Hamerspace 🔨

A compiler-style optimization pass for non-LLM machine learning models

Hamerspace is a model compression and optimization engine that orchestrates existing open-source toolkits to compress and optimize computer vision, audio, time-series, and tabular ML models.

Features

• 🎯 Goal-oriented optimization: Specify what you want (size, latency, accuracy) and let Hamerspace figure out how
• 🔧 Multi-backend orchestration: Automatically selects and composes tools from PyTorch, TensorFlow, ONNX, OpenVINO, TVM, and more
• 📊 Comprehensive benchmarking: Measures size, latency, and accuracy before and after optimization
• 🎨 Hardware-aware: Optimizes for specific target hardware (CPU, ARM, edge devices)
• 📦 Deployment-ready: Produces optimized model artifacts ready for production

Installation

pip install hamerspace

For full backend support (OpenVINO, TVM, bitsandbytes):

pip install hamerspace[full]

Quick Start

from hamerspace import Optimizer, OptimizationGoal, Constraints

# Load your trained model
optimizer = Optimizer.from_pytorch("model.pt")

# Define constraints
constraints = Constraints(
    target_size_mb=10,          # Must be under 10MB
    max_latency_ms=50,          # Must inference in <50ms
    max_accuracy_drop=0.02,     # Max 2% accuracy drop
    target_hardware="cpu"       # Optimize for CPU
)

# Optimize
result = optimizer.optimize(
    goal=OptimizationGoal.AUTO,
    constraints=constraints
)

# Save optimized model
result.save_model("optimized_model.onnx")

# View report
print(result.report)

Optimization Goals

• OptimizationGoal.QUANTIZE: Apply quantization (INT8, INT4)
• OptimizationGoal.PRUNE: Remove unnecessary weights
• OptimizationGoal.DISTILL: Knowledge distillation (requires training data)
• OptimizationGoal.AUTO: Automatically select best techniques

Supported Frameworks

Input Models

• PyTorch (.pt, .pth)
• TensorFlow (.h5, SavedModel)
• ONNX (.onnx)

Backend Toolkits

• PyTorch (quantization, pruning)
• TensorFlow (quantization)
• ONNX Runtime (quantization, graph optimization)
• OpenVINO (optimization for Intel hardware)
• Apache TVM (compilation and optimization)
• Hugging Face Optimum (hardware-specific optimization)

Architecture

┌─────────────────────────────────────────────────────┐
│                  Public API Layer                    │
│         (Optimizer, Constraints, Goals)              │
└────────────────────┬────────────────────────────────┘
                     │
┌────────────────────▼────────────────────────────────┐
│              Orchestration Layer                     │
│    (Strategy Selection, Pipeline Composition)        │
└────────────────────┬────────────────────────────────┘
                     │
┌────────────────────▼────────────────────────────────┐
│               Backend Layer                          │
│  ┌──────────┬──────────┬──────────┬──────────┐     │
│  │ PyTorch  │   ONNX   │ OpenVINO │   TVM    │     │
│  │ Backend  │ Backend  │ Backend  │ Backend  │     │
│  └──────────┴──────────┴──────────┴──────────┘     │
└─────────────────────────────────────────────────────┘

Advanced Usage

Custom Backend Selection

from hamerspace import Optimizer, Backend

optimizer = Optimizer.from_pytorch("model.pt")
result = optimizer.optimize(
    goal=OptimizationGoal.QUANTIZE,
    constraints=constraints,
    preferred_backends=[Backend.ONNX, Backend.OPENVINO]
)

Benchmarking Only

# Benchmark without optimization
metrics = optimizer.benchmark(
    hardware="cpu",
    num_runs=100
)
print(f"Latency: {metrics.latency_ms}ms")
print(f"Size: {metrics.size_mb}MB")

Export Optimization Config

# Save configuration for reproducibility
result.save_config("optimization_config.json")

# Reproduce optimization
from hamerspace import Optimizer
optimizer = Optimizer.from_config("optimization_config.json")
result = optimizer.apply()

Non-Goals

❌ LLM optimization (use specialized tools like vLLM, TensorRT-LLM)
❌ Training models from scratch
❌ Research or SOTA benchmarking
❌ Custom kernel development

Requirements

• Python 3.8+
• One or more supported backends installed

Contributing

Contributions welcome! Please see CONTRIBUTING.md for guidelines.

License

Apache License 2.0

Citation

@software{hamerspace2025,
  title={Hamerspace: Model Compression and Optimization Engine},
  author={Hamerspace Contributors},
  year={2025},
  url={https://github.com/yourusername/hamerspace}
}