Back to Portfolio
Open SourceInfrastructure

PyTorch Inference Framework

Production-ready inference with TensorRT acceleration. Handles batching, model versioning, and GPU memory management. Built under Kolosal AI.

What is it?

Running PyTorch models in production is messy: GPU memory leaks, no batching, manual versioning, and no API layer. This framework wraps all of that into a clean, deployable service.

Built as part of the Kolosal AI open-source toolchain, it compiles models with TensorRT for maximum GPU throughput, exposes them via a FastAPI endpoint, and ships in Docker so deployment is a single command.

What it does

  • TensorRT Acceleration — Compiles PyTorch models to TensorRT engines for maximum GPU throughput with minimal latency.
  • Dynamic Batching — Automatically groups incoming requests into optimal batch sizes to saturate GPU compute.
  • Model Versioning — Register and serve multiple model versions simultaneously, with instant rollback support.
  • GPU Memory Management — Pools and recycles GPU memory across requests to prevent OOM errors under load.
  • FastAPI Endpoints — Auto-generated REST API with async support, health checks, and OpenAPI docs.
  • Docker-ready — Ships as a Docker image. Deploy to any GPU-enabled environment with a single command.

Built with

PyTorchTensorRTFastAPIDockerCUDAPython

See the code

Full source, docs, and usage examples on GitHub.