Skip to main content

Configuration reference

LLMBoost runs great with defaults. When you need to tune, here's the surface.

Source of truth

llmboost serve --help always lists the complete, version-correct set of flags for your build. The tables below cover the ones you'll reach for most.

vLLM-compatible CLI

Migrating from vLLM? llmboost serve exposes a vLLM-compatible serve interface — the same flag names and semantics — so your existing launch scripts, configs, and tooling run against it unchanged. (Same idea as the OpenAI-compatible API: meet you where your stack already is.)

llmboost serve <Repo/Model-Name> [flags]
FlagDefaultWhat it does
--port N8000Port for the HTTP server.
--tensor-parallel-size NautoShard the model across N GPUs.
--max-model-len Nmodel maxCap context length (prompt + output).
--gpu-memory-utilization 0–1tunedFraction of VRAM for weights + KV cache.
--max-num-seqs NtunedMax concurrent sequences per batch.
--enforce-eageroffSkip graph capture (faster start, lower peak throughput).
--scheduling-policy fcfsfcfsRequest ordering policy.
--disable-llmboost-optsoffRun without LLMBoost's licensed optimizations (diagnostics).

For more detail, use llmboost --help (or, per command, llmboost <command> --help); llmboost serve --help always lists the complete, version-correct flag set.

Environment variables

VariablePurpose
HF_TOKENHugging Face token for gated/private model downloads.

Other advanced scheduler knobs are documented in llmboost serve --help. (Server log verbosity is fixed for now and not user-configurable.)

LLMBoost Hub config

LLMBoost Hub (lbh) manages images and models:

lbh list # models + status
lbh serve <Repo/Model> -- <serve flags> # flags after `--` pass to the server

See Run with LLMBoost Hub.

Licensing

LLMBoost activates a license on first start (a free trial is provisioned automatically) and then serves locally. Keep the host online for first activation. For enterprise seats and offline licensing, contact mangoboost.io/contact.