Configuration

From the MirbadLLM repository

This page is copied from docs/configuration.md when the site is built.

MirbadLLM reads standard ASP.NET Core configuration: appsettings.json, then environment variables, then command-line arguments. The master’s settings live in the MirbadLLM section. The same key has three spellings:

Where	Spelling	Example
appsettings.json	nested JSON	{ "MirbadLLM": { "Backup": { "Keep": 30 } } }
Environment variable	__ between levels	MirbadLLM__Backup__Keep=30
Documentation	: between levels	MirbadLLM:Backup:Keep

The container deployment sets the environment variables from deploy/.env through the compose files, so most operators only edit that file.

deploy/.env

Copy deploy/.env.example to deploy/.env and adjust it. The compose files read these variables:

Variable	Default	What it does
DOCKER_GID	none, required	Group id of /var/run/docker.sock (stat -c %g /var/run/docker.sock); the agent container joins this group to start model containers
MIRBADLLM_PORT	8080	Host port of the UI, /api, /v1, /metrics and /healthz
MIRBADLLM_ENROLL_PORT	8443	Host port of the enroll listener that joining nodes use (multi-node compose file)
MIRBADLLM_DATA	/srv/mirbadllm/data	Master state: the registry, API keys, nodes, the CA, history. Owned by uid 1654
MIRBADLLM_MODELS	/srv/mirbadllm/models	Checkpoints root, mounted at /models in the agent and the model containers
MIRBADLLM_CACHE	/srv/mirbadllm/cache	Download, compile and autotune caches
MIRBADLLM_BACKUPS	/srv/mirbadllm/backups	Scheduled backups, mounted at /backups. Owned by uid 1654
MIRBADLLM_MEMORY_RESERVE_GIB	8	Memory kept free for the operating system and other workloads when admitting a model
MIRBADLLM_VLLM_IMAGE	a locally built image name	vLLM image for models without their own image; the agent’s Hugging Face download helper runs in it too. Set it to an image the machine has
MIRBADLLM_API_KEY	empty	Optional static bearer key for /v1. Prefer keys created on the Access page
MIRBADLLM_REQUIRE_API_KEY	false	true = /v1 requires a key even before the first key exists
MIRBADLLM_ADMIN_PASSWORD	empty	Admin password for the UI and /api. Empty disables the login, and every page shows a warning
MIRBADLLM_BACKUP_PASSPHRASE	empty	Encrypts the secrets section of scheduled backups. Empty = scheduled backups carry no secrets. Keep a copy elsewhere: a restore needs it
MIRBADLLM_BACKUP_HOSTNAME	dgx-spark	Host name written into backup file names

Master settings

The settings below are the ones operators change most. Each class in the code documents the rest next to its property.

Key	Default	What it does
MirbadLLM:VllmImage	a locally built image name	Default vLLM image (see MIRBADLLM_VLLM_IMAGE)
MirbadLLM:MemoryReserveGiB	8	Seed for each node’s memory reserve on first boot
MirbadLLM:AdminPassword	empty	Admin login password
MirbadLLM:RequireApiKey	false	Require an API key on /v1 from the start
MirbadLLM:AutoStartOnRequest	false	Start a stopped model when a request names it; the request waits until the model is ready or times out
MirbadLLM:StartupTimeout	00:10:00	How long a start may take before the model is marked Failed
MirbadLLM:CrashRestart	true	Restart a model whose process or container died on its own
MirbadLLM:CrashRestartLimit	3	Consecutive crashes after which automatic restarts pause
MirbadLLM:MtpArchitectures	empty	Model architectures the vLLM image can run with MTP speculative decoding; the tuner offers MTP only for these or for a saved MTP configuration
MirbadLLM:Backup:DailyAt	03:00	Time of the daily scheduled backup, in the container’s local time (UTC unless you set TZ)
MirbadLLM:Backup:Keep	14	Number of scheduled backup files kept
MirbadLLM:Backup:IncludeCa	true	Put the master’s CA into the encrypted section of a backup
MirbadLLM:NodeApi:EnrollUrl	the browser’s host on port 8443	Enroll listener URL written into join commands; set it when nodes reach the master by another name or port
MirbadLLM:NodeApi:AgentBundleUrl	the address the browser used	Base URL that join commands download agent bundles from

Node agent settings

Each agent reads its own agent.json (section Agent), written by the installer. The examples in deploy/agent/ show every field the installers set:

• deploy/agent/agent.dgx.json: the agent container on a DGX Spark-style unified-memory host
• deploy/agent/agent.linux.example.json: a Linux PC with discrete NVIDIA GPUs (systemd service)
• deploy/agent/agent.windows.example.json: a Windows PC (Windows service, llama.cpp processes)
• deploy/agent/agent.macos.example.json: an Apple Silicon Mac (LaunchDaemon, llama.cpp and MLX)

In the agent container, environment variables with the prefix MIRBADLLM_AGENT_ override agent.json, for example MIRBADLLM_AGENT_Agent__MemoryReserveGiB=8. The operator guides explain the node settings that the master manages from the node’s page, such as the memory reserve and approved images.

Development settings

src/MirbadLLM.Web/appsettings.Development.json puts the master’s data in ./.data and points the enroll listener and the bootstrap agent at loopback. Two switches exist only for development, and both refuse to start outside the Development environment: the agent’s DevFakeRuntime (fake models, no Docker or GPU) and the master’s MirbadLLM:DevJoinTokens. See setup.md.