The model download commands provided in nano-vllm’s documentation and example scripts rely on the huggingface-cli download --resume-download syntax, which is no longer functional due to breaking changes in the Hugging Face Hub CLI (introduced in huggingface-hub v0.20+). In newer versions of the Hugging Face Hub library, the legacy huggingface-cli download command has been replaced by a simplified hf download command, and the --resume-download flag has been entirely removed—resume functionality for interrupted downloads is now enabled by default, with --force-download serving as the new flag for forcing a full re-download (opposite logic to the old --resume-download). This mismatch causes users running the latest huggingface-hub versions to encounter a "No such option: --resume-download" error when executing the provided download commands, creating unnecessary setup friction for nano-vllm.
To reproduce the issue, install the latest huggingface-hub via pip install --upgrade huggingface-hub, then run either the legacy command huggingface-cli download --resume-download Qwen/Qwen3-0.6B --local-dir ~/huggingface/Qwen3-0.6B/ --local-dir-use-symlinks False or the new simplified command hf download --resume-download Qwen/Qwen3-0.6B --local-dir ~/huggingface/Qwen3-0.6B/ --local-dir-use-symlinks False—both will throw the error noting that --resume-download is not a valid option, with only --force-download and --no-force-download available as alternatives. My environment includes WSL Ubuntu 22.04, Python 3.10.x, huggingface-hub v0.25.1 (new) / v0.19.4 (legacy), and nano-vllm v0.2.0.
The expected behavior is for nano-vllm’s documentation to reflect the current Hugging Face CLI syntax, ensuring users can download models without errors during setup. To resolve this, the documentation and examples should be updated to use the valid new command syntax: hf download Qwen/Qwen3-0.6B --local-dir ~/huggingface/Qwen3-0.6B/ --local-dir-use-symlinks False (resume is default, so no additional flags are needed). A compatibility note should also be added to clarify the difference between legacy and new CLI usage: for huggingface-hub <0.20, the old huggingface-cli download --resume-download command remains valid, while for huggingface-hub ≥0.20, users should use hf download (with --force-download to re-download files if needed). Optionally, a helper function could be added to the codebase to auto-detect the huggingface-hub version and generate the correct download command automatically. Reference for the updated Hugging Face Hub CLI syntax can be found in the official docs at https://huggingface.co/docs/huggingface-hub/main/en/cli#download.
The model download commands provided in nano-vllm’s documentation and example scripts rely on the
huggingface-cli download --resume-downloadsyntax, which is no longer functional due to breaking changes in the Hugging Face Hub CLI (introduced inhuggingface-hubv0.20+). In newer versions of the Hugging Face Hub library, the legacyhuggingface-cli downloadcommand has been replaced by a simplifiedhf downloadcommand, and the--resume-downloadflag has been entirely removed—resume functionality for interrupted downloads is now enabled by default, with--force-downloadserving as the new flag for forcing a full re-download (opposite logic to the old--resume-download). This mismatch causes users running the latesthuggingface-hubversions to encounter a "No such option: --resume-download" error when executing the provided download commands, creating unnecessary setup friction for nano-vllm.To reproduce the issue, install the latest
huggingface-hubviapip install --upgrade huggingface-hub, then run either the legacy commandhuggingface-cli download --resume-download Qwen/Qwen3-0.6B --local-dir ~/huggingface/Qwen3-0.6B/ --local-dir-use-symlinks Falseor the new simplified commandhf download --resume-download Qwen/Qwen3-0.6B --local-dir ~/huggingface/Qwen3-0.6B/ --local-dir-use-symlinks False—both will throw the error noting that--resume-downloadis not a valid option, with only--force-downloadand--no-force-downloadavailable as alternatives. My environment includes WSL Ubuntu 22.04, Python 3.10.x,huggingface-hubv0.25.1 (new) / v0.19.4 (legacy), and nano-vllm v0.2.0.The expected behavior is for nano-vllm’s documentation to reflect the current Hugging Face CLI syntax, ensuring users can download models without errors during setup. To resolve this, the documentation and examples should be updated to use the valid new command syntax:
hf download Qwen/Qwen3-0.6B --local-dir ~/huggingface/Qwen3-0.6B/ --local-dir-use-symlinks False(resume is default, so no additional flags are needed). A compatibility note should also be added to clarify the difference between legacy and new CLI usage: forhuggingface-hub <0.20, the oldhuggingface-cli download --resume-downloadcommand remains valid, while forhuggingface-hub ≥0.20, users should usehf download(with--force-downloadto re-download files if needed). Optionally, a helper function could be added to the codebase to auto-detect thehuggingface-hubversion and generate the correct download command automatically. Reference for the updated Hugging Face Hub CLI syntax can be found in the official docs at https://huggingface.co/docs/huggingface-hub/main/en/cli#download.