Configuration reference#

Every setting llmgw reads from a file, with the type it takes and what it is for.

Generated from the source. The descriptions are the comments on the fields themselves, so this cannot describe a setting that does not exist or miss one that does — and a test fails when it has drifted. Do not edit it; edit the comment and run go generate ./....

Where a setting takes another block, the block is listed under its own heading rather than repeated.

Every key#

Adapter#

name#

string

Name is what clients ask for, like a model name.

source#

string

Source is the adapter: a GGUF file for llama.cpp; for vLLM a PEFT adapter directory (a mirror version ending in "/", a path on the node, or a Hugging Face repository id). Fetched by each node as the model's weights are.

scale#

float64

Scale is how strongly it is applied; zero takes 1.

owner#

string

Owner is the part of the tenancy tree the adapter belongs to — org:ID, team:ID, project:ID, user:ID or key:ID — and only callers in it see it: to anyone else it is a model that does not exist. An adapter trained on an organisation's documents carries some of what they say in its weights. Empty is every caller the model allows.

Admission#

enabled#

bool

Enabled turns the layer on. Off, every request goes straight through, which is the v0.2.0 behaviour.

bands#

[]Band

default_band#

string

band_header#

string

BandHeader lets a client declare its band. Until authentication lands in v1.0.0 this is trusted, so the port must not be exposed to callers who should not be able to promote themselves.

queue_ttl#

time.Duration

queue_capacity_per_band#

int

queue_capacity_total#

int

QueueCapacityTotal bounds everything queued across bands. Reached, a request is not refused while a lower band has any queued: the newest of the lowest band is turned away instead, with a 503 and Retry-After. Zero takes half of all the bands' capacities together.

interval#

time.Duration

AgentSettings#

default#

map[string]any

Default is agent.yaml's keys for every node.

nodes#

map[string]map[string]any

Nodes are agent.yaml's keys for one node each, by node id, over Default.

Auth#

enabled#

bool

store#

string

Store is the SQLite file holding keys, tenancy and usage.

cache_ttl#

time.Duration

CacheTTL bounds how long a revoked key keeps working, traded against a database read per key per TTL.

negative_ttl#

time.Duration

NegativeTTL is how long a rejected key is remembered as rejected, so a client looping with a bad key cannot become a database read per request.

default_rpm#

int

DefaultRPM and DefaultTPM apply to a key that no scope gives a limit to.

Unlimited used to be the default, and a project created without limits is the easy path — the "default" project on this cluster still is one. The first sign of that is the bill or the outage. Zero keeps the old behaviour for anyone who wants it, but it should be a choice rather than an omission. A fallback, not a ceiling: an explicit limit at any scope wins, above or below.

default_tpm#

int

signup_open#

bool

SignupOpen serves the signup form. Accounts it creates are pending until an operator approves them, so this is not a way in — but turning it off means accounts are made with llmgw users create instead.

oidc#

OIDC

OIDC is sign-in through an identity provider.

Takes a OIDC block.

ldap#

LDAP

LDAP is sign-in against a directory, Active Directory above all, and the accounts kept in step with it.

Takes a LDAP block.

scim#

SCIM

SCIM is account provisioning by the identity provider.

Takes a SCIM block.

max_key_ttl#

time.Duration

MaxKeyTTL is how long any API key may live; a key asked for with no expiry is given this. Organizations, teams and projects can set shorter ones (llmgw keys ttl). Zero is no limit.

personal#

Personal

Personal is what a person may do on their own, outside any project: make personal keys and use the chat page personally, within these limits.

Takes a Personal block.

pricing#

map[string]Price

Pricing is what each model costs, for budgets. A model absent from the table costs whatever DefaultPrice says, which is nothing unless set.

default_price#

Price

Takes a config.Price block.

flush_interval#

time.Duration

FlushInterval and BatchSize decide when usage records are written.

batch_size#

int

buffer_size#

int

BufferSize bounds the queue. Beyond it records are dropped rather than blocking a request: losing accounting is bad, turning a stalled database into an outage is worse.

usage_retention#

time.Duration

UsageRetention is how long the record of each request is kept: who sent it, which replica answered, how long it queued — what ctl request reads. Zero keeps it for ever. Totals, charts and budgets are read from five-minute sums kept beside the records, which are not pruned, so this bounds the table that grows by a row a request without changing any figure a page shows. At least a day when set.

Batch#

max_file_bytes#

int64

MaxFileBytes bounds one uploaded file; 100 MB unless set.

max_lines#

int

MaxLines bounds the requests in one batch; 50 000 unless set.

concurrency#

int

Concurrency is how many of one batch's requests are in flight at once; 4 unless set. They queue in the lowest admission band, behind every interactive request.

retain_days#

int

RetainDays is how long a batch's answer and error files are kept; 30 unless set. Uploaded files are kept until deleted.

Cache#

max_bytes#

int64

MaxBytes bounds every project's answers together; 128 MB unless set. Past it, the answers used longest ago go first.

max_entry_bytes#

int64

MaxEntryBytes bounds one answer; a longer one is not kept. 1 MB unless set.

embedding_model#

string

EmbeddingModel embeds the questions of projects that reuse answers to similar ones (llmgw keys cache --similar): an embedding model in the catalog. Unset, those projects reuse only the same request's answer. The gateway asks it on its own account, as it asks a judge.

Canary#

model#

string

Model is the candidate: a catalog model, usually this one at newer weights, which clients cannot name themselves.

percent#

int

Percent of requests go to it, 0 to 100.

min_requests#

int

The candidate is taken out — rolled back — when, over at least MinRequests of its own, its error rate is MaxErrorIncrease above this model's (0.05 is five points), or its mean time to the first token is MaxTTFTRatio times this model's. Zero is 50, 0.05 and 1.5.

max_error_increase#

float64

max_ttft_ratio#

float64

Capabilities#

tools#

bool

vision#

bool

embeddings#

bool

images#

bool

Images is a model that makes pictures, asked for at /v1/images/generations. Implied by an external model whose API is ComfyUI's.

meshes#

bool

Meshes is a model that makes 3D models from a picture, asked for as a job at /v1/3d/generations. Implied by a ComfyUI model whose output is meshes.

ComfyUI#

workflow#

string

Workflow is a built-in's name (flux-schnell, the default, or flux-dev) or a JSON file on the control plane, in ComfyUI's API format.

checkpoint#

string

Checkpoint is the file the workflow's {{checkpoint}} names: the source's own name unless set, which a source that is a set of files must.

output#

string

Output is what the workflow makes: pictures (the default), asked for at /v1/images/generations; meshes, asked for as a job at /v1/3d/generations from a picture; or video, audio or files. Every kind is asked for as a job at /v1/workflows/runs too.

inputs#

[]Input

Inputs are what the workflow takes from a request at /v1/workflows/runs, each the placeholder of its name.

nodes#

[]CustomNode

Nodes are custom nodes the workflow needs, put in each replica's own custom_nodes at the commit named.

huggingface_repo#

string

HuggingFaceRepo is the repository a custom node loads the weights by name from (Stable Fast 3D's loader asks for stabilityai/stable-fast-3d): the replica's weights are shown to the node there, in the Hugging Face cache of the engines' user, so nothing is fetched and no token is needed when it runs.

environment#

string

Environment names the models that may share one ComfyUI replica: every model of an environment agrees on the custom nodes they both name, so one replica can carry all of them and ComfyUI moves between their weights itself. Empty is an environment of the model alone.

ComfyUIConfig#

dir#

string

Dir is ComfyUI's directory, where main.py is. Readable by the user engines run as.

python#

string

Python is the interpreter of its environment; Dir's .venv when unset.

Control#

interval#

time.Duration

Interval is how often desired and actual state are compared.

base_backoff#

time.Duration

BaseBackoff and MaxBackoff bound how fast a failing model is retried.

max_backoff#

time.Duration

unreachable_after#

time.Duration

UnreachableAfter and LostAfter stage node expiry. They are deliberately far apart: a network blip must not tear down replicas that still serve, and rebuilding a multi-node replica costs tens of minutes of loading.

lost_after#

time.Duration

port_range#

[]int

PortRange is the per-node range for engine HTTP and dist-init ports.

mgmt_interface#

string

MgmtInterface is the ethernet interface NCCL bootstraps over.

timezone#

string

Timezone is where models' scaling schedules are read, an IANA name like Europe/Istanbul; empty is the control plane's own.

eval_dir#

string

EvalDir is where quality suites are read from by name, for llmgw ctl eval MODEL --suite NAME, a model's eval gate and the console. The suites the binary carries (arithmetic, json-format) are found when it has none of the name; a suite here of the same name replaces one.

enabled#

bool

Enabled runs the control plane even when configuration declares no models.

Without it, a deployment whose catalog lives entirely in the database cannot start: control plane mode is inferred from the config file, so an empty models block means "router only", the process refuses to start for want of anything to serve, and the models it would have loaded are never read. That is a cluster that cannot be managed at runtime because it is managed at runtime. AgentListen cannot carry this signal — it has a default, so every deployment would become a control plane.

lease_ttl#

time.Duration

LeaseTTL is how long this control plane's claim to be the one placing work stands without being renewed.

Only meaningful with a shared store, where a second control plane can see the claim. Long enough to survive a slow query or a brief network stall, short enough that a standby takes over before anyone has finished reading the alert. Zero takes the default, 10s.

CustomNode#

repo#

string

Repo is the repository, https only.

commit#

string

Commit is the full commit the node is used at: a branch moves, and a node that changed under a model is a model that changed unasked.

requirements#

[]string

Requirements are the PyPI packages the node needs, by name and version range only. They are installed beside the replica, never into ComfyUI's environment: a package named here at the range named, and every other one at the version ComfyUI already runs. One named here is first on ComfyUI's path, so its range must suit ComfyUI too.

build#

[]string

Build are directories of the repository that are Python packages to be compiled rather than downloaded: C++ and CUDA extensions no index has (Stable Fast 3D's texture_baker and uv_unwrapper). Each is built as a wheel with ComfyUI's Python for the card the replica runs on, kept on the node for the next replica on the same kind of card, and installed beside the replica like its requirements.

Endpoint#

id#

string

url#

string

api_key#

string

APIKey is sent to the engine as a bearer token, for an engine started with one.

models#

[]string

engine#

string

Engine names the inference engine, which decides what its metrics are called. Empty means vLLM's names.

workflow#

string

Workflow is the ComfyUI workflow for engine: comfyui, a built-in's name or its JSON; flux-schnell unless set.

pool#

string

weight#

float64

priority#

int

EngineOverride#

gpu_class#

string

GPUClass picks nodes with a GPU of this class ("v100-32g").

nodes#

[]string

Nodes picks nodes by id.

labels#

map[string]string

Labels picks nodes carrying all of these.

engines#

map[string]string

Engines are laid over the cluster's: a version, or "" for none.

EvalGate#

suite#

string

Suite is a suite's name in control.eval_dir, or one the binary carries (arithmetic, json-format).

min_score#

float64

MinScore is the score, 0 to 1, to pass at; zero is the suite's own threshold.

judge#

string

Judge grades the suite's rubric cases, instead of the judge it names.

External#

url#

string

URL is the provider's base address, without /v1: https://api.openai.com.

model#

string

Model is what the provider calls it; requests are rewritten to it.

api_key_file#

string

APIKeyFile holds the provider's key, on the control plane. A file rather than the key itself, so the key is in neither the catalog, the store nor the audit log; empty for a provider that asks for none.

api_key_files#

[]ExternalKey

APIKeyFiles are several keys for the provider, in place of APIKeyFile: one endpoint each (external:MODEL#1, #2…), so a key that answers 429 or 401 is passed over while its breaker is open and the request goes to another. A provider's limits are per key.

peer#

string

Peer names the federated llmgw cluster this model is offered by, for a model that federation added (llmgw ctl federation add) rather than one written by hand. Such a model is named PEER/MODEL, comes and goes with the peer's own list, and is sent only while the peer is up.

api#

string

API is what the provider speaks: OpenAI's, unless this says comfyui — a ComfyUI server, asked for pictures at /v1/images/generations and sent the workflow below with the request's prompt in it. Model is then the checkpoint file the workflow loads, as ComfyUI lists it.

workflow#

string

Workflow is the ComfyUI workflow, in its API format: a built-in's name (flux-schnell, the default, or flux-dev) or a JSON file on the control plane.

output#

string

Output is what it makes: pictures, or meshes (see ComfyUI.Output).

ExternalKey#

file#

string

File holds the key, on the control plane, as APIKeyFile does.

weight#

float64

Weight is its share of the requests against the others'; zero is 1.

Facility#

budgets#

[]FacilityBudget

Budgets are the domains with a power cap.

margin#

float64

Margin is the share of every cap kept free, 0.1 for ten percent. Zero keeps none, for caps already given derated.

gpu_watts#

float64

GPUWatts is a card's draw when it does not report its power limit: an agent without nvidia-smi's power fields, or a node that has not sent a heartbeat yet. Zero counts such a card as drawing nothing.

cpu_core_watts#

float64

CPUCoreWatts is the draw of a core reserved by a replica with no GPU (cpu_cores). Zero counts such replicas as drawing nothing.

node_watts#

float64

NodeWatts is what a node draws besides its cards and reserved cores — idle CPUs, memory, fans, disks — counted for every node in a capped domain, busy or not.

measured_fresh#

time.Duration

MeasuredFresh is how old a node's power reading may be and still be counted against its domains. Zero takes two minutes.

FacilityBudget#

domain#

string

Domain is level:name, as the nodes' labels give it: rack:r12-07, pdu:r12-07-a, row:12.

cap_watts#

float64

CapWatts is what the domain may draw. For a dual-fed rack it is what one feed carries alone, derated as the facility derates it (usually 80%), since that is what the rack has to survive.

other_watts#

float64

OtherWatts is drawn in the domain by what llmgw does not place on: switches, storage, machines outside the cluster.

Federation#

name#

string

Name is this cluster as a peer reads it, in the federated_from tag sent with a request to a peer set to receive tags (ctl federation add --send-tags). Empty is this machine's host name.

interval#

time.Duration

Interval is how often each peer is asked whether it is ready, timed, and what it offers. A peer that does not answer is taken out of routing until it does. Zero is 10s.

FirewallSettings#

enabled#

bool

Enabled keeps llmgw's internal ports — engines, node exporter, the agents' stream, the mirror, the store — to the cluster's networks on every machine. False removes llmgw's nftables table everywhere, for a site whose own firewall does that job; the rules are kept, and true puts them back. Changed in the cluster's settings, it is in force on every node and control plane within seconds.

FreeTokenConfig#

binary#

string

Binary is its ft command, in the environment it was installed into. The agent installs one itself on a node whose GPUs it runs on; this is for one installed by hand.

GPUNeed#

min_compute_capability#

float64

MinComputeCapability is the oldest architecture that runs it: 7.5 is Turing, 8.0 Ampere (bf16, flash-attention 2), 8.9 Ada (fp8), 9.0 Hopper.

min_memory_gb#

int

MinMemoryGB is the card's own memory it needs.

GuardModel#

format#

string

Format is qwen3guard (Qwen3Guard-Gen: "Safety: Safe", "Unsafe" or "Controversial", and the categories), llamaguard3 (Meta's Llama Guard 3: "safe", or "unsafe" and the categories), shieldgemma (Google's ShieldGemma: Yes or No), or json ({"safe": false, "categories": ["..."]}).

categories#

[]string

Categories the model is asked about, for shieldgemma — each a policy sentence — and the names reported for llamaguard3's S1, S2, ….

Input#

name#

string

Name is the placeholder's, and the request's key.

type#

string

Type is one of text, int, float, bool, choice, image, mask, audio, video, file. The last five are files, sent as base64, a data URL, or the id of a file uploaded to /v1/files, and put in ComfyUI's input directory; the placeholder gets the name they were given there.

required#

bool

Required inputs must be in every request; the others take Default, or the workflow's own {{name|default}}.

default#

any

min#

*float64

Min and Max bound a number; for text, Max bounds its length.

max#

*float64

choices#

[]string

Choices are what a choice may be.

max_mb#

int

MaxMB bounds a file input; zero is the type's own bound.

doc#

string

Doc says what the input is for, to whoever fills it in.

JudgePolicy#

model#

string

Model is a catalog model with role: guard.

models#

[]string

Models are judges asked in turn, instead of Model: the first that answers gives the verdict. A judge on a GPU first and one on CPUs after it is the GPU's speed while it is there, and the CPUs' when it is not — busy, failed, or the node down.

input#

string

Input and Output are what an unsafe verdict does: flag or block. Empty does not ask the judge about that direction.

output#

string

categories#

[]string

Categories, when set, are the only ones that count: Llama Guard's codes (S1, S10) or the names the model reports.

controversial#

string

Controversial is what Qwen3Guard's third verdict counts as: safe (the default) or unsafe.

latency_budget#

time.Duration

LatencyBudget is how long each judge is waited for before the next is asked, or, after the last, the policy's fail applies. Zero is 250ms.

KVIndex#

tokenize_timeout#

time.Duration

TokenizeTimeout bounds the call that turns a request into token ids, made to one of the model's replicas before the request is placed. Past it the request is routed as tier 1 would. Zero takes 250ms.

min_blocks#

int

MinBlocks is the fewest leading blocks a replica must hold to be preferred for them; fewer is coincidence — a chat template's opening tokens — rather than a prefix worth following. Zero takes 4, which is 64 tokens at vLLM's usual block size of 16.

max_blocks#

int

MaxBlocks bounds how many of a prompt's blocks are hashed and looked up. Beyond it a longer match still helps, but the difference no longer decides anything. Zero takes 4096.

LDAP#

urls#

[]string

URLs are the directory's servers, tried in turn: ldaps:// (636), or ldap:// (389) with StartTLS.

start_tls#

bool

StartTLS upgrades an ldap:// connection before anything is sent. A password never crosses a connection without TLS: ldap:// without it is refused.

ca_file#

string

CAFile is the authority the servers' certificates are checked against, when it is not the system's: an AD Certificate Services root.

ca#

string

CA is the same authority written out, PEM: kept in the cluster's settings, so no file has to be put on every control plane. Used in place of CAFile when both are set.

bind_dn#

string

BindDN and the file holding its password: a service account that may read users and groups. Nothing is written to the directory.

bind_password_file#

string

base_dn#

string

BaseDN is where people are searched for, subtree included.

group_base_dn#

string

GroupBaseDN is where their groups are; empty is BaseDN's domain root.

user_filter#

string

UserFilter narrows who may sign in, ANDed with the login match: a group (memberOf=...) or an OU's people only. Empty is every enabled user object.

label#

string

Label is what the form says to sign in with: "Sign in with your account". Empty is "Active Directory".

create_accounts#

bool

CreateAccounts makes an account, active, for a person the directory vouches for and the gateway does not know. Off, only existing accounts sign in this way.

roles#

map[string][]string

Roles maps a staff role (viewer, operator, billing, admin) to the groups that give it, by name (CN) or whole DN. Set at every sign-in and sync: someone taken out of the group loses the role.

memberships#

map[string]string

Memberships maps a group, by CN or DN, to a place in the tree — org:NAME, team:ORG/TEAM, project:ORG/TEAM/PROJECT — joined at sign-in and kept by the sync, which also takes away a place whose group the person has left.

sync_interval#

time.Duration

SyncInterval is how often the leader reads the directory. Zero is 15m; negative turns the sync off, leaving only sign-in.

timeout#

time.Duration

Timeout bounds each round trip to a server. Zero is 10s.

Landing#

contact#

string

Contact is where someone without a key should go — an address, a URL, a channel name. Rendered as a link when it is one, as text otherwise.

notice#

string

Notice is an optional line of your own, for whatever the reader needs to know that this page cannot guess.

language#

string

Language is what the web portal speaks to someone who has not chosen a language with the switcher at the foot of every page and whose browser asks for none it speaks: "en" or "tr". Empty is English.

docs_sign_in#

bool

DocsSignIn serves the docs portal at /docs only to someone signed in. Off, the portal is public like this page: its client guide is for the person who has no account yet, and the operator documents describe llmgw in general — no address, model or key of this installation — which the project's repository publishes anyway. On, a visitor is sent to sign in, and without accounts the portal needs an API key.

License#

id#

string

ID names the licence for a machine: an SPDX identifier where there is one (MIT, Apache-2.0, CC-BY-NC-4.0), else a lowercase name of the licence's own (stability-community, tencent-hunyuan-community). What a person's or a site's choice of licences matches on. Name is for people.

name#

string

Name is the licence's own: MIT, Apache-2.0.

url#

string

notice#

string

Notice is what must accompany what was made: attribution, terms on the output, what the weights were trained on.

commercial#

*bool

Commercial says whether what is made may be used commercially: nil is not said, which is not the same as yes.

excluded_regions#

[]string

ExcludedRegions are where the licence grants nothing: ISO 3166 codes, and EU for the union (Hunyuan's: EU, UK, KR).

revenue_cap_usd#

int64

RevenueCapUSD is the yearly revenue above which the licence asks for another agreement (Stability's community licence: 1000000). Zero is none.

attribution#

string

Attribution is what a product using the output must show: "Powered by Stability AI".

LlamaCppConfig#

bin#

string

Bin is the llama-server binary. It is the server, not the CLI: the gateway talks OpenAI over HTTP and needs the one that serves it. Empty means the agent installs llama.cpp itself, at the version the control plane asks for. Setting it opts out of that.

parallel#

int

Parallel is how many requests a server handles at once, when a profile does not say. llama.cpp defaults to one, which would make every second request wait behind the first no matter what the router decided.

threads#

int

Threads is how many CPU threads a shard with no GPU computes on. Zero takes half this machine's CPUs.

Not all of them, which is llama.cpp's own default: two busy servers that each spin on every CPU fight for them, and on an 8-vCPU node that made 7B generation fall from 10 tokens a second to under 1 each. At half, one server alone is about a third slower and two together keep 5.5 each. A node that only ever runs one CPU replica does better with all of them; say so here.

Mirror#

api#

string

API is the mirror's management address, scheme included, for example https://models.internal:8082. Empty means no mirror is configured.

The management port, not the one nodes fetch bytes from. Those are two listeners on the mirror precisely so this one can be locked down without re-certifying every agent's fetcher.

peers#

[]string

Peers are the management addresses of mirrors kept in step with this one (llmgw mirror serve --peer): asked when it cannot be reached, and a node given the same file on one of them while it is down.

tls#

tlsconf.Config

TLS authenticates to the mirror's management API, with the same CA the agents use. Empty leaves the connection in plain HTTP, which is only acceptable on a network where nothing untrusted can reach that port — the mirror says as much in its own log when it starts that way.

Takes a tlsconf.Config block.

registry#

string

Registry is the model registry the mirror page browses, empty for the public Hugging Face. Set it to an internal mirror of the registry where the control plane cannot reach the internet — or leave the whole mirror block out, which turns the explorer off rather than giving it a search box that can only time out.

timeout#

time.Duration

Timeout bounds a management call. It does not bound a download: the mirror answers a fetch as soon as it has started one, and the bytes keep arriving long after this deadline has passed.

update_check#

time.Duration

UpdateCheck is how often the registry is asked whether the files of the models served from it have changed there: a repository can publish new weights under the same name. Only told, never applied — moving a model to them is llmgw ctl model update. 0 never asks.

Model#

name#

string

aliases#

[]string

description#

string

Description is what this model is for, in a line.

Written by whoever added it, because nobody else can say it. A registry offers a pipeline tag and a wall of README, and neither answers the question a person picking from a list is asking: which of these nine names should I send this request to.

It reaches clients through /v1/models, which is the only view most of them ever see. A gateway serving nine models by name alone makes that choice by guesswork.

source#

string

Source is the weights path or a HuggingFace id.

allow_pickle#

bool

AllowPickle lets the nodes fetch this model's weights, or its adapters', in a pickle format (.bin, .pt, .pth, .ckpt), which run code as they are loaded. Off, such a source is refused when the model is added and by the node before it fetches anything. For a model that has no safetensors or GGUF form and whose files were checked.

profiles#

[]Profile

scaling#

Scaling

Takes a Scaling block.

stopped#

*Scaling

Stopped is the scaling a model had when an administrator stopped it: kept in the catalog with none running (Scaling 0/0) until started again with this. Nil for a model not stopped.

Takes a Scaling block.

class#

string

Class is how much this model's capacity is worth against another's, when there is not enough for both: critical, standard or opportunistic. Empty is standard, or opportunistic for a model loaded on demand. See EffectiveClass.

owner#

string

Owner is the part of the tenancy tree whose capacity quota this model counts against, written org:ID, team:ID or project:ID. Empty is shared: held by the cluster, counted against nobody.

external#

*External

External is a model served by someone else — an OpenAI-compatible API outside the cluster — rather than by replicas placed here. Such a model has no profiles and nothing is placed for it; the router sends its requests to the provider. Most useful named in another model's fallbacks: used only when the cluster has no capacity left. Its prompts leave the cluster, so which keys may name it is worth an allowlist.

Takes a External block.

comfyui#

*ComfyUI

ComfyUI is how a model whose profiles run engine comfyui makes its pictures: the workflow, and the file it loads.

Takes a ComfyUI block.

license#

*License

License is the terms of what the model makes, for a model whose output somebody ships: shown with every result and written into a mesh.

Takes a License block.

adapters#

[]Adapter

Adapters are LoRA adapters served on this model's replicas, each asked for by its own name as though it were a model. Every replica loads every adapter, applying one only to a request that names it — unless MaxLoadedAdapters is set.

max_loaded_adapters#

int

MaxLoadedAdapters, above zero, has each replica hold at most this many adapters at once, loaded when a request first asks for one and unloaded least recently used first to make room. The router sends a request for an adapter to a replica that already holds it while that one has room to take it. vLLM profiles only: llama-server loads adapters at start and cannot load one afterwards. Zero loads every adapter on every replica at start.

fallbacks#

[]string

canary#

*Canary

Canary sends a share of this model's requests to another catalog model — the same model at new weights — to compare the two before moving to it (llmgw ctl model canary).

Takes a Canary block.

eval#

*EvalGate

Eval is a quality suite new weights must pass before they replace this model's: llmgw ctl model update and model promote run it against them first, and stop below its score.

Takes a EvalGate block.

routing_profile#

string

RoutingProfile selects how the router scores endpoints for this model.

slo#

SLO

Takes a SLO block.

context_length#

int

ContextLength is how many tokens one request may use, prompt and completion together, and MaxOutput bounds the completion alone. Zero means unknown, which is what every model was until now.

The gateway knew neither, so an over-long prompt was not refused: it was admitted, given a replica, and failed at the engine, and a client sizing its prompts from /v1/models had nothing to read. Worse, the number is not the one an operator writes in the engine's arguments — llama.cpp divides --ctx-size across its parallel slots, so a pool of 8192 across four slots is 2048 a request, and nothing said so.

max_output#

int

capabilities#

Capabilities

Capabilities is what this model can do beyond plain completion.

Takes a Capabilities block.

query_prefix#

string

QueryPrefix is put before a question an embedding model embeds for a knowledge-base search; documents are embedded without it. Models trained with an instruction on the query side want it, Qwen3-Embedding "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: " for one.

role#

string

Role "guard" is a model that judges the gateway's traffic for the guardrails rather than serving clients: hidden from /v1/models and the chat page, refused to any key that names it, critical by default, and always running. Guard says how its answers are read.

guard#

*GuardModel

Takes a GuardModel block.

origin#

*Origin

Origin is the registry repository the weights came from, at which commit, and each file's checksum there. Written when a model is served from the registry; for one that was not, worked out from the mirror's record of where it fetched the file. It is what lets the control plane say that the repository has published different weights since, and move the model to them.

Takes a Origin block.

Monitoring#

prometheus#

string

Prometheus is the server's address, for example http://monitoring.internal:9090. Empty leaves the charts off the dashboard rather than showing an error where one should be.

window#

time.Duration

Window is how far back the charts look. Zero takes an hour.

timeout#

time.Duration

Timeout bounds one query. It runs while somebody waits for a page, so a monitoring system that is slow today is skipped rather than waited for.

OIDC#

issuer#

string

Issuer is the provider's address; its /.well-known/openid-configuration says the rest.

client_id#

string

ClientID and the file holding its secret, as registered with the provider.

client_secret_file#

string

redirect_url#

string

RedirectURL is where the provider sends people back: https:///login/oidc/callback. Empty works it out from the request, which is right unless a proxy changes the address.

label#

string

Label is what the sign-in button says: "Sign in with ".

scopes#

[]string

Scopes asked for besides openid; empty is email and profile, and groups when Roles or Memberships use them.

groups_claim#

string

GroupsClaim names the ID token claim holding groups; empty is "groups".

create_accounts#

bool

CreateAccounts makes an account, active, for a person the provider vouches for and the gateway does not know. Off, only existing accounts sign in this way.

roles#

map[string][]string

Roles maps a staff role (viewer, operator, billing, admin) to the groups that give it. Set at every sign-in: someone taken out of the group loses the role the next time.

memberships#

map[string]string

Memberships maps a group to a place in the tree — org:NAME, team:ORG/TEAM, project:ORG/TEAM/PROJECT — joined at sign-in.

no_passwords#

bool

PasswordSignIn off refuses passwords on the sign-in form: people come in through the provider only. Administrators still can, so a broken provider does not lock everybody out.

Offline#

mirror#

string

Mirror is the mirror's file address, as nodes reach it: http://10.0.0.5:8081 — the port model files are served on, not the management API. Set, the cluster is offline: nodes install engines from it and never from their publishers, releases come from it unless releases names another place, and no registry is asked whether a model has newer weights. Python engines (vLLM, SGLang) come from PyPI, which no bundle carries; on an offline node they are installed by hand.

checksums#

string

Checksums is the file of engine checksums this control plane trusts, kept by llmgw bundle import and llmgw bundle trust; every engine file a node fetches from the mirror is checked against it. Read again whenever it changes. Empty is /etc/llmgw/offline.sha256.

OnDemandSettings#

evict_after_idle#

time.Duration

EvictAfterIdle is how long an on-demand replica must go unasked-for before a model that wants its card may have it. Zero keeps 2m.

min_lifetime#

time.Duration

MinLifetime is how long a replica that has just come up is safe from that, so two models do not take turns paying for loads. Zero keeps 5m. On one card shared by a picture model and a 3D model, 1m each brings the wait between a picture and its mesh to about a minute. Both are in force on the leader within seconds.

Options#

nvidia_smi#

string

NvidiaSMI is the path to nvidia-smi; empty means look it up on PATH. Worth configuring on nodes where the driver lives outside the default paths, which containerized driver installs often do.

ib_root#

string

IBRoot is the sysfs InfiniBand directory.

prefer_nvml#

bool

PreferNVML uses the NVML backend when this binary has one. NVLink topology is read from nvidia-smi topo -m on every build; NVML, where it is used, is the more exact source of the peers.

skip_gpus#

bool

SkipGPUs omits GPU discovery. A node without nvidia-smi is already found to have none, so this is for one whose driver is broken.

dcgm_exporter#

string

DCGMExporter is the dcgm-exporter metrics URL read for how much of each card's chip is in use, on cards NVML's GPM does not cover (before Hopper). Empty is http://127.0.0.1:9400/metrics; "off" reads none. A node without one costs a refused connection every five minutes.

Origin#

repo#

string

Repo is the repository, owner/name.

revision#

string

Revision is the commit the files were fetched at.

files#

[]OriginFile

Files are the weights fetched from it, with what the registry published as each one's sha256. A new commit that leaves them alone — a README edited — is not an update.

previous#

*PreviousOrigin

Previous is what the last update replaced, kept so it can be gone back to while its files are still on the mirror.

Takes a PreviousOrigin block.

OriginFile#

path#

string

sha256#

string

bytes#

int64

Pattern#

name#

string

regex#

string

action#

string

Personal#

disabled#

bool

Disabled refuses personal keys and personal chat; people then use the gateway only through projects they are members of.

rpm#

int

tpm#

int

budget#

float64

Budget is in whole units per Period ("day" or "month").

period#

string

Policy#

mode#

string

Mode: enforce acts, shadow only records what it would have done, off does nothing. Empty is enforce.

fail#

string

Fail: what happens when a policy cannot judge — a judge model down or over its latency budget. open serves, closed refuses. Empty is open.

no_store#

bool

NoStore keeps nothing of the request: no stored conversation, no memory notes, no prefix routing, no fallback outside the cluster.

input#

Rules

Takes a Rules block.

output#

Rules

Takes a Rules block.

tools#

Tools

Takes a Tools block.

max_output_tokens#

int

MaxOutputTokens is put in a request that asks for no limit; one asking for more is lowered to it.

system_prompt#

string

SystemPrompt is put before every request's messages.

drop_client_system#

bool

DropClientSystem removes the caller's own system and developer messages, so only the pinned prompt instructs the model.

require_capabilities#

bool

RequireCapabilities refuses a request offering tools, or sending images, to a model whose catalog entry does not say it takes them.

validate_structured#

bool

ValidateStructured checks an answer against the JSON schema the request asked for in response_format: one that does not fit is stopped — or, streamed a sentence at a time, flagged, having been sent already.

judge#

*JudgePolicy

Judge has a judge model on the cluster read the text too.

Takes a JudgePolicy block.

services#

[]Service

Services are outside guardrails asked over HTTP, beside the judge.

stream#

string

Stream: sentence scans a streamed answer a sentence at a time, releasing each once it passes; buffer_all holds the whole answer until it is judged. Empty is sentence.

Prefix#

enabled#

bool

max_nodes#

int

MaxNodes bounds the tree, MinMatch is the shortest match worth acting on in bytes, and TTL is how long an entry is believed — an engine evicts on its own schedule, so an old record may describe a cache that is gone.

min_match#

int

ttl#

time.Duration

match_ratio#

float64

MatchRatio is the share of the prompt that must match to be worth acting on. BalanceAbs and BalanceRel are the imbalance guard, and both must trip before affinity is abandoned.

balance_abs#

float64

balance_rel#

float64

PresetSettings#

auto#

bool

Auto has the leading control plane add a picture model (flux-schnell) and a 3D model (triposr) by itself, on demand, once a node has a card that fits one and the catalog has no model that makes the same kind of thing — so a new cluster's portal has its pictures and 3D pages without anybody writing a definition. Only where the cluster already installs ComfyUI (engines.comfyui). A preset somebody removed is not added again by itself. In force on the leader within a minute.

PreviousOrigin#

source#

string

revision#

string

files#

[]OriginFile

Profile#

name#

string

engine#

string

engine_version#

string

EngineVersion restricts the profile to nodes whose installed engine's version starts with it: "0.11" takes vLLM 0.11.0 and 0.11.1, not 0.10.2. Empty takes whatever version the node has.

gpu_class#

string

source#

string

Source is the weights this profile loads, when not the model's: an embedding model served by vLLM from its repository on a card and by llama.cpp from a GGUF on CPUs is one model with two sources. Empty loads the model's. Updates and canaries move the model's source; a profile with its own keeps it.

tp#

int

pp#

int

gpus_per_node#

int

GPUsPerNode defaults to TP for a single-node profile.

nodes#

int

Nodes above 1 makes this a multi-node profile, which requires a fabric.

args#

[]string

env#

map[string]string

min_memory_gb#

int

MinMemoryGB is the host RAM this profile needs to run.

The number that matters is the model's working set: the weights, plus the KV cache for the whole context pool, plus the engine's own overhead. For a GPU profile that lives in the cards and this is small; for a CPU profile it is the whole thing, and getting it wrong means an OOM on a node that reported itself healthy.

Zero means unconstrained, which is every profile that existed before this. A node that does not report its memory is not excluded either — an older agent must not have its nodes quietly emptied by an upgrade of the control plane.

cpu_cores#

int

CPUCores is the cores a replica of a profile with no GPU is given: reserved out of its node's CPUs by placement, as memory is, and what its engine computes on (llama.cpp's --threads).

Reserved because an engine takes every core it is told about, and two on one node that each think they have all of them spin against each other: on an 8-vCPU node, two 7B replicas fell from 10 tokens a second to under one each. Zero reserves nothing and leaves the thread count to the node, which is every profile from before this.

min_disk_gb#

int

MinDiskGB is the free space a node needs before it can hold these weights.

Only checked when the node does not already have them: a node holding the file needs no room for it, and requiring the space twice would keep a model off the one node best placed to run it.

Declared rather than derived, because the size of a model that has never been fetched anywhere is not knowable from here — and a wrong guess fails late, halfway through a download, which is the failure this exists to prevent. Zero means unconstrained, and a node that reports no storage is not excluded: unknown is not full.

node_selector#

map[string]string

NodeSelector restricts this profile to nodes carrying every one of these labels, with these values.

A CPU profile otherwise fits everywhere, because there is no card to be free — which is right for a 0.5B model and wrong for a 14B one. Nothing in the inventory says how much memory a node has, so the operator says which nodes are meant for a model instead of the scheduler inferring it. Agents have reported labels since the beginning; this is what reads them.

power_watts#

float64

PowerWatts is what one replica of this profile draws, in watts, split evenly over its nodes, counted against the power cap of every rack, PDU or other domain its nodes are in (facility.budgets in server.yaml). Zero derives it: each card at the power limit it reports, each reserved core at facility.cpu_core_watts. Only matters where a domain has a cap.

keep_within#

string

KeepWithin keeps a multi-node replica's nodes inside one rack, row, room or site, read from their llmgw.io/rack (and so on) labels: for a fabric whose leaves do not reach across racks, or cables that do not. Nodes without the label are passed over for such a profile. Empty is no constraint.

weight#

float64

Weight is the router's capacity weight. It should come from ctl bench rather than a guess (plan §6.2.1).

priority#

int

Priority orders profiles during placement; highest that fits wins.

replicas#

int

Replicas, set on any of a model's profiles, is how many of its replicas run on this one: seven on cards and four on CPUs, all answering under the model's name, rather than each replica on the best profile that fits. The model's scaling is then the sum, min and max alike; a profile left at zero runs none.

startup_timeout#

time.Duration

StartupTimeout bounds model loading. A 70B model across two nodes takes tens of minutes, so this is per profile rather than global.

kv_events#

bool

KVEvents has vLLM publish what its prefix cache holds, and the router route this model's requests by it: to the replica holding the most of a prompt's leading blocks (plan §7.5 tier 2). vLLM only.

Off by default, because it costs something on every request: the router asks a replica to tokenize the prompt before placing it, a round trip and the chat template run twice. It pays where prompts share long prefixes that the approximate tree guesses wrong — many replicas, conversations that move between them, evictions. The engine is started with sha256_cbor block hashing and a fixed PYTHONHASHSEED, which the router's hashing assumes; arguments overriding either leave the index empty, and the agent says so in its log.

gpu#

*GPUNeed

GPU says what card the profile needs, in place of naming its class: it stands for every class in the cluster that meets it. A catalog entry written once then places sensibly on any site's cards.

Takes a GPUNeed block.

comfyui#

*ProfileComfyUI

ComfyUI is this profile's own workflow, checkpoint and nodes, over the model's: an fp8 workflow where the card has fp8, fp16 where it has not.

Takes a ProfileComfyUI block.

ProfileComfyUI#

workflow#

string

checkpoint#

string

nodes#

[]CustomNode

Router#

activation_wait#

time.Duration

ActivationWait is how long a request waits for an on-demand model to load before the client is told to come back. Zero takes the default.

The one real decision in on-demand loading. Refusing outright makes the feature useless, since the first request for every model would fail; waiting without a bound turns a request into a job, and a client that budgeted thirty seconds ends up holding a connection for ten minutes.

scrape_interval#

time.Duration

scrape_timeout#

time.Duration

stale_after#

time.Duration

StaleAfter is how long an endpoint may go unscraped before being scored down (plan §7.6).

retries#

int

Retries is how many additional endpoints may be tried after a failure, and only while nothing has been written to the client (plan §7.7).

cooldown#

time.Duration

Cooldown is how long a failing endpoint leaves the candidate pool.

failure_threshold#

int

FailureThreshold is the consecutive failure count that opens a breaker. 1 turns the breaker into a pure cooldown.

default_profile#

string

DefaultProfile names the routing profile used when a model picks none.

prefix#

Prefix

Prefix turns on the approximate prefix tree (plan §7.5 tier 1). Off, the router uses session affinity alone, which is tier 0 and already most of the win for multi-turn traffic.

Takes a Prefix block.

kv_index#

KVIndex

KVIndex tunes the precise KV index (plan §7.5 tier 2). Which models use it is the catalog's to say, by kv_events on a vLLM profile; this is only how.

Takes a KVIndex block.

session_header#

string

SessionHeader carries the session id for affinity (plan §7.5 tier 0).

inject_usage#

bool

InjectUsage asks engines for usage even when the client did not, so token accounting works regardless of client behaviour.

tenant_label#

string

TenantLabel is the level of the tree the per-tenant request series are labelled by — llmgw_tenant_requests_total and its siblings: "org", "team" or "project", or "none" for no such series. A label per project on a cluster of thousands is more series than Prometheus wants, and a team is what an alert is usually about; "team" is the default.

tag_labels#

[]string

TagLabels are the request tags (metadata, x-llmgw-tags) counted in llmgw_tagged_requests_total and llmgw_tagged_tokens_total, by value. Only these: a tag is the client's to set, and every value of one is a series. The first 100 values of each are counted apart, the rest as "other". Empty counts none; the usage records keep every tag anyway.

max_body_bytes#

int64

MaxBodyBytes bounds the request body the router will read.

batch#

Batch

Batch is the Files and Batch APIs (/v1/files, /v1/batches): a file of requests served at the lowest priority, its answers written to another. Needs the store.

Takes a Batch block.

cache#

Cache

Cache bounds the answers kept for the projects that reuse them (llmgw keys cache). Which projects do, and for how long, is theirs to set; this is only how much memory it may take.

Takes a Cache block.

trusted_proxies#

[]string

TrustedProxies are the reverse proxies in front of the client port, as addresses or CIDR prefixes. Their X-Forwarded-For, -Proto and -Host are believed; anyone else's are not, since they decide the address written into the setup scripts and whose attempts are counted where.

active_active#

bool

ActiveActive has standby control planes answer the client API themselves, routing to the leader's replicas, rather than passing every request to the leader. More machines carry the traffic, which is the point. Rate limits are the cluster's: each router's counts reach the others through the leader every two seconds, so a key spread over N control planes may pass its limit by what they allow in those two seconds. A budget is checked against the store every half minute; cached answers and prefix affinity are split between the routers. Off by default. Which machine a client reaches is the load balancer's or DNS's to say; the floating address stays on the leader.

timeouts#

Timeouts

Takes a Timeouts block.

scoring#

Scoring

Takes a Scoring block.

admission#

Admission

Takes a Admission block.

Rules#

pii#

map[string]string

PII maps a kind (email, phone, iban, card, tckn, ip, secret) to what is done when it is found: flag, redact or block.

injection#

string

Injection is what is done on signs of prompt injection: flag or block.

deny#

[]Pattern

Deny are patterns, matched against the text as a model reads it.

max_chars#

int

MaxChars refuses text longer than this; zero is no limit.

restore#

bool

Restore, on input, puts what redact rewrote as placeholders back into the answer: the model reads "[EMAIL_1]", the client reads its own address. Only the values that request sent, only into its answer.

SCIM#

token_file#

string

TokenFile holds the bearer token the provider sends. Empty turns provisioning off.

SGLangConfig#

python#

string

Python is the interpreter of the SGLang virtualenv. SGLang is launched as a module rather than a console script, which is the documented entry point and avoids depending on the script being on PATH.

SLO#

ttft_p99#

time.Duration

TTFTp99 is the time to the first token 99 requests in 100 are within.

tpot_p99#

time.Duration

TPOTp99 is the time per output token, after the first, 99 streamed requests in 100 are within.

ScaleWindow#

days#

[]string

Days are mon … sun, or weekdays, weekends; empty is every day.

from#

string

From and To are HH:MM; a window whose To is before its From runs past midnight.

to#

string

min#

int

max#

int

Max, when set, is the window's ceiling for load scaling; zero keeps the model's own.

Scaling#

min#

int

max#

int

idle_timeout#

time.Duration

IdleTimeout is how long an on-demand model may go unasked-for before its replicas are stopped. Zero takes the cluster default.

Minutes rather than seconds, always. Stopping a replica throws away the KV cache prefix routing spent its effort filling, and a model that flaps in and out loses that entirely — so the saving has to be worth more than the cache, which it only is over a real idle period.

schedule#

[]ScaleWindow

Schedule changes min and max by the time of day, in the control plane's control.timezone: more replicas kept through working hours, fewer at night, an on-demand model kept loaded while people are at work. The first window that holds now wins; outside them all, Min and Max apply.

prewarm#

time.Duration

Prewarm starts a window this long before it opens, so its replicas have loaded by then: loading a model is minutes, and the people who arrive at nine are the reason for the window.

Scoring#

target_waiting#

float64

target_inflight#

float64

kv_ceiling#

float64

session_load_factor#

float64

Secrets#

vault#

secrets.Vault

Vault is how vault: references and vault-transit: are reached.

Takes a Vault block.

master_key#

string

MasterKey seals the secrets the store keeps for itself — what every engine's key is derived from — so a copy of the store, or of a backup, does not hold them: vault-transit:KEY seals with Vault's transit engine, which never lets the key out; otherwise a reference to 32 random bytes (openssl rand -hex 32), which then must be kept as safely as the store. Values from before it was set are sealed the first time they are read.

Server#

listen#

string

Listen is the client-facing OpenAI API address.

admin_listen#

string

AdminListen serves the dashboard, the management API, health and pprof.

Loopback by default, and it should stay there: pprof can dump the process heap, and API keys are in that heap in plaintext. Every comment beside this setting has always said not to expose it, while the default bound every interface — a default that contradicts its own documentation is a footgun, so the default now matches the advice.

metrics_listen#

string

MetricsListen serves Prometheus metrics and nothing else.

Separate from AdminListen because the reason that one is loopback is pprof, not metrics — and metrics are the one thing on it that something else genuinely needs to reach. Sharing a listener meant a fleet could either be monitored or keep its heap private, which is not a choice anybody should have to make.

Empty means off, and metrics remain available on AdminListen either way, so nothing that scrapes it today has to change.

agent_listen#

string

AgentListen is where node agents connect. Separate from the client port so the fleet-facing surface can be firewalled independently.

auth#

Auth

Auth turns on API keys on the client-facing port.

Takes a Auth block.

threads#

Threads

Threads keeps conversations and agent memory without API keys, for a gateway on one trusted machine. With auth on they live in the auth store, per project, and this block is ignored.

Takes a Threads block.

landing#

Landing

Landing configures the page served at "/".

Takes a Landing block.

agent_tls#

tlsconf.Config

AgentTLS authenticates the fleet-facing port. It deliberately does not cover the client-facing one: that surface is fronted by a publicly trusted certificate, while this one is mutual and private.

Takes a tlsconf.Config block.

max_connections#

int

MaxConnections bounds the client port's open connections; past it new ones wait in the kernel's backlog. 0 is no bound.

admin_token_file#

string

AdminTokenFile holds the token the admin API's /api and /debug paths demand. Empty is admin-token in the systemd unit's state directory (/var/lib/llmgw), or secrets/admin-token beside this file for a server run by hand. Made, 0600, when the server starts and it does not exist; sudo llmgw ctl and the console read it from there.

admin_remote#

bool

AdminRemote allows admin_listen on an address other than loopback. The token still guards the API, but it then crosses the network in clear.

debug_pprof#

bool

DebugPprof serves Go's profiler on the admin listener, behind the token. Off by default: a heap profile holds the API keys requests presented.

agent_insecure#

bool

AgentInsecure lets a control plane run its agent port without agent_tls: plaintext, and any machine that reaches it may register as any node — be given replicas, and so be sent users' prompts — or, in the other direction, a machine in the path may tell every agent what to run. Refused unless said here, and then the port listens on loopback unless agent_listen names an address. For one machine, a test, or a network nothing untrusted reaches.

monitoring#

Monitoring

Monitoring is where the dashboard reads its history from.

Takes a Monitoring block.

tracing#

obs.Tracing

Tracing sends traces to an OpenTelemetry collector: each inference request, and on a control plane each placement, replica start and model job. Off unless an endpoint is set here or in OTEL_EXPORTER_OTLP_ENDPOINT.

Takes a Tracing block.

mirror#

Mirror

Mirror points the control plane at a model mirror it can administer. Empty means there is none, and /admin/mirror stays unmounted.

Takes a Mirror block.

offline#

Offline

Offline is a cluster that reaches nothing outside itself: its nodes install engines, and upgrades fetch releases, from the mirror, out of a bundle carried in (llmgw bundle create, llmgw bundle import).

Takes a Offline block.

engines#

map[string]string

Engines are the engines every node installs, and at which version. The agents install them themselves — the right build for their hardware, checked against its publisher's checksum and seen to run — so a node is never trusted to already have what somebody once put there.

Unset keeps DefaultEngines. A version replaces the default; an empty one leaves that engine to be installed by hand.

engine_overrides#

[]EngineOverride

EngineOverrides give some nodes different engines from the rest: a vLLM built for CUDA 12 on the Volta cards CUDA 13 dropped, or none at all on nodes meant only for llama.cpp. Each picks nodes and lays its engines over the cluster's, in order, later ones last; an empty version removes that engine from the nodes it picks.

backup_dir#

string

BackupDir is where llmgw server backup writes, and where the newest archive's age is read for llmgw_backup_last_success_timestamp_seconds. Empty means /var/backups/llmgw-server.

guard#

guard.Config

Guard is the library of guardrail policies: what requests and answers are checked for, and what is done. Applied to every request through guard.default, and to parts of the tenancy tree by attaching them (llmgw ctl guard attach). Empty checks nothing.

Takes a guard.Config block.

secrets#

Secrets

Secrets is where secrets kept outside files are read from, and the key the store's own secrets are sealed under.

Takes a Secrets block.

notify#

notify.Config

Notify sends what needs a person — a model parked, a node lost, a budget spent, new weights published — to webhooks, signed, without a Prometheus and Alertmanager in between. Empty sends nothing.

Takes a notify.Config block.

mail#

mail.Config

Mail is where the gateway's few emails go out from: a link to set a forgotten password. Unset, none is sent and people ask an operator.

Takes a mail.Config block.

serving#

ServingSettings

Serving stops and starts the whole cluster's service (llmgw stop, llmgw start).

Takes a ServingSettings block.

firewall#

FirewallSettings

Firewall turns llmgw's firewall on every machine on or off.

Takes a FirewallSettings block.

slurm#

SlurmSettings

Slurm is a Slurm cluster run beside llmgw on some of its nodes, and which of them it may use now.

Takes a SlurmSettings block.

presets#

PresetSettings

Presets is how the models that come with llmgw reach the catalog.

Takes a PresetSettings block.

on_demand#

OnDemandSettings

OnDemand is how models that load when asked for take turns on the cards.

Takes a OnDemandSettings block.

agents#

AgentSettings

Agents are the nodes' agent settings: agent.yaml's keys but those an agent needs to start, for every node (default) and for one (nodes.). Sent to the agents as they change; an agent whose settings changed starts again on them, its engines serving throughout.

Takes a AgentSettings block.

firewall_networks#

[]string

FirewallNetworks are networks, besides those each machine reaches the cluster on, that llmgw's internal ports answer: a monitoring host on another subnet, nodes behind a router. Written into every machine's firewall by server init, node add and server add.

backup_recipients#

[]string

BackupRecipients encrypt every archive llmgw server backup writes, to whoever holds one of their keys: age recipients (age1…) or SSH public keys (ssh-ed25519 …, ssh-rsa …), as age takes them. An archive holds the certificate authority's key and every secret in the store, and is copied off the machine; encrypted, the copy is safe wherever it lands. Empty writes archives unencrypted, readable by root only.

backup_copy_to#

[]string

BackupCopyTo is where each archive is copied once written, so it survives this machine: a directory (a mounted share, another disk) or ssh://user@host[:port]/path, reached with BackupCopyKey and checked against /etc/llmgw/backup_known_hosts, which learns a host the first time. The same number of archives is kept there as here. Refused without backup_recipients: an archive leaves the machine encrypted.

backup_copy_key#

string

BackupCopyKey is the SSH private key for ssh:// copies.

releases#

string

Releases is where llmgw ctl upgrade fetches release binaries from, for this machine and — through their agents — every node. Empty means the project's own releases.

router#

Router

Takes a Router block.

control#

Control

Takes a Control block.

models#

[]catalog.Model

facility#

Facility

Facility is the power caps of racks, PDUs and other domains, and how a replica's draw is estimated against them. Where a node is comes from its llmgw.io/site, room, row, rack and pdu labels; holding a domain for maintenance is llmgw ctl facility hold, not configuration.

Takes a Facility block.

federation#

Federation

Federation is how this cluster takes part in federation with other llmgw clusters. The peers themselves are added at runtime, with llmgw ctl federation add, and kept in the store.

Takes a Federation block.

endpoints#

[]Endpoint

Endpoints is the phase-1 static replica list. It is only consulted when no model declares a profile, which is what makes a router-only smoke test against sim-engines possible without a control plane.

Service#

name#

string

Name says which service decided, in the verdict's rule.

url#

string

URL is POSTed the body.

header_files#

map[string]string

HeaderFiles are headers whose values are read from files each time, so a rotated key needs no restart: {Authorization: /etc/…/lakera}.

body#

string

Body is the request, a template given .Text; {{json .Text}} writes it as a JSON string. Empty is {"input": }.

blocked#

string

Blocked is where in the answer the service says the text is unsafe: a dotted path, numbers for list items (results.0.flagged). True, a number above zero, a non-empty list or string other than "false" is unsafe.

reason#

string

Reason is where it says why, for the verdict; optional.

on#

string

On: input, output or both. Empty is input.

action#

string

Action on unsafe: block or flag. Empty is block.

timeout#

time.Duration

Timeout bounds one call. Empty is 3s; past it the policy's fail says.

ServingSettings#

stopped#

bool

Stopped takes every model down and answers every request for one with 503 (system_stopped), leaving the control planes, agents and mirror running and every model's definition as it is: llmgw stop sets it, llmgw start clears it, and the models come back as their catalog entries ask. In force on every control plane within seconds.

SimConfig#

bin#

string

Bin is the llmgw binary itself; the engine is one of its subcommands. Empty means whatever started this agent, which is right almost always.

SlurmSettings#

enabled#

bool

Enabled runs Slurm on the nodes below. False stops llmgw keeping it; what was installed is left installed, and every node is llmgw's.

cluster_name#

string

ClusterName is Slurm's ClusterName. Empty is "llmgw".

controller#

string

Controller is the llmgw node whose agent runs slurmctld. It needs no GPU, and need not be one of Nodes.

nodes#

[]string

Nodes are the llmgw nodes that run slurmd. Each stays llmgw's, drained in Slurm, until it is also in Lent.

all_nodes#

bool

AllNodes runs slurmd on every llmgw node, those that join later too, besides any named in Nodes. Each is llmgw's until it is lent, as any.

lent#

[]string

Lent are the nodes of Nodes Slurm may run jobs on now. A node put here is drained of llmgw's replicas first, as ctl drain does, and resumed in Slurm once none is left; a node taken out is drained in Slurm and given back to llmgw when its last job ends.

partition#

string

Partition is the one partition every node is in. Empty is "gpu".

StrataConfig#

dir#

string

Dir is Strata's source directory, where setup.py and serve/ are.

python#

string

Python is the interpreter of its environment; Dir's .venv when unset.

engine#

string

Engine is the directory of its prebuilt engine, where the strata binary and the CUDA libraries it ships are; Dir's engine/ when unset.

TRTLLMConfig#

python#

string

Python is the interpreter of the environment TensorRT-LLM is installed in: /bin/python for a virtual environment, python3 in NVIDIA's container. Naming it is what makes this node offer the engine; empty offers nothing.

The interpreter rather than trtllm-serve, because llmgw starts the server through a few lines of its own that make it demand the gateway's key, which trtllm-serve has no option for.

Threads#

store#

string

Store is the SQLite file. Empty means nothing is kept.

retain_hours#

int

RetainHours is how long idle conversations and notes are kept. Required with Store: a store with no expiry keeps everything anyone typed, by omission.

Timeouts#

dial#

time.Duration

response_header#

time.Duration

ResponseHeader bounds the wait for headers on a streaming request. It is a liveness check there, because headers arrive before generation starts.

generation#

time.Duration

Generation bounds a non-streaming request, where the engine sends no headers until the completion is done. Setting response_header alone used to cap both, so 60s read as a liveness guard and behaved as a limit of about a hundred tokens on a slow model.

first_token#

time.Duration

idle#

time.Duration

total#

time.Duration

Tools#

allow#

[]string

Allow, when set, are the only tool names a request may offer and a model may call.

deny#

[]string

Deny are tool names never offered or called.

on_denied#

string

OnDenied is strip (the tool is removed from the request) or block. Empty is strip.

max_calls_per_turn#

int

MaxCallsPerTurn stops an answer calling more tools than this.

validate_arguments#

bool

ValidateArguments checks a call's arguments against the parameters schema the request declared for the tool.

Tracing#

endpoint#

string

Endpoint is the collector, for example http://otel.internal:4318. Empty falls back to OTEL_EXPORTER_OTLP_ENDPOINT, and to no tracing.

sample_ratio#

float64

SampleRatio is the share of traces kept, 0 to 1, of those nobody upstream already decided. Zero takes 1. A request with a sampled traceparent of its own is kept, and an agent keeps every start the control plane kept, whatever their own ratio says.

VLLMConfig#

bin#

string

Bin is the vllm executable.

launch_mode#

string

LaunchMode selects the multi-node mechanism: "mp", the default, or "ray".

mp is what vLLM documents today — every node runs serve with its rank, the workers headless — and needs nothing but vLLM. ray is the older way, kept for a vLLM too old for mp's multi-node flags; it needs Ray on every node and has not been checked against a real cluster.

ray_bin#

string

RayBin is used when LaunchMode is "ray".

Vault#

addr#

string

Addr is Vault's address, like https://vault.internal:8200; empty takes VAULT_ADDR.

token_file#

string

TokenFile holds a token; empty takes VAULT_TOKEN, or logs in with AppRole when RoleIDFile and SecretIDFile are set.

role_id_file#

string

secret_id_file#

string

ca_file#

string

CAFile verifies Vault's certificate; empty uses the system's.

namespace#

string

Namespace is Vault Enterprise's or OpenBao's namespace.

Webhook#

url#

string

format#

string

Format is json (the event, with X-Llmgw-Signature when SecretFile is set) or slack ({"text": …}, which Slack, Mattermost and Rocket.Chat incoming webhooks take).

events#

[]string

Events are the kinds sent here; empty is all.

secret_file#

string

SecretFile holds the key the body is signed with: X-Llmgw-Signature is sha256=.

accounting.Price#

prompt_per_1k#

int64

completion_per_1k#

int64

cached_per_1k#

int64

CachedPer1k is what a prompt token costs that the engine or provider read from its cache instead of computing: the part of prompt_tokens the answer counts as cached_tokens. Zero charges those at PromptPer1k, as before there was a rate for them.

per_image#

int64

PerImage is what one picture costs, for a model that makes them: a picture's cost is the card's time, which its prompt's few tokens do not measure.

per_mesh#

int64

PerMesh is what one 3D model costs: minutes of a card, which no token count measures.

admission.Band#

name#

string

priority#

int

Priority orders bands; higher is served first.

saturation_ceiling#

float64

SaturationCeiling is the pool saturation above which this band waits. 1.0 means "never hold this band back".

queue_ttl#

time.Duration

QueueTTL, set, is how long this band's requests may wait instead of the controller's: a band for bulk work — a knowledge-base import, a backfill — whose requests had rather wait ten minutes behind people than be turned away after thirty seconds.

admission.Config#

bands#

[]Band

default_band#

string

queue_ttl#

time.Duration

QueueTTL drops a request that has waited this long. Sending the GPU work the client has already given up on helps nobody.

queue_capacity_per_band#

int

QueueCapacityPerBand bounds each band's queue. Rejecting immediately beats accumulating silently.

queue_capacity_total#

int

QueueCapacityTotal bounds everything queued, across bands. When it is reached, a request is not refused while a lower band has requests queued: the newest of the lowest band is turned away instead (displacement), since the oldest have waited longest and are nearest their TTL. Zero takes half of every band's capacity together.

interval#

time.Duration

Interval is how often the dispatcher re-checks saturation. Releases also wake it, so this is only the backstop for saturation falling on its own.

agent.Config#

node_id#

string

NodeID defaults to the hostname. It must equal the common name of the agent's certificate, so set it only where the two differ.

server#

string

Server is the control plane, and Servers is the same thing when there is more than one. Both are accepted: every configuration written before there could be a second control plane names one address, and none of them should have to change.

The order is tried in order, and a refusal moves to the next. Only the leader accepts an agent, so a standby is skipped in the time it takes to be told so — which is what makes failover a reconnect rather than an edit to every node's configuration.

servers#

[]string

hostname#

string

Hostname defaults to the machine's.

mgmt_addr#

string

MgmtAddr is the address engines are reached on. It defaults to this machine's address on the route to the control plane, which is the one the control plane and the router can reach.

labels#

map[string]string

Labels are free-form facts about a node for a profile's node_selector. Nothing needs them: memory, GPUs, engines and disk are measured, and a label restating one of those goes stale the day the machine changes.

mgmt_interface#

string

MgmtInterface is the ethernet interface NCCL bootstraps over. It defaults to the interface holding MgmtAddr, then eth0.

fabrics#

[]inventory.FabricAttachment

Fabrics labels this node's HCAs. It is explicit on purpose: OpenSM's default subnet prefix is the same on every fabric, so nothing can be inferred from the wire (plan §5.2).

heartbeat_interval#

time.Duration

reconnect_base#

time.Duration

ReconnectBase is the first reconnect delay; it backs off from there.

reconnect_max#

time.Duration

adopt_ready_timeout#

time.Duration

AdoptReadyTimeout bounds how long an adopted engine is given to answer a health check before the agent stops waiting on it.

Shorter than a startup timeout on purpose: this engine is already running, so it is not loading a model, it is either serving or it is not.

tls#

tlsconf.Config

TLS authenticates both ends of the control-plane stream. Leaving it unset keeps the plaintext behaviour, which only belongs on a trusted network.

Takes a tlsconf.Config block.

model_dirs#

[]string

ModelDirs are the directories this node keeps model weights in, beyond WeightsCache, which is always looked in. Storage differs at every installation, so llmgw observes it rather than owning it: a Lustre mount, an image exported over the fabric, local NVMe and a copy made by hand all look identical from here, and placement only needs to know whether the file is present.

engines_dir#

string

EnginesDir is where the agent installs engines, one directory per build. Defaults to /opt/llmgw/engines.

weights_cache#

string

WeightsCache is where a remote source is downloaded to. Defaults to the first entry of ModelDirs, because a node that says where its models live has already answered this question, and otherwise to /var/lib/llmgw-weights.

Fetching is the simple arrangement: every node keeps its own copy. The exotic ones need no code — weights on a shared filer or an image exported over the fabric are simply present, and present means nothing is fetched.

discovery#

discovery.Options

Takes a Options block.

engines#

engines.Config

Takes a engines.Config block.

supervisor#

supervisor.Config

Takes a supervisor.Config block.

tracing#

obs.Tracing

Tracing sends this node's part of each replica start — the weights fetched, the engine launched, the wait for it to answer — to an OpenTelemetry collector, inside the control plane's trace of that start. Give it the collector server.yaml's tracing names. Off unless an endpoint is set here or in OTEL_EXPORTER_OTLP_ENDPOINT.

Takes a Tracing block.

config.Band#

name#

string

priority#

int

saturation_ceiling#

float64

queue_ttl#

time.Duration

QueueTTL is how long this band's requests may wait for room; the admission's queue_ttl unless set. Long for a band of bulk work, which had rather wait behind people than be turned away.

config.Price#

prompt_per_1k#

int64

completion_per_1k#

int64

cached_per_1k#

int64

CachedPer1k is the cost of a prompt token the engine or provider read from its cache (the answer's cached_tokens, counted within the prompt's). Zero charges them at prompt_per_1k.

per_image#

int64

PerImage is the cost of one picture, for a model that makes them (/v1/images/generations), on top of its prompt's tokens.

per_mesh#

int64

PerMesh is the cost of one 3D model (/v1/3d/generations).

engines.Config#

disabled#

[]string

Disabled are engines this node never installs or advertises, whatever the control plane asks: vLLM on a node of cards no build of it runs on, that is meant for llama.cpp alone.

vllm#

VLLMConfig

Takes a VLLMConfig block.

sglang#

SGLangConfig

Takes a SGLangConfig block.

llamacpp#

LlamaCppConfig

Takes a LlamaCppConfig block.

trtllm#

TRTLLMConfig

TRTLLM is TensorRT-LLM, which the agent does not install: a node offers it only when its interpreter is named here.

Takes a TRTLLMConfig block.

sim#

SimConfig

Takes a SimConfig block.

comfyui#

ComfyUIConfig

ComfyUI makes pictures; offered by a node whose ComfyUI is named here.

Takes a ComfyUIConfig block.

strata#

StrataConfig

Strata serves a few mixture-of-experts models from RAM and one card; offered by a node whose Strata is named here, or that installed it.

Takes a StrataConfig block.

freetoken#

FreeTokenConfig

FreeToken serves mixture-of-experts models larger than a card from RAM and one card of 8.0 or newer; installed by the agent where the cards allow it, or named here.

Takes a FreeTokenConfig block.

guard.Config#

enabled#

*bool

Enabled turns every policy on or off at once: false checks nothing, whatever is defined or attached. Unset is on.

default#

[]string

policies#

map[string]Policy

mail.Config#

smtp#

string

SMTP is the relay, host:port: 587 with STARTTLS, 465 with TLS from the start, 25 as the relay allows. Empty sends nothing.

from#

string

From is the sender, llmgw@example.com or "llmgw llmgw@example.com".

username#

string

Username and PasswordFile log in to the relay, when it wants that. The password is a secret reference: a file, env:NAME or vault:path.

password_file#

string

base_url#

string

BaseURL is the gateway's public address, which links in mail point at: https://llm.example.com. Never taken from a request, whose Host a sender chooses and could point a reset link at their own site.

notify.Config#

webhooks#

[]Webhook

host#

string

Host names the gateway in messages; empty is this machine's name.

supervisor.Config#

work_dir#

string

WorkDir holds pid files and per-shard scratch space.

log_lines#

int

LogLines is how many lines of engine output to retain per shard.

stop_timeout#

time.Duration

StopTimeout is how long a process gets to exit after SIGTERM before it is killed. Engines need time to free GPU memory cleanly.

health_interval#

time.Duration

HealthInterval is how often a starting shard is polled.

rendezvous_timeout#

time.Duration

RendezvousTimeout bounds the wait for a multi-node rendezvous point to start listening.

alive_grace#

time.Duration

AliveGrace is how long a non-serving rank must stay up before it counts as ready. There is no external signal for a worker, so staying alive is the only evidence available.

run_as#

string

RunAs is the user engines run as when the agent is root; empty is llmgw-engine, "root" keeps them root.

tlsconf.Config#

ca#

string

CA is the bundle used to verify the other side. In a real deployment this is whichever authority issued the fleet's certificates.

cert#

string

Cert and Key are this component's own identity.

key#

string

insecure#

bool

Insecure disables TLS entirely. It exists for local runs against sim-engines; anything using it must stay on a trusted network.