Hi, What optimizations would you recommend for increasing model download speeds from Hugging Face?
Hi,
There are multiple optimizations depending on the model, some of the quick and easy changes:
-
set
HF_HOMEto your SFS (Shared Filesystem) or NVMe drive (read about Cache management: https://huggingface.co/docs/datasets/en/cache) -
use
HF_HUB_ENABLE_HF_TRANSFER=1
Use hf_transfer, more on their docs site: https://huggingface.co/docs/hub/en/models-downloading
HF transfer helps usually if you have 100s of 4-9 gb files. Sometimes in HF there are these huge 45 gb files (example Qwen Image, Wan 2.2) They make the downloads quite slow.
Most good llm models have shards of size 9gb so can make excellent use of HF_TRANSFER
For reference