Adding a node#

Written for whoever runs the cluster.

The short version#

On the control plane:

sudo llmgw node add ubuntu@10.0.0.145

That is all a machine that already exists needs — bare metal, or a VM you built. install.md says what it does. On Proxmox, the machine itself can be built first by the script below, on the hypervisor.

The two halves#

Two scripts, because the two halves of onboarding are done by different people with different access: one needs the hypervisor, the other needs the certificate authority.

1. The machine#

On the Proxmox host:

llmgw-provision-node --vmid 144 --name llmgw-node04 --ip 10.0.0.144/24

Name it after the node id. The VM name, the hostname, the node id and the certificate's common name are all the same string. They do not have to be, and every time they are not, somebody spends an afternoon working out which of the four a message is talking about.

It imports a cloud image and lets cloud-init do the rest, which is why a node takes a minute rather than an afternoon. It refuses to touch an existing VMID: adopting a machine is not a decision a provisioning script should make.

Releases carry static linux/amd64 and linux/arm64 binaries. Download one and pass it with --binary rather than copying whatever happens to be on the machine you are running onboarding from — otherwise whoever runs the script quietly becomes the build server, and the fleet ends up running whatever was in their working tree.

2. The identity#

On the control plane, which holds the certificate authority — this is the part that cannot be delegated, because it issues the credential by which the control plane believes the node is who it says it is:

sudo llmgw node add ilker@10.0.0.144 --fabric ib-b:mlx5_0,mlx5_1

It checks the machine with llmgw doctor, issues a client certificate whose common name is the node id (the hostname), installs it along with the CA root, trusts that root system-wide, installs the binary and the unit, starts the agent, and waits for the node to join.

If it does not join, the agent's log on the node says why. A certificate whose common name does not match the claimed node id is refused by design — that check is what makes a node's identity something it proves rather than something it picks.

It needs SSH from the control plane to the node. Provisioning copies the hypervisor's authorized_keys into the new machine; with ssh -A to the control plane and sudo --preserve-env=SSH_AUTH_SOCK, the same key that reaches the hypervisor reaches the node.

deploy/bootstrap/onboard-node.sh is the script this replaced, for a machine holding an authority but not a current llmgw.

The CPU model matters, and its default is wrong#

Nodes are provisioned with --cpu host. Proxmox's default, kvm64, hides AVX, AVX2, FMA and even SSE4.2 from the guest, and llama.cpp then loads its scalar backend instead of the one built for the host's instruction set. Measured on node05, an 8-core VM on a Xeon E5-2680 v4, same weights, same threads, the only change being the CPU model:

Prompt processing is worse still: 7.6 tokens a second against 143.8, a factor of nineteen.

It is worth calling out because nothing looks broken. The node registers, the engine starts, replies are correct — every number is simply small, and the obvious conclusion is that CPU inference is hopeless rather than that the VM cannot multiply efficiently.

The trade is live migration: a VM pinned to the host CPU cannot move to a machine with a different one. For a node whose job is arithmetic that is the right way round, and --cpu overrides it where it is not.

Engines#

The script installs none, and does not need to: once the agent is connected, the control plane tells it which engines to have and it installs them itself — llama.cpp on every node, vLLM on every node with an NVIDIA GPU it can run on, SGLang where the control plane names a version. Each is checked to run before the node offers it, so a node never advertises an engine that cannot start. install.md says what vLLM needs of the node.

--llamacpp-bin, --vllm-bin and --sglang-python are for an engine you installed yourself. For the agent to use yours rather than install its own, also set that engine's version to "" under engines in server.yaml.

Weights are the other thing a node needs. llmgw fetches them for a model whose source is a URL, one copy per node, which covers the simple case; anything larger — a shared filesystem, a cache — stays the operator's arrangement, and llmgw is content to find the files already there.

The sim engine needs neither, which is what makes it useful for exercising everything else.

Checking an engine after an upgrade#

Engines rename and remove flags between versions, and when that happens the failure arrives minutes into a model load, in the engine's own error message, on a node nobody was watching.

llmgw ctl engine-check vllm
llmgw ctl engine-check llamacpp --help-output "$(llama-server --help 2>&1)"

It builds the widest command llmgw would ever produce — a multi-node one, since a flag that only appears there is exactly the one nobody notices has gone — and checks every flag against the engine's own --help. No GPU required: a CLI prints its help on any machine.

Some programs do not give up their whole help for --help. vLLM prints a summary of section names and expects --help=all; checking against the summary reports every flag as missing, confidently and wrongly. engine-check knows that one. If you add an engine, check that what you get back is the full list.

vLLM also refuses to do anything at all on a machine with no accelerator, including print its help, unless VLLM_TARGET_DEVICE says what to assume. engine-check sets it.

Give it the whole help. Truncating llama.cpp's 725 lines to 400 reports five flags as missing that the running cluster uses; the command warns when the text it was given does not document --help itself, which is the cheapest sign it was cut short.

Fabrics#

--fabric is how a node says which InfiniBand fabric it is attached to. The name is checked against what the adapters report — every port on one fabric sees the same subnet managers — so two independent fabrics given one name are placed as two, and the cluster page says so; give each fabric its own name. The agent installs infiniband-diags, which that check needs, if it is missing.

Before a node with a fabric joins, run sudo llmgw doctor on it: it checks the subnet manager, GPUDirect RDMA, the fabric manager and IMEX on NVLink machines, and the declared fabrics against the adapters. fabrics.md has the rest.

Omit it for a node with no fabric. Such a node is its own island: it can host single-node replicas and will never be given a multi-node one.

Two adapters per fabric is the default; --fabric ib-b:mlx5_2,mlx5_3 names them explicitly. An adapter must never appear under two fabrics — the agent refuses that configuration: a replica given adapters on two fabrics starts, passes its health checks, and hangs on its first all-reduce.

Where the driver is#

--nvidia-smi /opt/nvidia/bin/nvidia-smi
--ib-root /sys/class/infiniband

Both default to the usual places. Containerized driver installs routinely put nvidia-smi somewhere else, and an agent that cannot find it registers with no GPUs — which looks exactly like a node that has none.

Removing a node#

llmgw ctl drain llmgw-node04     # move its replicas off, one at a time
llmgw ctl nodes                  # wait until it holds none

Then stop the agent. Draining brings each replacement up before stopping the original, which is the difference between a rolling move and an outage.

The node's certificate stays valid until it expires. There is no revocation list — if a machine is decommissioned rather than retired, reissue the authority or accept that a certificate naming a node that no longer exists can still connect and claim to be it.

Testing without hardware#

Every part of this works on machines with no GPUs and no InfiniBand. Point --nvidia-smi at a script that prints the CSV discovery asks for, point --ib-root at a sysfs-shaped directory, and use engine: sim in the catalog.

The simulated engine rendezvouses for real: rank 0 binds the dist-init port and does not serve until every worker has joined, and workers serve nothing. That exercises the multi-node launch sequence, the fabric constraint, and failure and rebuild — which is most of what goes wrong — without exercising NCCL, which is the rest.