Triton-hf-builder: Generate NVIDIA Triton repositories from Hugging Face checkpoints

Hi everyone! I’m sharing triton-hf-builder, an open-source Python library and CLI that generates NVIDIA Triton model repositories from local Hugging Face checkpoints.

It creates model configuration, runtime adapters and dependency files for Transformers, vLLM and TensorRT-LLM. Repository generation runs without a GPU and does not load model weights.

The project includes a Docker quickstart that downloads a small checkpoint, starts CPU serving and sends an inference request.

pip install triton-hf-builder

Currently in beta, licensed under Apache-2.0. Model and hardware compatibility depends on the selected serving backend.

Feedback, bug reports and experiences with different model architectures are welcome!