runner

1# ==============================================================================
2# POST-SESSION PROCESSING - diarization, chronicle, state
3# ==============================================================================
4
5# Only used to download gated pyannote weights once. Inference is entirely
6# local; nothing about a session ever leaves the network.
7HF_TOKEN={{ vault_runner.huggingface_token }}
8HF_HOME=/models
9
10# LiteLLM fronts Ollama with an OpenAI-compatible API.
11DND_LLM_BASE_URL=http://{{ ansible_default_ipv4.address }}:{{ llm_stack_litellm_port }}/v1
12DND_LLM_API_KEY={{ vault_runner.lite_llm_master_key }}
13DND_LLM_MODEL={{ dnd_chronicle_model }}
14
15DND_CAMPAIGN_DIR=/media/dnd/campaign
16DND_TRANSCRIPT_DIR=/media/dnd/transcripts
17DND_RECORDING_DIR=/media/dnd/recordings
18
19# cuda when the driver matches torch's build, cpu otherwise. CPU still runs
20# faster than realtime, so it is a genuine fallback rather than a failure.
21DND_DIARIZE_DEVICE={{ dnd_diarize_device }}
22