anydeploy 0.2.1

CLI tool and library to export ML models to production formats and containerize them with Docker

anydeploy

Deploy ML models anywhere

Export ML models to production formats (ONNX, TFLite, TorchScript) and deploy them locally or at the edge.

anydeploy makes model deployment easy. Convert your trained models to optimized inference formats, benchmark performance, validate correctness, generate serving code, and containerize everything -- all from a single CLI or Python API.

Edge-first deployment. Supports ONNX Runtime (CPU/GPU/edge), TFLite (mobile/edge), and llama.cpp (local LLM serving). All deployment targets work completely offline.

Built and maintained by Viet-Anh Nguyen at NRL.ai.

Installation

# Core (CLI + config + benchmarking)
pip install anydeploy

# With specific framework support
pip install anydeploy[torch]      # PyTorch + TorchScript
pip install anydeploy[onnx]       # ONNX + ONNX Runtime
pip install anydeploy[tflite]     # TensorFlow Lite
pip install anydeploy[serve]      # FastAPI serving

# Everything
pip install anydeploy[all]

Quick Start

CLI

# Export a PyTorch model to ONNX
anydeploy export model.pt --format onnx --input-shape 1,3,224,224

# Export to TFLite
anydeploy export model.pt --format tflite --input-shape 1,3,224,224

# Benchmark an exported model
anydeploy benchmark model.onnx --runs 100

# Serve a model with FastAPI
anydeploy serve model.onnx --backend fastapi --port 8000

# Generate a Docker container for deployment
anydeploy dockerize model.onnx --base python:3.11-slim

Python API

import anydeploy

# Export a model
anydeploy.export(model, format="onnx", input_shape=(1, 3, 224, 224))

# Benchmark performance
result = anydeploy.benchmark("model.onnx", runs=100)
print(f"Mean latency: {result.mean_latency_ms:.2f} ms")
print(f"P95 latency:  {result.p95_latency_ms:.2f} ms")
print(f"Throughput:   {result.throughput:.1f} inferences/sec")

# Validate exported model against original
report = anydeploy.validate(original_model, "model.onnx", test_input)
print(f"Max difference: {report.max_diff}")
print(f"Passed: {report.passed}")

# Generate Dockerfile and serving code
from anydeploy.config import DockerConfig
docker_cfg = DockerConfig(base_image="python:3.11-slim")
anydeploy.dockerize("model.onnx", docker_cfg)

# Register a custom exporter
from anydeploy.export.base import BaseExporter
class MyExporter(BaseExporter):
    def export(self, model, output_path, config=None):
        ...
anydeploy.register_exporter("myformat", MyExporter)

Export Format Comparison

Format	Framework	Hardware	Optimization	File Size
ONNX	Any (via ONNX Runtime)	CPU, GPU, Edge	Graph optimization	Medium
TFLite	TensorFlow	Mobile, Edge	Quantization	Small
TorchScript	PyTorch	CPU, GPU	JIT compilation	Large

Serving

anydeploy generates production-ready serving code for multiple backends:

# FastAPI server for ONNX/TFLite/TorchScript models
anydeploy serve model.onnx --backend fastapi --port 8000

# llama.cpp server for GGUF language models (edge LLM deployment)
anydeploy serve model.gguf --backend llamacpp --port 8080

FastAPI Backend

Creates a FastAPI application with:

• /predict endpoint accepting JSON or binary input
• /health health check endpoint
• Automatic input validation
• Configurable batch size

llama.cpp Backend

Creates deployment scripts for serving GGUF language models locally:

• Shell script to launch llama.cpp server
• Dockerfile for containerized LLM serving
• OpenAI-compatible /v1/chat/completions endpoint
• Works on CPU, GPU, and edge devices

Docker Deployment

Generate a complete Docker setup for your model:

anydeploy dockerize model.onnx --base python:3.11-slim --port 8000

This creates:

• Dockerfile with optimized layers
• serve.py FastAPI application
• requirements.txt with pinned dependencies

Extensibility

anydeploy uses a plugin architecture. You can register custom exporters and serving backends:

import anydeploy
from anydeploy.export.base import BaseExporter

class CoreMLExporter(BaseExporter):
    format_name = "coreml"

    def export(self, model, output_path, config=None):
        # Your export logic
        ...

    def validate_model(self, model):
        return True

anydeploy.register_exporter("coreml", CoreMLExporter)

See CONTRIBUTING.md for details on adding new exporters and backends.

Local-First / Edge AI

This package is designed for edge and local deployment. All export formats (ONNX, TFLite, TorchScript) produce models that run completely offline. The llama.cpp backend enables local LLM serving without any cloud dependencies.

# Export for edge deployment
anydeploy export model.pt --format onnx       # ONNX Runtime (CPU/GPU/edge)
anydeploy export model.pt --format tflite      # TFLite (mobile/edge)

# Serve an LLM locally
anydeploy serve model.gguf --backend llamacpp

Contributing

Contributions are welcome! Please see CONTRIBUTING.md for guidelines.

License

MIT License. See LICENSE for details.

Links

• NRL.ai
• GitHub
• PyPI