hf-mem 0.0.1

A CLI to estimate inference memory requirements for Hugging Face models, written in Python

hf-mem

[!WARNING] hf-mem works, but is still experimental. Ideally there should be a minor release anytime soon making things a bit more stable (and aesthetic).

hf-mem is a CLI to estimate inference memory requirements for Hugging Face models, written in Python. hf-mem is lightweight and free for third-party dependencies. It's recommended to run with uv for a better experience.

hf-mem lets you estimate the inference requirements to run any model from the Hugging Face Hub, including Transformers, Diffusers and Sentence Transformers models, as well as any model that contains Safetensors compatible weights.

Usage

uvx hf-mem --model-id HuggingFaceTB/SmolLM3-3B

References

• Safetensors Metadata parsing