Docs / Reference
Environment variables
Configure caches, transport behavior, native execution, sampling, sessions, and Hugging Face progress.
Complete public inventory
| Variable | Effect |
|---|---|
FERMION_CACHE_DIR | Move model downloads made by the fermion CLI. |
HF_HOME | Move the complete Hugging Face cache. |
FERMION_KEEP_TRANSPORT=1 | Keep the coded transport after verified local expansion. |
FERMION_DOWNLOAD_ALL=1 | Download the entire model repository without filtering. |
FERMION_NO_TV4Z=1 | Skip coded transport and download the raw container. |
FERMION_TV4Z=/path | Use an explicit transport decoder. |
FERMION_NATIVE_BIN=/path | Use an explicit native runner. |
FERMION_THREADS=N | Set the native decode thread count. |
FERMION_MAX_THREADS=N | Raise the default hard thread ceiling of 16. |
FERMION_SEED=N | Pin the native seed for sampled decoding. |
FERMION_SESSION=0 | Disable the resident session runner. |
FERMION_SESSION_CTX=N | Override the resident session KV capacity. |
HF_HUB_DISABLE_PROGRESS_BARS | Honor the Hugging Face progress-bar preference. |
Example configuration
shell
export FERMION_CACHE_DIR=/mnt/models/fermion
export FERMION_THREADS=12
export FERMION_SESSION_CTX=16384
fermion serve