nvidia-nemo/automodel is a PyTorch distributed training library that lets you train Hugging Face models without format conversion. Its targets are language models and vision-language models (LLM·VLM), and the repository description states that Hugging Face support is included by default in a PyTorch Distributed native implementation. The latest release is v0.6.0, and the repository was created on May 21, 2025. The language is Python, and the license is Apache-2.0.
The README describes this library as a DTensor-native SPMD library for large-scale training. SPMD is an approach that divides the same work across multiple devices with a single program, and DTensor is PyTorch's approach to handling tensors sharded across multiple devices. It can train Hugging Face LLMs and VLMs without format conversion, and includes recipes and examples for LLMs, VLMs, diffusion, and retrieval. It also includes FSDP2 (data parallelism that shards parameters across devices), pipeline and context parallelism, and components related to LoRA, QLoRA, and quantization. The fine-tuning recipes cover LoRA and full fine-tuning, FSDP, and expert parallelism (a form of parallelism that distributes operations across multiple devices). It has 1,031 stars.
That said, the evidence is limited to the repository description and the README's own claims. There are no performance figures, no comparison with other distributed training tools, and no real-world adoption cases in the given information. There are also no release note entries for v0.6.0, so what this version changed cannot be confirmed. In the accompanying evaluation, novelty, technical, and quality scores were not computed except for Momentum 2.5. What is worth expecting now is closer to a pipeline reference with a well-organized training configuration than to any specific performance advantage.