Forge is an AI-powered, on-demand benchmarking tool. Describe tasks in plain English. Forge generates implementations, executes them in Docker, and visualizes real-world performance.
Everything you need to objectively measure and validate ML implementation performance.
Containers are spun up per library to ensure zero cache pollution or state leakage between benchmark runs.
Native support for Apple Silicon GPU profiling. Compare PyTorch CPU vs MPS backends on local Mac hardware.
Compare your local results directly against industry baselines from MLPerf and DS-1000 datasets.
Forge is designed to be library agnostic. We currently support the big three for Phase 1, with JAX and TinyGrad integration arriving in Phase 2.
Pre-install all ML libraries in an isolated Docker image.
Add Forge to your claude_desktop_config.json to enable natural language benchmarking.
"mcpServers": {
"forge": {
"command": "python",
"args": ["/path/to/forge/mcp-server/server.py"]
}
}
Prompt Forge with your task and compare results instantly.
forge benchmark "compare numpy vs pytorch for 1024x1024 matmul"