Onboard a node with a script, and find two examples that never worked

New nodes had been joined by hand, four times, by me. There was no procedure,
so there was nothing to be wrong — and nothing to be right either. Two scripts
now: provision-node.sh makes the VM on Proxmox, onboard-node.sh gives it an
identity. They are separate because they are run by different people with
different access, and the second must run where the certificate authority is,
which is the one part that cannot be delegated.

Running them turned up why phase 4 was never verified. FabricAttachment
carried no koanf tags, so the `name:` key the shipped example documents bound
to nothing and `rate_gbps` was never read at all. An agent configured exactly
as instructed refused to start with "fabrics[0]: name is required" — naming
the key the operator had just written. Every fabric in this project had been
configured in a Go test, where the tags do not matter.

The server example did not load either: its models had no source, so it failed
validation the moment anything tried to parse it.

Both examples are now parsed by a test. A published example that has never
been loaded is a guess, and these two had been guesses since they were
written.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Ilker ManapandClaude Opus 5 committed 2026-09-16 12:28:36 +02:00
1 parent a075ff4120
commit c1ac2f022f
11 files changed
+687 -9

No files matched your search

+36
View File
@@ -206,15 +206,51 @@ control:
port_range: [20000, 20999]
mgmt_interface: eth0
# Each model needs a source — a weights path or a HuggingFace id — and at least
# one profile if the control plane is to place it. Without profiles the model is
# catalogued but never started, which is the router-only mode below.
models:
- name: llama-3.3-70b
source: meta-llama/Llama-3.3-70B-Instruct
aliases: [llama]
fallbacks: [qwen2.5-32b]
routing_profile: chat
scaling: { min: 1, max: 4 }
profiles:
# Two nodes of four H100s each. A multi-node profile requires a fabric,
# and a replica never spans two of them.
- name: tp8
engine: vllm
gpu_class: h100-80g
tp: 8
nodes: 2
gpus_per_node: 4
priority: 100
weight: 8
startup_timeout: 30m
# Where a whole node is free, one node is cheaper to schedule and faster
# to start than two.
- name: tp4-a100
engine: vllm
gpu_class: a100-40g
tp: 4
priority: 50
weight: 3
startup_timeout: 20m
- name: qwen2.5-32b
source: Qwen/Qwen2.5-32B-Instruct
aliases: [qwen]
routing_profile: batch
scaling: { min: 1, max: 2 }
profiles:
- name: tp2
engine: sglang
gpu_class: h100-80g
tp: 2
priority: 100
weight: 2
startup_timeout: 15m
# A static replica list, used only when no model declares a profile. That is the
# router-only mode for testing against sim-engines; with the control plane