOT:

• A Python wrapper (pure Zig)
• A general actor framework (ML-focused)
• A distributed database (though it could be)
• A web framework (compute only)

Success Metrics

Performance

• < 1% overhead vs native CUDA
• Linear scaling to 10,000 GPUs
• Sub-millisecond failure detection

Reliability

• 99.999% uptime for training
• Zero data loss on failures
• Automatic recovery < 1 second

Usability

• Single binary deployment
• Zero configuration defaults
• Drop-in PyTorch replacement