[programming][machine_learning] Ollama environment variables

[LukeShortCloud]

Load more than one model (default is 1):

OLLAMA_MAX_LOADED_MODELS=3

Force Vulkan support (default is 0 for OLLAMA_VULKAN):

OLLAMA_VULKAN=1
# Disable AMD ROCm.
HIP_VISIBLE_DEVICES=-1
ROCR_VISIBLE_DEVICES=-1
# Disable NVIDIA CUDA.
CUDA_VISIBLE_DEVICES=-1

https://github.com/ollama/ollama/issues/13589

Run Ollama with ROCm on unsupported AMD GPUs:

• Although AMD users can use HSA_OVERRIDE_GFX_VERSION, it is complicated to find the correct version. The official Ollama recommendation is to use Vulkan instead.
  • https://www.conroyp.com/articles/running-ollama-ubuntu-unsupported-amd-gpu-performance-guide
  • https://docs.ollama.com/gpu

[LukeShortCloud]

Keep models loaded forever (default is 5m for 5 minutes, can also specify hours such as 24h, or 0 to unload immediately):

OLLAMA_KEEP_ALIVE=-1

https://docs.ollama.com/faq

[LukeShortCloud]

Increase the context length manually (default is 0 for automatic, can specify for 32768 for 32K, etc.):

OLLAMA_CONTEXT_LENGTH=32768

Related: #1209

[LukeShortCloud]

Use OLLAMA_HOST to change the server IP address (default is 127.0.0.1). You can use 0.0.0.0 to listen on all addresses.

https://ollama.readthedocs.io/en/faq/?h=default+context#how-can-i-expose-ollama-on-my-network