dgx-spark-playbooks/nvidia/station-healthcare-agent/assets/.env.example
2026-05-26 18:25:53 +00:00

40 lines
2.3 KiB
Plaintext

# Clinical Intelligence — Environment Variables
# Copy to .env and fill in your values. Never commit .env itself.
# ─── Required ────────────────────────────────────────────────────
# NVIDIA NGC API Key — used for two things:
# 1) Docker login to nvcr.io (pulls the OpenFold3 NIM image).
# Run `make ngc-login` after editing this file, or:
# echo "$NGC_API_KEY" | docker login nvcr.io -u '$oauthtoken' --password-stdin
# 2) Runtime credential injected into the OpenFold3 container.
# Get one at: https://ngc.nvidia.com/setup/api-key
NGC_API_KEY=nvapi-XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
# ─── LLM Model ──────────────────────────────────────────────────
# Ollama model to pull and use for inference (~86 GB on disk)
OLLAMA_MODEL=nemotron-3-super:120b-a12b
# ─── GPU Selection ───────────────────────────────────────────────
# Index of the GPU to dedicate to each container. Find with:
# nvidia-smi --query-gpu=index,name --format=csv,noheader
# On multi-GPU stations (e.g. RTX PRO 6000 + GB300) you MUST set both
# to the GB300 index — the RTX PRO 6000 (98 GB) cannot fit Nemotron 3 Super
# safely, and OpenFold3 crashes when more than one GPU is exposed.
LLM_GPU=0
OPENFOLD_GPU=0
# ─── Service Ports ───────────────────────────────────────────────
# If host Ollama is already bound to 11434, change OLLAMA_PORT to e.g. 11435.
# `make setup` sources this file so the sandbox provider picks up the override.
OLLAMA_PORT=11434
PROXY_PORT=11435
OPENFOLD_PORT=8000
GATEWAY_PORT=18789
# ─── OpenShell Sandbox ───────────────────────────────────────────
SANDBOX_NAME=clinical-sandbox
# ─── FHIR Server ────────────────────────────────────────────────
# Public synthetic data server (no PHI, no auth required)
FHIR_BASE_URL=https://r4.smarthealthit.org