GLBs carried the licence's name, URL and notice in asset.extras.license.
A pipeline that sorts assets by licence needs what a machine matches on:
the id, whether commercial use is allowed, the regions it excludes, the
revenue cap and the attribution. The whole licence is written, as
/v1/models gives it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Stable Fast 3D takes the light out of the picture it sees, and a dark
basalt (median 0.17) came back a light stone dimly lit (0.67). The save
node takes albedo_match, the picture and its mask, and scales the base
colour texture in linear light by the ratio of the object's medians, so
marks keep their contrast; texels nothing maps to are left alone. On the
game team's rock it is 0.18.
A workflow's {{name|false}} default now fills a BOOLEAN input with a
boolean rather than the word.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A site whose catalog had no flux-schnell or triposr showed no Images or 3D
page, and nothing said they were one click away. The cluster page now has
a Pictures and 3D models card under the models: each preset, what it
makes, the classes of card here it fits, whether and by whom it was
added, and why the leader has not added it by itself; an Add button, which
says first when it would install ComfyUI on the GPU nodes and restart the
control planes; and the switch for presets.auto. The console's models
screen shows the same under its table: A adds one, asking first in the
same case, and P turns automatic adding on or off.
`ctl model add PRESET` and `ctl model presets` go through the presets API,
so the control plane keeps who added one, and `ctl model presets auto
on|off` flips the setting; against a control plane from before, add sends
the entry as v0.14.4 did.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A customer's portal had no Images or 3D pages: nobody had written a
definition for either, on a cluster whose V100 runs both. The leading
control plane now adds flux-schnell and triposr, on demand, once a node
has a card that fits one, the cluster installs ComfyUI, and the catalog
makes nothing of that kind. presets.auto (a live setting, on by default)
turns it off; a preset somebody removes is recorded as declined and not
added back until added again by hand. Each addition is in the audit log.
GET /api/v1/presets says how each preset stands, the ComfyUI the cluster
installs, and why one is not added; POST /api/v1/presets/{name} adds one
(setting engines.comfyui where there is none); PUT /api/v1/presets/auto
switches it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
flux-schnell and triposr come with llmgw as presets: their workflow,
custom nodes, packages and licence as they were found and tried on the
PoC, and a profile that says the card it needs rather than a class, so the
entry places on any site's cards. TripoSR's workflow is built in, with the
LLMGWPrepareForMesh node that keeps a picture's background out of the
mesh, which the copy in the document had left out.
`ctl model add PRESET` adds one, and has the nodes install ComfyUI when
the cluster's settings name none; `ctl model presets` lists them; the
console's add-model dialog offers them beside a definition file. A
ComfyUI build is now also checked to keep its order on a Turing card, so
a node's installed cu13 is not replaced by the cu126 build.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PyTorch's CUDA 12.8 and 12.9 builds start at Turing since torch 2.8, as
CUDA 13 does, so on a V100 both ComfyUI builds were installed, 10 GB each,
and failed on the card with "no kernel image". The cu129 build now needs
7.5, and a third, on PyTorch's CUDA 12.6 torch, carries Volta and Pascal.
The open kernel module, NVIDIA's default since 560, writes "Open Kernel
Module for x86_64 580.82.07" in /proc/driver/nvidia/version, and the word
after "Module" was taken as the driver: "for". With the driver unknown,
every build was tried whatever driver it needs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The VLM is to come first on the one card. Its class makes it take the card
from flux, triposr or sf3d when asked for, but idle eviction ignored class:
a 3D job two minutes into a lull would have taken the card from it. Idle
eviction offers only replicas of no higher class than the model wanting
room, and the wait such a model is told runs to the holder's idle timeout.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
sf3d's first jobs failed with a 401 from stabilityai/stable-fast-3d. The
weights were on the node, and the gate links them into the Hugging Face
cache when told the repository — but the agent passed on only the custom
nodes, so nothing was linked and the node went to the network for a gated
repository.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Removed and added again to give it a licence, qwen3-vl-8b came back loading
by itself: its name was still on the wanted list from before. A model that
leaves the catalog leaves the list too.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A 1280x720 PNG sent to qwen3-vl-8b was estimated at 138 thousand prompt
tokens and refused against its 8k context before reaching the engine, which
makes about a thousand of it. Parts are counted by kind: text by its text,
a picture or a sound as a thousand tokens.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Stable Fast 3D's texture_baker and uv_unwrapper are C++/CUDA packages no
index has; on the PoC they were built and installed into ComfyUI's
environment by hand. nodes[].build names such directories, and the gate
builds each as a wheel with ComfyUI's Python, for the compute capability
torch reports for the replica's own card, keeps it on the node by the
node's commit, the architecture and the directory, and installs it
beside the replica with the node's requirements. A copy in ComfyUI's
environment does not pin it. A node with no CUDA toolkit says so.
comfyui.huggingface_repo shows the replica's weights to a node that
loads them by repository name, in the engines' Hugging Face cache, so a
gated repository needs no token on the node.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A game sold in the EU cannot ship what a licence excluding the EU made. On
the profile page a person leaves out licences one by one, or by rule:
every licence forbidding commercial use, every licence granting nothing in
a region. The rules cover models added later. The choice reaches every
principal through the resolved limits, so permits() leaves such models out
of /v1/models, requests, jobs, fallbacks and the portal's pickers, and a
request for one is refused 403 licence_excluded saying which choice did.
Kept in user_excluded_licences (migration 0046), copied with the store and
gone with the person.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A catalog entry meant for many sites cannot know their GPU class names:
the PoC has a Quadro RTX 6000, a customer an A100, another an L40S. A
profile may now say gpu: {min_compute_capability, min_memory_gb} in
place of gpu_class, and stands for every class in the cluster that meets
it, named <profile>@<class>, at its own priority; placement still sees
exact classes. ProfileByName and EngineProfile resolve such a name to
the profile on that class, so every caller that maps a replica back to
its profile works unchanged. A ComfyUI profile may carry its own
workflow, checkpoint and nodes, and with its own source an fp8 workflow
and weights go to cards with fp8 and fp16 ones to Turing.
A profile by need is left out of placement until the cluster's cards are
known: with no class it would have read as a CPU profile.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A model's catalog entry now says what its workflow takes and makes:
comfyui.inputs, each a placeholder of its name with a type (text, int,
float, bool, choice, image, mask, audio, video, file), bounds, choices
and a default; and comfyui.output, which gains video, audio and files
beside pictures and meshes. POST /v1/workflows/runs with {model, inputs}
checks a request against them before a job is written — a required input
left out, an unknown one, a number out of bounds, a song sent as a
picture, another owner's file id are each refused with what was wrong —
and the job worker replays it to a replica, where the files go into
ComfyUI's input directory, the workflow is filled and run, and the first
file of the model's output kind is the job's result. /v1/models lists
each ComfyUI model's workflow, so a client can build its form.
A PNG result carries the licence in an iTXt chunk, as a GLB does in
asset.extras. A file whose server named no type is typed by its name
before its bytes, so a WAV is audio/wav, not audio/wave.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A ComfyUI model may name an environment: the models of one environment
agree on every custom node they both name and on every package range,
so one replica can carry all of them and ComfyUI moves between their
weights itself rather than the card being handed over with a cold start
each time. The catalog refuses an environment that cannot share a
Python, naming both models; Catalog.Environments gives each one's
models, the union of their nodes and their weight sources, for
placement to build on.
A licence gains an id (SPDX where there is one), whether its output may
be used commercially, the regions it excludes, a revenue cap and the
attribution it asks for, so a person's or a site's choice of licences
can be matched without reading prose.
BACKLOG gains section H: ComfyUI as a platform.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
After a control plane restart flux, adopted and asked for by nobody, had no
last use and so was never idle enough to evict: a 3D job waited out its
whole fifteen minutes with no estimate. It counts as used at adoption, as it
is already counted as wanted then; a replica in use is reported by the
router within seconds.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A 3D job leased by the standby control plane, whose router has no endpoints
and which a job replayed in the process never leaves, was told "loading"
for seven minutes while the leader served the jobs behind it. Jobs and
batches are claimed only while leading.
Five jobs asked for while cuda01's agent restarted failed for good as
"cannot be placed": re-registering, the node had no engines for a few
seconds. A job asks again for five minutes before that is its end.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Stable Fast 3D's own save node hands back the GLB as base64 in its UI
output, which the gate does not read. LLMGWSaveMesh writes each mesh to
output/ as a GLB and names it, as a picture's save node does. The
prepare node also gives the object's mask now, which SF3D's sampler
takes.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A 3D job waiting for the picture model to give up the one card had an empty
progress; its client could not tell waiting from stuck. The control plane
says which model holds the cards and when it lets go, plus the waiting
model's last start; the 503 carries it as headers, and a job keeps it as
progress {state: waiting_gpu, held_by, eta_seconds}.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
flux evicted to make room for triposr was placed again two seconds later:
still on the wanted list, and first in the catalog, it took back the card
it had just given up, and the 3D job went on waiting. Its last replica
evicted, it leaves the list until it is asked for again.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A 3D job asked for straight after a picture, on a cluster with one card,
failed at once with "cannot be placed": flux held the card, too freshly
started to evict. The cards coming back when an on-demand model goes idle
makes it a wait, so the request is told to retry and the job waits.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
LLMGWPrepareForMesh, llmgw's own ComfyUI node, written by the gate into
every replica's custom_nodes from the binary: TripoSR's own preparation
(tsr/utils.py) — the background taken away, by the picture's alpha where
it has one and by rembg (MIT; u2net, Apache 2.0) where it does not, the
object cropped to its bounds, set square at a ratio of the side (0.85),
on grey. A picture sent as it was had a white or black shell round the
object, whatever was behind it made into the mesh.
/v1/models listed only models with a replica up, so one that loads when
asked — triposr, flux-schnell — was invisible to the clients that would
have asked for it. Those are listed now, with loaded: false; every entry
says loaded.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
/v1/openapi.json is the API clients use — OpenAI's endpoints, Anthropic's,
and what llmgw adds — in OpenAPI 3.1, generated from internal/clientapi's
route table by the generator the admin API's document comes from, now
told each document's title, server, token and error. A test reads the
router's mux.HandleFunc calls and refuses a route described and not
mounted, or mounted and not described. Asked for before there is a key,
so it needs none.
/docs/api shows it in Swagger UI 5.33.1 (Apache 2.0, served from the
gateway as the page's policy requires): Authorize takes a key, Try it out
sends to this gateway. Staff are offered the admin API's document too,
from /docs/api/admin.json, which nobody else is given.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The precise KV index and vLLM's LoRA affinity were built in v0.11.2 and
never run against a real vLLM; facility-aware placement in v0.11.2 less
temperature and cooling: marked [~], a mark now explained at the top of
the roadmap. Control-plane HA is done, by leader election on PostgreSQL
rather than Raft, and the admin UI changes things since staff roles
(v0.11.2) and the settings page (v0.14.1).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Tenancy, prefix affinity and benchmarking were built in v0.9 and after.
What phase 5 of llmgw-plan.md has open is precise KV routing proven on a
real vLLM, the block cost picker, defragmentation across NVLink domains
and prefill/decode disaggregation. Control-plane HA by Raft is left out:
HA is there already, by leader election on PostgreSQL.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The table said v0.4.0 was released, which has no tag of its own; the
changelog's table says, rightly, that it was released as part of v0.5.0.
The list of what came since v0.11.0 stopped at v0.14.1.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A finished job kept the last word heard from ComfyUI, the node running;
it is now said done before the job is marked completed.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
/3d beside /images, while some model makes meshes: a picture chosen, the
job asked for through the same path an API client's takes, its progress
shown (waiting for the card, n ahead; making it, the node and percent),
and the mesh in a viewer — model-viewer 4.4.0 (Apache 2.0, with three.js
and lit), served from the gateway since the page's policy loads scripts
from nowhere else — with a download and the model's licence. The mesh is
served to its owner from /3d/jobs/{id}/file; another project's person is
told there is no such job.
A job made from the portal acted for a key that does not exist: the
worker now rebuilds a session's principal from the account that made it,
while it is active and still in the project the job is charged to.
ctl 3d list|get|cancel, over /api/v1/client-jobs (/api/v1/jobs is the
control plane's own work, which ctl jobs shows).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
TripoSR's checkpoint loads with transformers 4, and ComfyUI runs 5: of a
package a node names, its range decides, and of everything else what
ComfyUI runs.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
G2: the gate keeps a websocket open to its ComfyUI under a client id of its
own, puts that id on every workflow it lets through — ComfyUI tells only
the client that queued a workflow how it goes — and keeps the last word on
each: the node running, its step of how many. It answers at
/progress/{prompt_id}. While a job's request waits, the router reads the
queue (queued, n ahead) and the gate (running, node, step, percent) and
writes it into the job, which a client reads back. Only into a job of the
same key: a tag naming another's job writes nothing.
G3, the rest of the plumbing: comfyui.nodes names custom nodes by https
repository and full commit, with the PyPI packages they need. The gate puts
each in the replica's own custom_nodes at that commit and installs the
packages beside the replica, constrained to the versions ComfyUI's
environment already has, which it never changes. The catalog refuses a
branch for a commit and anything but package specs for requirements. An
OBJ a node saves leaves the gateway as a GLB, colours kept, so every mesh
can carry its licence.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
G3, the part that needs no model: a ComfyUI model's comfyui.output (or
external.output) says pictures or meshes. A request for a 3D model — a
job at /v1/3d/generations — carries the picture as base64; at the replica
it is uploaded to ComfyUI's input directory (the gate lets /upload/image
through), the workflow is filled in with its name, the seed and the
resolution, and the mesh the workflow saves comes back with its type,
which is the file the job keeps. comfy.Client reads every list of files an
output names, not "images" only, and types a file by its extension where
ComfyUI sends octet-stream.
A model's license (name, url, notice) is said in /v1/models and in the
job, and written into a GLB's asset.extras, since the file travels without
the API; a mesh it cannot be written into is refused. per_mesh in
auth.pricing, usage.meshes in the rows, the sums and the export
(migration 0045). A picture asked of a mesh model, or a mesh of a picture
model, is refused saying which it makes.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The catalog had one source a model, so an embedding model served by vLLM
from its repository on a card and by llama.cpp from a GGUF on CPUs had its
entry swapped by hand halfway through the metallurgy import. A profile's
source is what its engine loads and placement looks for; the model's is
the default, and what updates and canaries move. ctl model add checks
every source, the GGUF check reads each llama.cpp profile's, and ctl
weights collects none a profile loads. An empty source leaves a profile's
fingerprint as it was, so running replicas do not read as stale.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A knowledge-base import sends far more than the cluster can take at once,
and every band waited the same 30 s. A band now has its own queue_ttl, so
one for bulk work waits behind people instead of being turned away.
And a request turned away because the cluster stayed busy — the queue
full (503) or waited as long as its band may (504) — now carries
Retry-After: 10. Without it llmgw-kb, which retries only what says when
to, counted each such passage as failed; that is what turned parts of the
metallurgy import away.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
POST /v1/3d/generations answers 202 with a job; GET says its status,
progress and result file; cancel stops it in the engine's queue too. The
request is kept as sent and replayed by whichever gateway holds the lease
through the path a client's request takes, so it is charged and judged
like any request (tagged job=<id>), and a model still loading is waited
out for as long as the job has. Migration 0044. Nothing makes 3D models
yet; the door refuses every model until the ComfyUI adapter learns meshes.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Section G of the backlog: a jobs table of its own with a leased worker,
progress read from ComfyUI's websocket by the gate, the ComfyUI adapter
widened from pictures to meshes with custom-node installs, and the pages.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
When notifications were last sent is not worth carrying: at worst one is
sent again an hour early.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each kept its own hour of quiet per event, so a budget spent on a team
both control planes served was announced by each. An event is now claimed
in the store before it is sent (notify_sent, migration 0043; one statement,
so two asking at once cannot both win), and only the claimer sends it. A
store that cannot be asked sends it anyway.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The cookie that carries a sign-in across the provider was signed with a
key each process made for itself, so with two control planes behind a
balancer a callback reaching the other one was refused as tampered — half
of all sign-ins. The key is now the store's oidc-state secret, sealed like
the others and the same on every control plane.
Changelog: the fixes since v0.14.2.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The five-minute, hourly and daily sums gain images (migration 0042), and
the CSV and JSON exports a column for them: once request rows were past
their retention, an export said what pictures cost but not how many there
were. The pictures made before are added to the sums from their rows.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Engines wrote to a pipe of the agent's. An agent that restarted and
adopted them left that pipe with no reader, and an engine's next write
past its buffer failed with EPIPE — in a Python engine, inside whatever it
was reporting on. Each process now writes to its own file in the agent's
work directory (0600), followed into the shard's log; an adopted engine's
file is read from the start, so its output from before the restart is in
its log too. A file past 64 MB is emptied in place; the files go with the
shard.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
ComfyUI wrote straight to the agent's pipe. An agent that restarts and
adopts the replica leaves that pipe with no reader, and ComfyUI's next
progress bar failed with a broken pipe, failing the workflow. It now
writes to a pipe of the gate's, which reads for as long as ComfyUI writes
and passes on what it can.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
engines.comfyui in the cluster's settings has every NVIDIA node build
ComfyUI: the release's source archive from GitHub, checked against the
digest the recipe holds — GitHub publishes none for a tag's archive, so a
version without one is refused — unpacked beside a Python environment
whose dependencies are its requirements file resolved as PyPI stood the
day after the release. torch is PyPI's CUDA 13 build, or PyTorch's CUDA
12.9 one on an older driver, and the install is checked on the card.
A requirements file that names anything but PyPI packages is refused.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A control plane that has just started or become leader has heard no
requests, and read the on-demand replicas its nodes report as wanted by
nobody: it drained every one of them at every failover. An adopted replica
of an on-demand model now gets a whole idle timeout from its adoption.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
ctl models said "not placed" for a model with min 0 woken by a request and
serving; it now says "ready (on demand)". And the docs: pictures on the
cluster's own cards, with engine comfyui behind its gate.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
ComfyUI lists its base directory's custom_nodes before anything else and
exits when there is none.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A model whose profiles run engine comfyui is placed on a node's card like
any other, and makes pictures there. ComfyUI asks for no key, so the agent
starts it behind llmgw comfyui-gate: ComfyUI on the loopback, the gate on
the replica's port demanding the replica's key and letting through only
what the router asks — queueing, waiting, fetching, cancelling. The gate
answers the agent's /health and /v1/models, shows ComfyUI the fetched
weights by their own names, and gives each replica its own base directory.
ComfyUI runs as the gate's child, so its card memory is the replica's.
A node offers the engine when agent.yaml names its ComfyUI
(engines.comfyui.dir). The model's comfyui: block names the workflow and,
for a source that is a set of files, the checkpoint.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
/images offers the models that make pictures, and the chat no longer
offers them. Like the chat, the browser holds a session and never a key;
the page's request is rebuilt from its own fields and sent down the path
every /v1/images/generations request takes. Size, how many, seed, steps
and a negative prompt; each picture shown with its seed, a download link
and a way to make it again. Nothing is kept but the usage. The menu
offers the page while some model makes pictures.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
auth.pricing takes per_image, charged on top of a prompt's tokens; a model
priced only per picture is not free, so a spent budget still stops it.
Each usage row keeps how many pictures were made (migration 0041), and
ctl request and the admin API's request view show them.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
ComfyUI saves the whole workflow — prompt, checkpoint, nodes — as PNG text
chunks in every picture. They are taken out before the picture is sent.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
An external model with api: comfyui is a ComfyUI server. A request for
pictures goes through the same key, allowlist, limit, admission and guard
checks as any other; at the replica it becomes a workflow — a built-in
(flux-schnell, flux-dev) or a JSON file in API format, with {{prompt}},
{{width}}, {{seed}} and the like filled in — queued, waited on, and its
pictures sent back as b64_json. A client that leaves takes its workflow
off the queue or interrupts it. ComfyUI is never scraped for metrics.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Qwen3-Embedding and models like it are trained with an instruction on the
query side only. The prefix goes before the question a knowledge-base
search embeds, at answer time and in the admin API's search; documents
reach /v1/embeddings from llmgw-kb as they are.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
It sat in localStorage under one fixed key, so the next person to sign in
on the same browser opened the chat on the last one's conversation. It is
now kept under the account's id, anyone else's is dropped when the page
loads, and signing out sends Clear-Site-Data: "storage" so nothing the
pages kept outlives the session.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
llmgw-kb sends many passages in one /v1/embeddings request. Their sizes
were added together and compared with the context, so a batch of passages
that each fit was refused as one over-long prompt. The longest input is
what has to fit; the total still counts for usage and limits.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The leading control plane tells the mirror over its API (PUT /v1/firewall,
kept as .firewall.json beside the models and applied by the same path unit
control planes use), and keeps the state in the store's cluster_state for
members that run no llmgw, which a root timer reads from the member's own
copy. The mirror also lets in the network of every node and control plane
the leader knows. Tried on the PoC: off, on and a network added, on every
kind of machine.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
llmgw_org, llmgw_team and llmgw_project rename and move in place;
llmgw_key holds the key in a sensitive secret and is made again when
revoked or expired outside Terraform; llmgw_membership puts a person in
the tree. Tried with OpenTofu against the fake admin API, drift and
import included.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
What llmgw org, keys and users did against the store on a control plane
can now be done remotely and by automation: the tree read, added to,
renamed, moved and emptied; keys issued (the key in the answer, once),
listed and revoked, with the scopes' lifetime limits; people put in and
taken out of the tree, their keys in projects they no longer reach
revoked with it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A search asking for more than 10 passages got at most 40 candidates from
HNSW, pgvector's default ef_search. The full-text arm scored every chunk
holding a word of the question before limiting, over the whole
organization's GIN index; it now takes candidates by ts_rank from the
collection's own partial index and scores those. New collections index
halfvec on pgvector 0.7 and later (up to 4000 dimensions).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each Patroni member gets a share of its own RAM in its own config.yml:
shared_buffers a quarter (15% beside a control plane or agent),
effective_cache_size, maintenance_work_mem and work_mem scaled to it, all
clamped. It is read again on every server init --patroni, so a machine
given more memory gets the use of it. A single-machine PostgreSQL init
installs is sized the same way, as a shared machine, instead of a fixed
96MB.
The step itself went in with 85025df by accident, in an early form that
restarted PostgreSQL before Patroni had taken the reload, leaving the old
shared_buffers with a restart pending. It now waits for Patroni to report
the restart pending, restarts with restart_pending (replicas first, the
leader last), and restarts a member it finds pending from an earlier run.
Applied to the PoC: pg01 978MB, ctl02 (control plane beside it) 587MB,
both from 128MB; streaming, lag 0, gateways unaffected.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
An adapter now has an owner (org, team, project, user or key): only callers
there see it, and to anyone else it is a model that does not exist, absent
from /v1/models too. A training job's owner defaults to the one collection
its pairs come from, must be able to read every collection they come from,
and pairs from a knowledge base are never made public. ctl train publish
keeps the GGUF or PEFT files on the mirror, which accepts uploads now, and
sets the owned adapter after the usual architecture check.
The knowledge bases' cache_salt follows the collections searched rather
than the organization, so two teams of one organization no longer share
cached passages.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
pgEnv passed sslmode=verify-full but dropped sslrootcert, so libpq looked
for ~/.postgresql/root.crt and both backup and restore failed against the
Patroni store. sslrootcert, sslcert and sslkey now go along as PGSSL*.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A job names the catalog model the adapter is for, the safetensors weights to
train on, the node and the MiB of a card it needs, and question/answer pairs.
The control plane sends it when the node's agent is there and nothing else
trains on it; the agent trains with its installed vLLM's torch and a pinned
overlay of peft and accelerate (no second environment), as the engines'
user, on the card with that much free — beside an engine, if it leaves that
much — converts the adapter to GGUF with the installed llama.cpp's own
conversion code when the model is served by llama.cpp, and sends what it
made back in pieces, kept in the store to download. Progress, loss and peak
card memory are reported as it goes; a node that goes fails its job.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A chat model of the cluster reads each chunk and writes questions a person
might ask of it with the answers it holds, streamed as NDJSON a chunk at a
time, for people to correct before an adapter is trained on it. A chunk that
gave no pairs says so on its own line.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
agents.default and agents.nodes.<id> in the cluster's settings hold
everything agent.yaml can say but the agent's id, control planes, address and
certificate, which stay in the file and are refused there. The control plane
sends each node its settings on register and, a second apart, when they
change; an agent whose settings changed checks they load beside its file,
keeps them and starts again on them, engines serving throughout. An edited
agent.yaml that loads restarts the agent too. History, audit, ctl config, the
console and the portal's settings page cover them as the rest of the
settings.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
/admin/settings edits the document llmgw ctl config edit opens: checked as a
control plane would load it, refused when made from a version since replaced
(the edit kept), its history viewed and rolled back as a new version, every
change audited by size and never by contents, since it can hold webhook
secrets. Administrators only; in English and Turkish.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A request given passages carries a cache_salt derived from its organization
when the replica chosen is vLLM, so how fast an answer starts cannot say what
another tenant's documents held. Other engines are not sent the field.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A Documents picker on /chat offers the collections the person may read in
the chosen project, the default when there is one, or none; the sources an
answer was given are listed under it. The page's choice becomes the
X-Llmgw-Knowledge header; one the page sent itself is dropped. In English
and Turkish.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
On an organization, team, project or key, the narrowest winning: llmgw ctl
kb default, POST /api/v1/kb/defaults, and the console's Knowledge screen. A
request's own X-Llmgw-Knowledge overrides it and none turns it off. A default
never refuses a request: a collection the caller cannot read is left out, and
a failed search answers without passages. Removing a collection takes it out
of every default.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A request naming collections in X-Llmgw-Knowledge has its last question
embedded on the caller's account, the best passages found and added to the
system message, numbered, with an instruction to cite them and to say when
they do not hold the answer; x-llmgw-sources says which were given. Done
before the guardrails judge the request, never served from the answer cache,
and a collection the key may not read is answered as one that does not
exist. Metrics llmgw_kb_retrievals_total and llmgw_kb_retrieval_seconds.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The best text match took the whole text weight, however little of the
question it matched: one everyday word ("most") put the remote-work page
above the travel one for "the most I can spend on a hotel". Now a chunk
gains the weight in proportion to the share of the question's words it
holds, counted once whatever their case.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Fused by rank, a full-text match on a question's everyday words outvoted a
far nearer meaning: on the PoC, "how much leave do fathers get" found the
annual-leave page before the paternity one (cosine distance 0.44 against
0.25). Now a chunk scores its similarity to the question plus up to 0.3 for
a full-text match, each search brings four times k candidates, the text
search drops Turkish and English stop words, and an operator's search says
when the embedding model did not answer rather than quietly searching text.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Collections listed with owner, model, documents and grants; A makes one; on a
row its documents, grant, revoke, a search to try, delete. docs/knowledge.md
says what administrators do and how people use llmgw-kb on Windows, macOS
and Linux: what it reads, what leaves their machine, where its ledger is.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Collections made for an organization, team, project or person, their
dimension asked of the embedding model; granted and revoked by path, never
outside their organization; their documents listed and removed; and a search
an operator can try. In the OpenAPI document.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Its own binary for Windows, macOS and Linux, carrying no server code. It reads
the files there, cuts them, has the gateway embed them with the person's key
and sends text and vectors for what changed only: a SQLite ledger on the
machine remembers each document's file, size, time, digest and what happened,
so a rerun reads nothing unchanged, a file skipped stays skipped until
--retry, and --delete removes only what this machine sent under the paths
given. --dry-run says what would leave the machine. Commands: login,
collections, sync, status, history, formats.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
What llmgw-kb, run by a person with their own key, reads and writes. A
collection the key may not read answers as one that does not exist; writing
needs ownership or a write grant on the key's chain; a personal key names its
project in X-Llmgw-Project and must be a member of it. Changes are audited by
document name.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
PDF, Word, Excel, PowerPoint, OpenDocument, RTF, EPUB, HTML, MHTML, mail with
its attachments, notebooks, subtitles, zip archives and text and source files
in any common encoding, in pure Go; legacy Office formats through LibreOffice
and scanned pages through Tesseract when they are installed. Chunks of about
500 tokens under their headings, overlapping by a paragraph.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Collections owned by an organization, team, project or person; grants that
add up down the tree and never cross an organization; chunks in PostgreSQL
with pgvector, partitioned by organization, an HNSW index per collection, and
row-level security as a second line behind each query's filter. Search fuses
nearest-neighbour and full-text ranks. Off on SQLite and without pgvector,
saying why. server init and the HA database setup install and enable it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
judge ask showed Qwen3Guard's controversial verdict as "safe (Violent)"; it
now says controversial, which a policy's judge.controversial counts. The
stray-judge note called a judge no policy asks any more "served as a model
clients use"; a model with role: guard is now said to be a judge no policy
asks.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The replica was stopped and the pass went on to the other models before
placing its replacement, so a model waiting for room took the freed cards.
On the PoC a GPU judge waiting for the only card took it from
qwen3.8-35b-a3b-distill, restarted there after its definition changed,
until the judge was stopped by hand. With the waiting model first in the
catalog it happened one pass later. The replacement is now placed in the
same step that stops the old replica.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every replica's engine demands its own key, and the check sent none, so it
failed with a 401 on every engine in the cluster. The key is asked of the
admin API for the replica at --endpoint, as bench does, or given with --key.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Llama Guard 3 1B could not tell ordinary coding requests from harmful ones:
on the PoC it scored "continue" 0.51, "continue the code above" 0.89 and
"how do I kill all child processes in bash" 0.92 unsafe, against 0.97 to
1.00 for harmful requests, at Q4_K_M and Q8_0 alike, with its own template.
A site had "continue" refused. Qwen3Guard-Gen 0.6B gave the ordinary ones
0.15 at the most and the harmful ones 1.00, in 0.8 GB on CPUs.
- format qwen3guard: its prompt written out as its template writes it, for
requests and answers, and its Safe / Unsafe / Controversial verdict read;
controversial counts as the policy's judge.controversial says, safe by
default.
- judge.models: judges asked in turn, the first that answers deciding; a
failure is recorded against the judge that failed. guard judge use sets
the order, replacing categories none of the judges names.
- known judges (Qwen3Guard 0.6B/4B/8B, Llama Guard 1B/8B): guard judge
enable serves any of them, --gpu the 4B on the roomiest GPU class as a
standard model; the leader fetches and serves one a policy names that the
catalog lacks.
- server init installs qwen3guard-0.6b with the baseline's categories.
- a judge reads a long text by its start and end, which fit its context,
rather than overflowing it and refusing the request under fail: closed.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every model of a stopped system asks for no replica and so has what it asks
for. Until the control plane took the setting up, seconds later, start
--wait saw 18 of 18 satisfied and returned at once, with nothing started.
It now counts only once the control plane says it is in the state asked
for, both ways.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The registration carries each shard's state as it is then. An adopted
engine is health-checked after that, and nothing reported the ready that
followed, so when the agent registered first the control plane kept the
replica starting while it served. After llmgw restart on the PoC two
replicas stayed so, one the only replica of qwen2.5-coder-14b. The
session's watcher now reports every shard the registration called not
ready once it is.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
One standby that ssh could not log in to stopped the whole restart before
this control plane's own. It is now said and left, the rest restarts - the
leader last - and the command fails at the end naming what was not
restarted and how to do it there.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Issue #3 asked for commands to stop and start the system, and 05aa21c gave
only one machine's services. Now:
- llmgw stop sets serving.stopped in the cluster's settings, live: the
reconciler wants every model at no replicas, freeing every GPU, and the
gateway answers every POST under /v1/ with 503 system_stopped. The control
planes, agents and mirror keep running, and no model's definition changes.
- llmgw start clears it; every model is placed again as it asks.
- llmgw restart has every agent restart over its stream (RestartAgent), then
the other control planes over ssh, standbys before a leader, then this
one; engines keep serving.
- llmgw status is the cluster's summary, saying when it is stopped.
One machine's services moved to llmgw service status|start|stop|restart.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
There was no command to stop or start llmgw on a machine (issue #3). These
act on the llmgw units the machine has - mirror, server, agent - or those
named, in dependency order, saying what each did. stop leaves a node's
engines serving unless --engines.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
SetCompute on an engine adopted from an earlier agent called writeState,
which read the command line from sh.cmd - nil for a process this agent did
not start - and the agent crashed at every start on the PoC's nodes until
fixed. An adopted shard now keeps the command line its state file had.
engine-check looked for llama-server one directory too shallow; the agent
installs it under <version-variant>/llama-<build>/. It now looks at both,
preferring a CUDA build on a machine with NVIDIA GPUs. Run on cuda01: a
token computed on CUDA0, 25 of 25 layers offloaded.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Checked against the installed llama.cpp (b10991): at its default verbosity
it prints neither the devices it uses nor the layers it offloads, so the
start-up output said nothing and every replica's ON column stayed empty.
The agent now takes compute from nvidia-smi - the GPUs the shard's processes
hold memory on, and how much - or CPU, for every engine, falling back to it
when llama.cpp's output says nothing; adopted engines get it too, and each
session reports it once per shard as a compute-only status. engine-check
runs its GPU token at -lv 4, where llama.cpp does say where its layers went.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
It was sent once, with READY, and kept only in the leader's memory: a
failover or an agent restart lost it. The agent now keeps it with the
shard's state on disk (ShardSpec.Compute, set by Supervisor.SetCompute) and
reports it with each running shard when it registers (RunningShard.compute),
which adoption carries onto the replica.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The settling window ran from process start, so a standby that took over
after a minute placed at once, before the agents reconnected and reported
what they run: it started second copies of replicas it did not know of and
drained the healthy ones when they were adopted. Seen on the PoC: a leader
restart replaced llama-guard-3-1b and a qwen2.5-1.5b replica. Reconcile now
reopens the window on the first pass after standing by.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
/forgot mails a link to set a new password when the cluster's mail settings
say how - SMTP relay, sender, and the public base_url the link points at,
never the request's Host - and says the same whether or not the address has
an account. /reset sets it once, closing every session. Links are kept by
their hash in password_resets, good for an hour, a newer one replacing an
older. Without mail, /forgot says to ask an operator, and llmgw users
reset-link makes the link to hand over. Single sign-on accounts get none.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
doctor sums findings up as two readinesses - the node's, which decides
whether it joins, and multi-node, its fabrics for replicas spanning nodes -
prints fabric findings in a section of their own, and gives both in --json.
node add says the multi-node one. NodeView.multi_node says whether a node's
declared fabrics are up, naming adapters that are down, and ctl nodes shows
it beside the fabric.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A model's replicas restarted one at a time, always. RestartPace says how
many restart at once: MaxSurge beside themselves (replacement first),
MaxUnavailable stopped first (no room needed, out while loading); the
reconciler begins that many per model per pass. The admin API takes
max_surge and max_unavailable on POST /api/v1/restart.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
NodeView carries engine_versions and compute_capabilities; ctl nodes shows
llamacpp b10991 and 6000-24g (7.5). A profile's engine_version restricts it
to nodes whose engine version starts with it, and explain-placement says
which version a node has, or which engines it has when it lacks the one
asked for.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Seeing a device is not computing on it. For llama.cpp on a machine with
NVIDIA GPUs, engine-check now loads a small GGUF (--model, or the smallest in
the weights cache) with every layer on the GPU, asks for one token, and
reads from llama.cpp's own output where the layers went: a build that
computes on the CPU fails, naming the fix. With no --bin it checks the build
the agent installed under /opt/llmgw/engines.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
llama.cpp's start-up output is read as a shard becomes ready: the devices it
used, its backend, and the layers it offloaded. That goes up with READY in a
new ShardStatus.compute, is kept on the shard, and is shown by the admin
API, ctl replicas (ON) and the console (COMPUTE). A shard given GPUs whose
engine says it found no usable GPU or offloaded no layer is stopped and
failed, even where nvidia-smi cannot be asked.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
guard.enabled (unset is on) makes every policy check nothing while false,
live: llmgw ctl guard off/on set it in the cluster's settings, and ctl
status, ctl guard list and the guardrails page say when they are off.
A refusal by a judge names the hazard it saw - judge: S6,Specialized
Advice - and the verdict is kept with the request in a new usage column,
guard_rule, shown by ctl request, the request log and the guardrails page.
llmgw ctl guard judge ask has a judge read a text, past the cache, and shows
its raw answer and how it is read; the endpoint is kept out of the audit so
no text is recorded.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
llmgw ctl guard judge status lists each judge model the policies ask:
in the catalog, as a judge, replicas ready, its last failure to answer, and
what is wrong in words. enable serves the missing one through model serve -
Llama Guard 3 1B on CPUs, role guard, under the name the policies give it -
with --replace for a model of that name that is not a judge; the console has
it under Do. GET /api/v1/guard/judges, and the summary's judges, put the
problem on ctl status and the console's Overview.
A Llama Guard file served from the mirror page is added as a judge, under
the name the policies ask, rather than as a model clients use the policies
never find. ctl says plainly when the API answers with a page.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The cluster's settings in the store, in force when changed; the firewall
switch; llmgw wipe; the mirror's address one the nodes reach.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
server init and mirror init took the first non-loopback address for the
mirror's --base, Docker's 172.17.0.1 on a machine with Docker; nodes asked
their own bridge for every model and were refused. primaryAddr takes the
default route's source, else the first address on an interface that is not
a container's, bridge's or overlay's. node add has the node reach the
mirror's files during the checks and fails, naming the address and the fix,
when it cannot.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
firewall.enabled (default true) is a cluster setting, applied live: the
leader sends a Firewall message to every agent when it registers and when
the setting or firewall_networks changes, and the agent (root) takes llmgw's
nft table out or installs it. A control plane writes the state to
/var/lib/llmgw/firewall.json, which llmgw-firewall-sync.path applies as root
with `llmgw firewall sync`; server init and server add install the units,
`llmgw firewall sync --install` on one set up before. node add and server
add put no firewall on a machine while the setting is off.
`llmgw wipe` removes llmgw from a machine after listing what goes and asking
for its name: units, engines' processes, the nft table, /etc/llmgw,
/var/lib/llmgw, the agent's state, the sudo grant, the CA from the trust
store, keepalived's configuration when llmgw wrote it, and the users.
--models, --store (Patroni member, etcd, or the local database) and --all
take more. The console has it under Do.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Everything in server.yaml that is not one machine's own is now a versioned
document in the store (cluster_settings), the same on every control plane.
The first start against a store imports the file's cluster keys, with the
baseline guard policy when there is none; after that those keys in a file
are ignored, with a warning.
Each control plane polls the document's version and its own server.yaml's
hash. A change is validated and applied: guard live (Gate.Replace), anything
else by a graceful restart that systemd brings back, standbys first and the
leader 30s later. A version that does not load on a machine is not applied;
at start the newest one that loads is used, so a bad edit cannot keep every
control plane down.
llmgw ctl config show/get/set/unset/edit/history/rollback, the admin API
under /api/v1/settings, and the console's Settings screen, which layers the
store's document over the machine's file, write it. Commands beside the
control plane that read cluster keys (init, node add, server add, backup,
upgrade) read the layered configuration; init --bundle writes offline to
the store once it holds the settings.
A standby that becomes the leader re-reads models, engine choices and
facility holds from the store.
Retires appendGuardIfMissing: the import does it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The control plane's firewall answered only the networks it was installed
with, so a node on another subnet was set up, could not reach 7443 and
showed as lost. node add now admits the node's network (or its address,
when routed) into the local ruleset, keeping everything already there.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A binary installed by hand made ctl upgrade say this machine already ran
the release and leave the old server running. And a node had 15 seconds
to answer an upgrade the control plane gives five minutes, so a node that
was upgrading stopped the run.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
An upgrade replaces the binary and leaves server.yaml alone, so a control
plane installed before guardrails stayed unguarded. Run again, init now
appends the baseline guard section (and so fetches the judge) when the file
has no guard key at all, keeping the old file as server.yaml.bak-<time>.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A --context-length 32768 --ctx-pool 32768 model got llama.cpp's default four
slots of 8192, and refused what the gateway admitted; --parallel is now
what the pool holds of the context. max_output was 1024 for every model,
which cut an editor's agent off mid-file; it is now a quarter of the
context, from 1024 to 8192. --tools writes capabilities.tools, which the
Continue configuration now reads, as it reads max_output.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
It told Continue every model uses tools and nothing of its context: a model
whose calls come back as text stalled Continue's agent, and Continue sent
prompts far over a CPU replica's 4096 tokens, each refused. tool_use is now
written for a model whose catalog entry says tools, and the context length
with a fitting maxTokens whenever the catalog knows it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Run on the standby, it upgraded the leader first, then itself, and the
leadership moved twice.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A proxy left pointing at a standby showed an empty cluster: the portal and
the admin pages read the standby's own memory. A standby now passes its
client port to the leader's, all but /health, /ready and /leader, which
must describe the machine a load balancer is choosing; and its admin port
passes the dashboard as well as the API. The leader trusts the control
planes' X-Forwarded-For, so clients keep their own address.
The changelog's candidate section is "Unreleased" again: release notes
for an rc read that section, and v0.11.3-rc.1 got none.
The e2e wait for two endpoints is 30s: under the race detector on a busy
runner, 15s was missed once.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The console never mentioned them. A Guardrails view shows each policy with
its mode, fail behaviour and checks, whether its judge models have replicas
ready, where policies are attached (attach and detach from it), and what
they did in a window. The admin API gains GET /api/v1/guard/events, and GET
/api/v1/guard says what each policy does.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The change to read it each time left the constructor storing the address
itself, so the file read each time was one named 10.0.0.161.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
One node at a time was hours on a fleet of hundreds. Nodes now go one, then
two, four, up to --parallel at once, each wave waited on whole; a wave that
does not come back stops it. The other control plane is upgraded first over
ssh, so a leader restarted last hands over to a standby on the release.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
With the agent server closed rather than drained, the process could exit
before Run's release had reached the store, and the standby waited out the
lease. The lease is now given up first, synchronously, and nothing takes it
again after.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
server add records it after restarting the new control plane, which then
never named it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The console and ctl on a standby showed its own empty view: nodes and
replicas live in the leader's memory. A standby now passes /api/ to the
leader over the agent port, with the cluster's certificates both ways; the
leader answers only a certificate naming a control plane the store lists,
and only while it leads. Control plane certificates are issued for client
use as well, and server add reissues older ones.
An agent's reconnect backoff starts again after a session that worked, so
each failover it lives through no longer makes the next one slower.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
keepalived's file holds its password and is root's alone; the control plane
runs as llmgw and could not read it, so the CONTROL line never named the
floating address.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A control plane being stopped held every agent's stream for the thirty
seconds of a graceful shutdown, so the standby led within four seconds and
had no nodes for thirty. The agent server is now closed, not drained. A
control plane that loses its lease while it lives on lets its agents go,
rather than keeping nodes the new leader cannot reach.
ctl status and the console's Overview say which control plane leads, which
stand by, and whether the floating address is on the leader. The console
adds a control plane (Do → Add a control plane), asking for the machine and
the floating address. server add says when the mirror would refuse the new
control plane.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Conflicts with facility-aware placement (config, control, admin API routes,
store test tables) all keep both; the generated reference and OpenAPI
document regenerated.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Conflicts with the LoRA work: vLLM's arguments, the engine profile, router
metrics and the scheduler's request all keep both; agent.pb.go regenerated
from the merged proto; the two test fakes of vLLM given separate names.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
ctl federation add/list/remove/sync keeps peers in the store (migration
0029); their /v1/models are offered here as external models named
PEER/MODEL and kept in step, a model still named by a fallback kept
withdrawn. Peers are checked at /ready, timed, and taken out of routing
while down. Requests to a peer carry x-llmgw-hop and are refused (508) when
they come back to a cluster they passed through; residency and no_store
keep requests off peers; X-Request-Id and, with --send-tags, attribution
tags go along. ctl model fallbacks sets overflow to a peer. Console view,
metrics, admin API routes and docs.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Plan §7.5 tier 2, opt-in per profile with kv_events. The agent starts vLLM
publishing KV cache events to a loopback port it listens on (pure-Go zmq4,
msgpack in both of vLLM's encodings), with sha256_cbor hashing and a fixed
PYTHONHASHSEED; it mirrors each engine's blocks and sends net changes up the
existing stream, with a full copy on every new one and a reset on gaps,
restarts and clears. The control plane hands them to the router, for the
sending node's own replicas only. The router tokenizes through a replica's
/tokenize, hashes the tokens as vLLM does (checked against cbor2/hashlib
golden values), and scores replicas by the share of leading blocks held,
falling back to the approximate tree. New metrics beside
llmgw_prefix_decisions_total. Untested against a running vLLM.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Nodes say where they are with llmgw.io/site, room, row, rack and pdu labels,
from agent.yaml or `llmgw ctl label` (kept in the store, over the agent's).
Placement spreads a model's replicas across racks and PDUs, keeps a
multi-node replica within one domain when its profile says keep_within,
places nothing in a domain held with `llmgw ctl facility hold` (optionally
draining it), and refuses a shard that would take a domain past its power
cap from facility.budgets, counting declared, derived and measured draw.
Explain-placement names the domain, and keeps the rejections of a placed
replica; `ctl nodes` gains LOCATION and --domain; `ctl facility` lists the
domains; the console's Nodes view shows and sets location and holds.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A vLLM model can carry PEFT adapters, registered at start, or with
max_loaded_adapters held at most N per replica: the router loads an adapter
through vLLM's runtime LoRA API when a request first wants it, sends the
adapter's requests to a replica that holds it while that one is not
saturated, loads each cold adapter once, and unloads the least recently used
idle one to make room. llama.cpp keeps loading every adapter at start.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The sampler sent the key given for --target, which for the gateway is a
client's, so llama.cpp and TensorRT-LLM refused every read and the run
reported an idle engine. The engine key now comes from the admin API for the
replica at --metrics' address, or --metrics-key; unreadable metrics stop the
run before any load, and a step that read none shows why.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The router read SGLang replicas with vLLM's names, and SGLang was started
without --enable-metrics, so it served no /metrics: every scrape failed as
if unreachable and the replica was routed to blind. It now gets
--enable-metrics and a mapping of its own (sglang:num_running_reqs,
num_queue_reqs, token_usage, the token counters and num_requests_total, as
0.5.21 publishes them). engine-check asks python -m sglang.launch_server
--help instead of Python's own help, and SGLang is given --tp-size rather
than the --tp argparse only takes as an abbreviation.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Conflicts: server.go imports (both kept); agent.pb.go regenerated with buf
from the merged proto.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A bundle carries what a cluster would download — llmgw's release with its
signed checksums, every llama.cpp build per architecture with GitHub's
checksums, model files pinned and checked against Hugging Face's — into a
directory or a tar whose manifest sha256 is carried apart. Import checks
everything, the release against the signing key, before anything is put in
the mirror. With offline.mirror set, nodes install llama.cpp from the
mirror against checksums the control plane holds and sends over the agent
stream; releases default to the mirror for node add, ctl upgrade and agent
upgrades. server init and the setup form install from a bundle.
The Helm chart runs one control plane with SQLite on a claim or an external
PostgreSQL, Secrets for the agent certificate and admin token, Services,
Ingress and ServiceMonitor; deploy/container/Dockerfile builds its image.
GPU nodes stay outside Kubernetes, as docs/kubernetes.md explains.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every template and handler message in internal/web goes through a JSON
catalogue under internal/web/locales/<code>/, with English as the fallback.
The language is the person's choice (a cookie set by /_lang/{code}), else
Accept-Language, else landing.language. Tests refuse a key some language
lacks, a translation with other arguments than English, and template text
left in English.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Conflicts: docs.go (manual, OpenAPI and suites all kept), modelserve Job
(trace and eval both kept). The eval routes are added to the OpenAPI route
table, update and promote documented with their optional body, and the
document regenerated.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Conflicts: config imports and docgen sources (both kept); agent.pb.go
regenerated with buf from the merged proto (MIG devices and trace context).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A suite is a set of cases in YAML or JSON lines, each scored by exact,
contains, regex, a number within a tolerance, a JSON Schema subset, or a
judge model with a rubric; weighed into one score against a threshold.
internal/bench/eval runs one on the control plane through Router.Replay, as
the gateway's own requests, and keeps every run and its cases in the store
(migration 0028: eval_runs, eval_cases).
- llmgw ctl eval MODEL --suite FILE|NAME, --history, --run, --suites; exits
non-zero below the score. Suites by name from control.eval_dir, or the
two built in from deploy/eval (arithmetic, json-format).
- model update and promote take --eval SUITE --min-score N, or a model's
standing gate (ctl model gate, eval: in the catalog): a promote asks the
canary first and leaves it below the score; an update tries the new
weights as a canary given no share, then promotes or removes it.
- The console's model menu: ask a suite with its progress, its runs, its
gate.
- The audit middleware no longer hands a handler a body cut off at 64 KiB.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
terraform-provider-llmgw/ is a module of its own, built on
terraform-plugin-framework, so the gateway does not carry its dependencies.
llmgw_model is a model added at runtime: scaling, class, owner and adapters
change in place, anything else replaces it; the definition is kept as
written rather than as the catalog filled it in. llmgw_quota and
llmgw_guard_attachment keep the scope as written beside the ID it resolved
to, and import by that ID. Data sources llmgw_models and llmgw_nodes.
The client lists every endpoint it calls and refuses any other; a test
checks each, and every field it sends and reads, against
docs/admin-api.openapi.json, so a change to the admin API that breaks the
provider fails here. Acceptance tests run Terraform against an admin API in
memory when TF_ACC is set. CI vets and tests the module.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A reconcile pass that does something is a trace, with placement (where it
went, or every candidate turned down), scale-up/down, drain, stop, eviction,
preemption, reaping and restarts beneath it. A replica's start is a span open
from placement until it serves; the start command carries its context in a
new ControlMessage.trace field, and the agent continues the trace through
the weights fetch, engine launch, rendezvous and readiness wait. Stops,
drains, adoption and engine installs are traced on the node, model jobs are
a trace with a span per step linked to the replica's start, and agent.yaml
takes the same tracing block as server.yaml. Off, every span is the no-op one
and commands carry nothing.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Engine trtllm runs trtllm-serve on one node, through a launcher of
llmgw's that demands the replica's key and turns on the Prometheus
endpoint its queue and cache gauges are read from; the router reads it
at /prometheus/metrics. Brought by the operator (engines.trtllm.python),
not installed; checked by ctl engine-check trtllm. Written against the
documented CLI, server and metrics and tested with stand-ins: not yet
run on a datacenter GPU.
mirror import-ollama copies a model's GGUF, and a vision model's
projector, from Ollama's directory into the mirror, checked against the
digest Ollama named each blob by; mirror import does the same for any
local file. The console's Mirror view imports from Ollama with I.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Discovery reports a card in MIG mode as its instances instead of the whole
card, on both paths: nvidia-smi (-L names them, -q -x gives their memory,
ids and processes) and NVML (GetMigMode, GetMaxMigDeviceCount,
GetMigDeviceHandleByIndex, behind an interface so it is tested with a fake).
Each instance has its MIG UUID, its own memory, and a class from the card's
and the profile: a100-80g-mig-3g.40gb. Placement numbers instances
1000 + 100 x card + instance and every view shows them as card/migN.
A replica on MIG runs on exactly one instance: a process opens one MIG
device and instances have no NVLink between them, so tp, pp, gpus or nodes
above one on a MIG class are refused in catalog validation and again in
placement, and the planner offers one instance or none. The agent gives
the engine CUDA_VISIBLE_DEVICES=MIG-<uuid> (no Vulkan device variable on
MIG). Memory and processes are per instance; heat, power and wear the
card's; utilisation is left out rather than reported as zero. A card in MIG
mode with nothing carved stays out of placement with the reason, and
llmgw doctor reports each card's mode and instances and warns of the states
in which nothing can run. ctl nodes, ctl status, the console and the plan
count devices by class. docs/install.md has a MIG section.
Written against fakes of nvidia-smi and NVML output; not yet run on MIG
hardware.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
llmgw keys cache --project P --ttl 1h [--any] | --off, and Cache answers… on
the console's project menu. The same request — project, endpoint, canonical
model and the body without stream, stream_options, user, metadata, store and
service_tier, keys sorted — is answered from the gateway's memory for the
TTL. Only reproducible answers (temperature 0 or a seed; embeddings, rerank,
score) unless --any; only a whole 200 from the model asked for, not a
fallback's, a canary's, or one a guardrail changed.
Looked up after the limits and the input guardrails; the output policy
judges a reused answer; no_store and stored conversations bypass it.
Streamed answers are kept put back together and streamed back as chunks.
x-llmgw-cache: hit|miss|bypass, Age on a hit, Cache-Control: no-cache or
x-llmgw-cache: no to bypass. A hit is recorded with no tokens and no cost,
marked cache_hit (migration 0027) and shown by ctl request. An LRU bounded
by router.cache.max_bytes and max_entry_bytes; llmgw_cache_requests_total,
llmgw_cache_bytes, llmgw_cache_entries.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
adminapi.Routes lists every endpoint with its query parameters and Go values
of the bodies it reads and answers with; internal/docgen/openapi turns that
into docs/admin-api.openapi.json by reflection, the way encoding/json encodes
the types, with the fields' comments as descriptions. A test fails when an
endpoint is registered and not in the table, another when the committed
document is stale. The binary embeds it and serves it with the token.
Handlers that wrote map literals and decoded anonymous structs now use named
types, so the document can see them. GET /api/v1/models/{name} answers with a
model's whole definition as it is in force, which a tool managing it needs.
docs/admin-api.md: reaching the API, the token, what a request must carry.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The manual the binary carries, rendered on the client port: a sidebar of
the documents in the console's reading order (now one list, llmgw.Manual,
for both), heading anchors, links between documents kept in the portal and
links to files outside the binary sent to the repository, no script.
Public by default; landing.docs_sign_in sends visitors to sign in first.
docs/clients.md shows the official OpenAI and Anthropic SDKs against
llmgw in Python, TypeScript, Go and curl, with llmgw's headers; the
portal's index, the landing page and the account page link to it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Audit log and staff roles, OIDC and SCIM, key lifecycle, usage export,
notifications; server init writes baseline guardrails with two Llama Guard
judges failing closed, and keys of at most 90 days.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A starter model with no room failed init after ten minutes and left the
firewall and the judge uninstalled.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
server init (and the setup screens) write guard.default [baseline] —
secrets refused, cards masked, injection recorded, the judge refusing
S1-S4, S7, S9-S12 in prompts and answers, fail open after 5s — fetch the
judge into the mirror and run it on a node's CPU. --no-guard-model leaves
it out. The closing text names what init cannot do: TLS on the client
port and backup recipients.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The README's status stopped at v0.9 and listed gaps since closed; onboarding
still said there was no SSO and no roles.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
notify.webhooks sends signed JSON or Slack-compatible messages, queued,
retried, de-duplicated hourly; ctl notify test checks each webhook.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Found on the PoC: llama.cpp takes a conversation ending on the assistant's
turn as a reply to continue, so Llama Guard answered "Rome.safe" and every
answer judged was served unjudged.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Found on the PoC: a streamed list came out as "1. 2. 3." — each item cut at
one release and dropped at the next, which sent the engine's own events.
A property test now cuts an answer every way and checks it arrives whole.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A cluster set up before the rules moved into the shared configuration kept
them in each member's config.yml, which the TLS step never looked at — and
then said its remote rules were hostssl. Found on the PoC: plaintext
connections were still accepted.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Members' etcd speaks TLS to peers and clients with a certificate from the
cluster's authority and asks clients for one; Patroni reaches it with the
member's certificate and serves its API over TLS. A running plain-HTTP
cluster is moved with Patroni paused, so PostgreSQL serves throughout:
Patroni's lease-free keys are carried over to the new etcd.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A new internal/clientip says who sent a request and how: the connection's
address, scheme and Host, or through a configured proxy what it forwarded.
The base URL in pages and setup scripts, isHTTPS for cookies, and the
per-address limits of auth and the sign-in forms use it. /setup scripts are
no-store.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A wrong password from anywhere counted against the address, so ten locked
its owner out. Attempts are now counted per address and client, with a
looser bound per address alone. The sign-in form carries a token matched
by a cookie (login CSRF). POST /login and /signup are rationed per client
address, a hash waits at most ten seconds for a turn, and server init
writes signup_open: false.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Resolve now checks the owner's state for a personal key; the authenticator
treats an inactive owner as a revoked key. Suspension stops the keys within
the cache TTL, reinstatement brings them back.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Patroni's REST API accepted changes from anyone on port 8008, and
PostgreSQL ran with sslmode=disable between machines. Setup now keeps a
fourth secret for the API (carried to members of older clusters), issues
each member a certificate from the cluster's authority, writes
restapi.authentication into each member's config and reloads it, and
turns ssl on with hostssl remote rules through an authenticated PATCH of
the shared configuration. The DSN becomes sslmode=verify-full against the
authority's certificate whenever the cluster has TLS, so a control plane
without the authority's key reconnects too. server add authenticates its
PATCH and lets the new control plane in by hostssl. Tried on the e2e
cluster: unauthenticated PATCH 401, plaintext refused, replication
streaming, the control plane on verify-full.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The admin listener answered any local process: models added, nodes
upgraded, engine keys listed, the heap dumped. /api and /debug now demand
the token the server writes 0600 to its state directory on first start;
ctl, the console and node add send it (sudo, LLMGW_ADMIN_TOKEN or
--token-file). pprof is off unless debug_pprof. admin_listen off loopback
is refused unless admin_remote. Dashboard pages, /metrics and /health are
unchanged.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The supervisor drops to the run_as user (default llmgw-engine) when the
agent is root: its credential and groups, a 0700 home under the work
directory as the working directory, HOME and XDG_CACHE_HOME there. A node
without the user keeps engines as root and logs why. node add creates the
user in the video and render groups.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
env is an allowlist of tuning families less path, plugin, library, address
and credential names (NCCL_*_PLUGIN, CUDA_INJECTION64_PATH and the like
got through the old denylist); llama.cpp's LLAMA_ARG_* are checked as
flags. args refuse --trust-remote-code, --log-file, --slot-save-path,
--tool-parser-plugin, --middleware, --allowed-local-media-path and others
that read, write or load, and what the agent sets itself. Checked in the
catalog and again on the node before an engine starts.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Without agent_tls the control plane served plaintext h2c on every
interface, and any machine reaching it could register as any node. The
server now refuses to start that way unless agent_insecure: true, and then
binds loopback when agent_listen names only a port.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
trusted.pub now holds the release signing key, so self-upgrade, agent
upgrade and node onboarding refuse a release whose checksums are not signed
with it. Onboarding fetched the release on the node and compared it with a
checksum from the same server; the control plane now fetches it for the
node's architecture with release.Fetch, checks signature and checksum, and
hands the bytes over. Release addresses must be https (loopback excepted).
LLMGW_RELEASE_KEY overrides the key only in development builds. The CI
release job fails instead of publishing an unsigned release while a key is
built in.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
govulncheck found eleven standard-library vulnerabilities reachable from
this code at Go 1.26.3 (net/http, crypto/x509, os.Root, net/textproto); the
go directive raised to 1.26.6 fixes every one. make vuln runs govulncheck;
make check runs it with the tests and vet.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Five minutes at five-second steps is sixty samples on a chart a hundred and
thirty dots wide, and resampling by count drew them on its right half.
Points are now put in the column of their time, held across the columns
between samples, a series that began late starting where it began.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Monitor draws CPU, memory, GPUs, requests, tokens, latency, network and disk
as Braille charts over 5m/1h/24h/7d, then per node, card and model
sparklines; node and model pages have their own. History is read from
monitoring.prometheus through POST /api/v1/metrics/range, or without it
from what the console saw. The control plane serves the cluster's machines
for Prometheus http_sd at /prometheus/targets on its metrics port, and node
add installs prometheus-node-exporter.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The console starts a serve or update job over the admin API and stays,
instead of closing to follow the download. Jobs lists them with their
progress and steps; the sidebar counts those running. ctl model serve and
update take --detach; ctl jobs lists jobs and follows one.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Ranking what was free first made every file come out on the CPU on a cluster
whose one card was in use. The plan now ranks by hardware alone, free now
only breaking a tie; a model served onto cards that are busy is added and
waits for them, the job ending as waiting rather than failed.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The control plane reads the model's shape from one GGUF header or its
config.json and lists, for every file, the best way it runs here: one card,
several of a node's (by layer for llama.cpp, tensor-parallel for vLLM),
several nodes' over one fabric (vLLM, pipeline-parallel between them), or a
CPU node; what each card holds and whether it is free now. ctl model plan
prints it; model serve takes its file and way unless told, gives a GGUF on
GPUs a GPU profile, and with --fallbacks adds the next best ways. The
console shows it per file and asks for the way and the fallback.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
model serve takes --projector, fetched beside the weights and loaded by
llama.cpp with --mmproj, and --format safetensors with --gpu-class and --tp,
which fetches the weights with their config and tokenizer and runs them on
vLLM. The console offers both, F switching between GGUF and original
weights. The mirror view is one row per version with its files on Enter.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The mirror keeps where it fetched a file beside it, and reports that and the
checksum after it restarts. A model whose files the mirror cannot trace is
found on the registry: a repository holding a file of that name that has the
commit the mirror's directory is named after, with every file at the same
size, the most downloaded first. ctl model origin and the console set one by
hand, checked the same way. Commits compare shortened.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every document and a page for every command, built from their help, read
like info: topics, contents, find. ? opens the console's own page, which is
new and covers every screen.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A model served from the registry records its origin: repository, commit and
each file's checksum; one served before is worked out from the mirror's
record of where it fetched the file. Every mirror.update_check (6h) the
control plane asks whether those files changed there. ctl model updates
lists it, ctl model update fetches beside the old files, checks, swaps and
replaces replicas one at a time, ctl model rollback goes back. The console
shows it on Models and offers the same.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Serve a model is a view of the registry's GGUF models, trending or searched;
Enter on one lists its files, and the one chosen is served with
llmgw ctl model serve. A form of one field is applied by Enter on it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Accounts, API keys and Organizations views read the store and change it by
running the users, keys and org commands in-process, so their checks and
messages are the ones shown. A Mirror view lists what the mirror holds and
which models use it, downloads into it and removes from it. Models gain
adding from a definition file.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Enter on a row of Models, Nodes or Replicas opens what can be done to it,
as forms and questions over the admin API. A Quotas view sets and clears
capacity quotas. Do gains serving a model, upgrading the cluster,
restarting every replica and a backup.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The summary's GPU, fabric and idle tables were written to standard output
regardless of the writer they were given, so on a cluster with a GPU the
console printed the GPU table across its full screen every three seconds.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The console gains live views of nodes, GPUs, models, replicas, usage and
agent logs, and a settings tree over every key of server.yaml (agent.yaml on
a node), generated from the configuration structs: values with their
defaults and doc comments, edited in dialogs, saved only once the service's
own loader accepts them, comments kept. The generator now tells blocks apart
by package, which brings the TLS, engine and supervisor settings back into
the reference.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
server add gives another machine this llmgw, a certificate, the
configuration and store address, lets it into the store through Patroni, and
starts it; --vip keeps a floating address on the leader with keepalived,
following a new /leader endpoint. Nodes learn every control plane from the
leader and notice a dead one by HTTP/2 ping instead of after the kernel's
fifteen minutes. The setup, console, node and init screens are drawn full
screen in one frame: title bar, sidebar, bordered panel, key bar, dialogs.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The setup screens can reach a database machine by a join code instead of
ssh: they listen on 7444 and show a short code; llmgw join dials out,
checks the setup by fingerprint, admits itself with the code's secret, and
runs what the setup sends until it is done. A joined machine is one more
onboard.Remote, so checks and the Patroni install run over it unchanged.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
With nothing installed, sudo llmgw asks for one machine or three, takes how
to reach the other two (ssh agent, key or password, and a sudo password used
only for the install), checks every machine first, shows the plan and its
command line, and installs. On a control plane it opens a console with the
cluster's state and the next steps. Machines are reached with Go's ssh so a
password works; the store's etcd and access rules use this machine's routed
address rather than a client-facing name.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
init --patroni user@host ... installs PostgreSQL, Patroni and etcd on the
given machines over ssh, with this one as the third etcd vote, makes the role
and database, and moves the store in use into the cluster. Each step checks
first, so it changes nothing on a cluster that is up. For a store of several
hosts the control plane publishes each member's role, state and lag from
Patroni, a stopped member as missing, prints it in ctl status, and alerts on
it. Making an administrator on PostgreSQL sent a bool to an integer column;
it sends an integer now.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
deploy/ha documents PostgreSQL under Patroni with a three-member etcd, as
running on the reference cluster, and the address that follows the primary.
Backups read such an address as libpq host and port lists and prefer a
standby, rather than resolving the list as one name.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Requests and tokens by tenant, the shares failing and refused by a limit,
first-token p95, budgets spent this period and the day's heaviest users,
filtered by a tenant chooser. Installed beside the cluster dashboard.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Requests by outcome, tokens and first-token time per tenant, at the level
router.tenant_label names (team by default), and every budget's share spent
this period. Rules alert on budgets at 80% and 100%, tenants hitting limits,
failing, spiking past their weekly average, or waiting on first tokens.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Percentiles are read from the histograms kept with the usage sums and shown
per group on the usage tab and in the people page's panels. Requests refused
by a limit are recorded as 429s against their owner, free of charge, so a team
hitting its limits shows as such. Usage can be grouped by model.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Charts, totals and budgets summed the request table on every view. Usage is
now also summed in the writing transaction into five-minute, hourly and daily
tables per key, person, project and model, with latency and first-token
histograms. Charts read the coarsest tier their slices allow; totals and
budgets compose rows, slices, hours and days, and stay exact. Request records
can be given a retention without moving any total.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
An organization's members, limits and use are a panel of the people page,
where they can also be changed, so the administrator's menu loses its
Organizations entry and /org redirects there. Managers keep /org.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A team's members, quota and spend were on three pages that linked to each
other. They are three tabs of one page now, each at its old address, and
choosing anything in the tree opens its limits, capacity and use beside it,
editable in place.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A model with replicas on a card and on CPUs answered at either speed with no
way to tell which. The picker names where each model runs, each reply names
the hardware that served it, and every answer carries x-llmgw-hardware.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A card with something else on it when the agent started stayed out of
placement after that process let go of it, until the agent restarted. The
agent now sends what processes other than its own engines hold on each card,
and the control plane replaces the registration verdict with it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every minute the agent reads nvidia-smi's full report, its NVLink error
counters and the kernel's XID lines. They are published per card,
shown on the node page and the GPU map, charted in Grafana, and five
alerts watch them.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The cluster page draws every card coloured by utilisation, memory or
temperature, and lists replicas whose every card is under 5% busy.
llmgw_gpu_assigned says which replica, model and owner hold each card.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
They were read once, at agent start, and published unchanged since.
Each sample now carries utilisation, memory and bandwidth in use, the SM
clock and uncorrected ECC errors. The node page charts each card, the
cluster page the fleet by class, Grafana has a GPU row, and three GPU
alerts are added.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each engine has a list of builds, best first, from what its project
publishes and how each was compiled: llama.cpp CUDA 13.3, CUDA 12.8
(driver 570 for the GPUs it ships as PTX), then Vulkan, never the CPU
build on an NVIDIA node; vLLM CUDA 13, then CUDA 12.9, and for GPUs
older than every build of the version asked for, the newest built for
them (V100: 0.19.1). The node takes the first its hardware meets and
that installs and computes on the GPU, moving on when one does not, and
reports the version it runs. GPU processes are read from nvidia-smi's
full report, so Vulkan engines count as on their GPU.
Refs #2.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
tp on a llama.cpp GPU profile is read as gpus_per_node rather than
refused, and a Hugging Face repository page as a vLLM source as the
repository id, so models written that way are placed after the upgrade
instead of dropped from the catalog. Engines adopted after an upgrade
are checked for computing on their GPUs too.
Refs #2.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Read from vLLM's release pipeline: every x86 build from 0.20.0 has
kernels for compute 7.5 and up, the CUDA 12.9 one included, and 0.19.1
is the last built for 7.0. PyPI's build is CUDA 13 from 0.20, not 0.26.
A V100 asked for a newer vLLM is refused, naming 0.19.1 for
engine_overrides; the CUDA 12.9 fallback is for older drivers on
Turing and newer. The cu129 wheel is found among the release's assets,
since its manylinux tag differs between versions.
Refs #2.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A node that cannot run the CUDA 13 build installs the same version's
CUDA 12.9 build from vLLM's GitHub release, with PyTorch's cu129 torch;
the install computes on the GPU before it counts. The build can be
chosen per node — auto, cu13, cu129, none — on the cluster page, with
ctl engine-build, or over the admin API; the control plane keeps the
choice and tells the node at once.
Refs #2.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The explorer fetches a safetensors set with its config and tokenizer;
the mirror gives out a version's manifest only once nothing is
downloading and every shard its index names is there, with each file's
sha256, which nodes check; mirror add takes several URLs; a repository
page is refused as a source.
Refs #2 (item 9).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A taken address on PostgreSQL fell through to a generic error, because
the store looked for SQLite's "UNIQUE". It checks code 23505 too. The
done page tells someone whose address had an account to sign in or ask
for a reset, without saying which case it was; other failures give a
logged reference.
Refs #2 (item 6).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A RoCE port without a v2 GID and a missing ib_uverbs failed doctor, and
node add, whatever fabric used them. They are now failures only for a
declared fabric and warnings otherwise; node add passes its --fabric
list to doctor so those are checked as declared.
Refs #2 (item 5).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A restart with no room beside it waited the drain grace period before
stopping the old replica, and stopped it even when the replacement
would fit nowhere. It now stops at once when the replica alone blocks
its replacement, gives the restart up (saying why) when nothing would
fit, and ctl restart --in-place stops first on request.
Refs #2 (item 3).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
engine_overrides give nodes picked by GPU class, id or labels their own
engine versions; node add --disable-engine keeps one off a node; what
vLLM cannot do on a node is a warning rather than a refused node; the
CUDA requirement follows the version; a hand-named binary is no longer
replaced by an installed one.
Refs #2 (items 4 and 7).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The form offered no placement and named every GGUF profile cpu. It now
lists the fleet's GPU classes (the freest chosen) and the CPU, a GPU
count and parallel requests, names the profile from the placement, shows
what the model will run as, and refuses the CPU while GPUs are free
unless that is ticked.
Refs #2 (item 1).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
tp on a llama.cpp profile was a width nothing placed. gpus_per_node is
now the count: placement reserves that many cards on one node and the
engine splits layers across them. tp, pp and nodes are refused with
what to write instead.
Refs #2 (item 8).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
llama.cpp whose CUDA backend cannot load runs on the CPU without an
error. The CUDA build is now asked for its devices when installed, a
shard on GPUs must hold memory on each before it is ready, and
engine-check reports the devices. Shards get CUDA_DEVICE_ORDER=PCI_BUS_ID
so CUDA numbers GPUs as placement does.
Refs #2 (item 2).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
--lora with --lora-init-without-apply leaves an adapter's scale at 1,
and llama-server applies it at that scale to any request without a
"lora" list of its own: every plain request for the model was served
with its adapters on. Found live on qwen2.5-coder-7b-instruct. Each
adapter is now loaded with --lora-scaled FILE:0.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
llmgw ctl model adapters lists a model's adapters and adds or removes
them through POST /api/v1/models/{name}/adapters, which reads each new
file first and refuses one that is not an adapter for the model's
architecture, as the mirror page's card does.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
An adapters view in the explorer, filtered by the model a held model was
made from; GGUF files told apart as adapter or model by their first bytes;
held adapter files marked; and a model's card attaches and detaches held
adapters of its architecture, refusing models and mismatched architectures.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A model's adapters are fetched by each node with its weights and loaded
unapplied at start; a request naming an adapter reaches the model's
replicas with llama.cpp's per-request lora field. Adapter names share the
model namespace, are listed, and follow allowlists on their own. The
router's config seed and pushed catalog now build the name map in one place.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A catalog model with external {url, model, api_key_file} and no profiles
is published as an endpoint by the control plane, the key read from its
file each time; the router rewrites the request's model to the provider's
name and sends its key. It is never scraped nor stale, and nothing is
placed for it. Named in fallbacks, it serves when the cluster has no room.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The pool-priority filter kept the highest-priority pool however full it
was. It now takes the first pool, by priority, that is not saturated by
admission's own measure, and the preferred one when all are.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A total bound across bands (admission.queue_capacity_total, default half of
all bands' capacity together); reached, a request is not refused while a
lower band has requests queued, and the newest of the lowest band gets a 503
with Retry-After instead.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
ed25519 over checksums.txt: llmgw release keygen makes the pair, CI signs
with the RELEASE_SIGNING_KEY secret when it is set, and release.Install
refuses an unsigned or wrongly signed release when the build embeds a
trusted key (internal/release/trusted.pub, empty for now) or
LLMGW_RELEASE_KEY names one.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Candidates tried on the way to finding no room were listed as whose replicas
a placement would stop.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The quotas page gains a capacity section over the same API methods the CLI
uses; a model's card sets its class, owner, memory and cores, applying only
what changed. Headroom counts the replicas a model could get by preempting
what its class outranks, and the card and ctl model target use it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Collects rc.1 to rc.14: the mirror page in tabs with a card per model,
one line per record in every admin table, reserved CPU cores and required
memory, capacity classes and preemption, tenant capacity quotas, sizing by
measured throughput, and node sensor readings. Adds `llmgw upgrade` for the
mirror host, which `ctl upgrade` does not reach, and says so when it ends.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every agent samples, at most every ten seconds: nvidia-smi's temperature,
HBM temperature, power draw, enforced limit and throttle reasons per GPU;
hwmon's CPU package and board inlet temperatures; and CPU package power
from RAPL's energy counters. Sent with heartbeats, published as metrics
and shown on the node page; unmeasured sensors are absent, never zero.
Observe-only, step 0 of docs/facility-aware-placement.md, which now
describes the agent source and what these readings can and cannot say.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A design document for step 5 of resource allocation, deferred: location
labels and rack spreading, power budgets with estimated then measured draw,
thermal and cooling conditions with hysteresis, opt-in shedding through
preemption, pluggable facility sources, and a rollout that is useful at
each step and safe when data is missing.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
It said "stopping a drained replica without a replacement after=5m0s" when
it had been given the 30-second preemption grace.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Idle times are unknown after a control plane restart, which left the choice
between a 1-core and a 4-core replica to map order.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Candidates were added in order until the model fitted, and every one added
was stopped, including those tried first that freed room on the wrong node.
A set that works is now pruned to the replicas it needs.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Stored ctl bench results reach the reconciler: requests per replica past
the measured SLO concurrency scale a model up, a replica is taken back only
if the rest stay under 70% of theirs, and among profiles of equal priority
the most measured tokens per GPU or core is tried first. ctl model target
sizes a model for a throughput against the room there is.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A model has an owner (org, team or project). Quotas guarantee a scope
GPUs, cores and memory and may limit it; what a scope holds is summed from
its models' replicas and those beneath it. A replica within its owner's
guarantee is not taken for another scope; what is held beyond a guarantee
is borrowed and taken back for a scope within its own. A limit stops a
model before it starts, and headroom stops there too. ctl quota and ctl
model owner set them; metrics and a CapacityAtLimit alert report them.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A model's class (critical, standard, opportunistic; unset follows its
scaling) decides what may be stopped to make room for it: a lower class
for a higher, and within a class only what another grew into under load,
for a model below its minimum. Victims are tried on a copy of the
inventory and stopped only if they make the room. Preempted replicas are
drained with a short grace and their model is held back until the room is
used. ctl model resources and class move under ctl model.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A profile's cpu_cores is reserved out of its node's CPUs by placement, as
memory is, and becomes the engine's thread count. Agents report their CPU
count in registration and heartbeats. New CPU models must declare memory;
model views show reserved against measured memory with warnings, and
ctl model resources sets both for a stored model.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Headers never wrap, free text is clipped with the whole on hover, full-width
rows are only notes and forms. The cluster page hides columns empty on every
row and shortens node names; People shows two memberships and a count.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The replica count, its nodes and the mode selector sit side by side instead
of stacking in a fixed-width cell, and node names drop the prefix every node
shares, with the full name on hover.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Ready of wanted, how many are starting, and the nodes, from the fleet data
the card already reads, so the list answers it without a click.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The Held tab's model name opens a card: each replica with its node, state,
age, CPU and memory, the room for more found by dry-running the scheduler
on a copy of the inventory, and a form for the always-running and
under-load replica counts. An always-running count placement cannot meet
is refused with the reason. Headroom is also on the admin API.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
One long page made every visit wait on Hugging Face for a list it might not
show. Each tab now has its own address and does only its own work; only
Held reloads while a download runs, and a refused fetch returns to the form
it came from.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
CPU replicas start on half the machine's CPUs, not all of them: two busy
llama.cpp servers each spinning on all 8 vCPUs fell to 0.8 tok/s apiece,
where at 4 threads they keep 5.5. llamacpp.threads sets it per node.
Inference requests can be traced over OTLP/HTTP (tracing.endpoint), with
admission, schedule and upstream spans and traceparent sent to the engine.
Security: onboarding no longer puts the cluster root in the system trust
store and removes it where it was; the mirror's mutual-TLS API can be
limited to named certificates (--api-client); init.log no longer keeps the
administrator's password; a warning when auth is off on a public listen;
the recorded actor is sanitised and sudo-aware; Ray's dashboard is off, and
the remaining engine gaps are documented.
reconciler.go and agent.go are split by topic, the status mapping of each
failure is tested, and the plan's checklist is brought up to date.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A model's max never did anything: desiredFor returned the minimum, or one
for an on-demand model someone was asking for, whatever the queue. Six small
CPU nodes run out of slots quickly, and a model with room to grow stayed at
its minimum while requests waited.
The router already scrapes each engine's running and waiting requests. The
control plane now reads them per model on every tick, alongside when each
replica was last used, and the reconciler keeps how many replicas load has
added above a model's base. One or more requests waiting per ready replica,
held for a minute, adds one; ten minutes with nothing waiting removes one of
those again. It never goes above max or below the base, moves one step at a
time, and does not add while the last addition is still placing or starting
— on this fleet a replica is minutes of weights and gigabytes of memory, and
a spike should not leave a pile of replicas behind it. Decisions are logged
with the load behind them.
Nothing live changes until a model's max is raised above its min.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Adding a model took four steps and a YAML file — plan it, fetch it into the
mirror, write its catalog entry with a memory figure somebody estimated,
scale it — and whether it would load was found out at the end.
`llmgw ctl model serve OWNER/REPO` does it as one job in the control plane,
where the registry, the mirror, the catalog and the fleet already are:
- the registry: the GGUF file set matching --quant (Q4_K_M by default) or
--file, leaving out vision projectors and importance matrices, and naming
what the repository does have when nothing matches; gated repositories are
refused;
- the mirror: fetched at the repository's revision, with progress, and each
file's sha256 compared with what the registry published — the mirror is
the one place a corrupted download is caught once rather than loaded on
every node;
- the check: whether the nodes' llama.cpp loads the architecture, and the
memory it needs from its own header;
- the catalog, then a wait for a replica that stops early, with the reason,
when the model is parked or fits on no node;
- a smoke test: one short request to the replica, with its engine key, and
the answer and tokens per second.
ctl starts the job and prints each step as it changes; a download of hours
does not depend on the terminal. --on-demand adds the model without starting
a replica. Jobs are kept in memory until the control plane restarts.
The catalog name a repository suggests moved into the registry package, so
the mirror page and this agree on it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every upgrade so far has been a script of mine copying a binary to each
machine over ssh, restarting it and polling until it came back — which is
the job, but not something the product could do, and which needed ssh into
every node.
`sudo llmgw ctl upgrade VERSION` on the control plane does it. This machine
first: the release for its platform, checked against the release's
checksums, the old binary kept beside it, llmgw-server restarted, and a wait
until every node that was ready has reconnected. Then each node in turn: the
control plane tells its agent the version over the stream it already holds;
the agent fetches and checks the release itself, swaps it in with a rename,
answers, and exits with status 75 so systemd starts the new binary. Engines
keep running, as for any agent restart, and are adopted again.
A node is done when it is back on the new version, ready, with as many ready
replicas as before; one that is not within the timeout stops the run before
the next node, so a bad release reaches one machine. Nodes already on the
version are skipped, so running it again resumes.
The fetch-and-verify is one package now, shared with onboarding's idea of
what a release version looks like.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
SDKs and editors pace themselves from OpenAI's x-ratelimit-* headers, and
the gateway sent none: only Retry-After, and only on the refusal they would
have avoided. A limit can be set at several levels of the tenancy tree, each
counted in its own pool, so the headers report the one with least left —
the one that will refuse first — after this request's own charge, on
refusals and successes alike. The windows slide, so the reset figure is the
next minute boundary, when a whole window has moved on. A limit that is not
set is not advertised.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The serve form asked for weights plus half again plus two gigabytes. That
ignores the one thing that varies most: the KV cache. At the same 16k pool,
the models on the live mirror need from 0.44 GB of cache (Qwen2.5-1.5B, two
KV heads) to 5.25 GB (Gemma-2-9B, 256-wide heads) — so the flat guess
overcounted the small models, which on 7 GB nodes means refusing placements
that fit, and undercounted the ones with wide or many KV heads, which means
an engine killed for memory.
The GGUF header already carries the shape. The reader now also takes each
architecture's block_count, head counts, embedding length and key/value
lengths, reading past general.* until the tokenizer — and a header whose
tokenizer comes first and runs past the fetched bytes still returns what it
has rather than failing. The estimate is weights, plus the f16 KV cache for
the context pool (llama.cpp's --ctx-size, which --parallel divides rather
than multiplies), plus a gigabyte and five percent for the engine. It reuses
the header the servability check already fetched, and the form says what it
is made of. Checked against all twelve GGUF models on the live mirror, whose
shapes read back as their published architectures.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
"Why was this request slow?" had no answer. Nothing tied a request to the
replica that served it, the time admission held it went out in a response
header and was never kept, and the only per-request log lines were for
disconnects, early stream ends and upstream retries.
Every request now has an id: the client's X-Request-Id when it is a plain
string of reasonable length, otherwise one the gateway makes. It is set on
the response before anything else, so a request refused as malformed can be
named too, and forwarded to the engine. One access log line per request
carries it with the model asked for and served, the replica, status, time in
the queue, time to the first token, total time, tokens, and the key,
project and user.
Usage records gain the id, the replica, queue time and time to first token,
tried against a restored copy of the live store before being shipped: 591
rows kept, columns and index in place. `llmgw ctl request ID` reads one back
through the admin API.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A review of v0.10.4-rc.4, and a look at the live network, found:
- Alertmanager listening on every interface: from any node its API took
silences, with no login. Bound to loopback on the monitoring host, and the
monitoring README now installs it that way; Prometheus beside it is its only
client, and one Alertmanager needs no cluster port.
- Login and signup unbounded in memory. Each attempt runs argon2 with 64 MiB,
and nothing limited how many ran at once; the per-address throttle kept one
entry per address ever tried. Hashing now takes one of a few slots and
waits for one otherwise. The throttle expires entries whose window has
passed when it fills, and if it is still full it refuses new addresses:
evicting live ones instead would let a flood of made-up addresses reset the
throttle on the address actually being guessed.
- Metric labels taken from the request. A refused request counted the model
name it asked for, so any key holder could mint a series per request.
Requests are counted under the catalog's name, or "unknown".
- Two in code written this week. A SQLite backup made its snapshot with the
default mode in the listable backup directory, a full copy of the store
readable while the backup ran; it is now made in a private directory, and
the unit sets UMask=0077. And archive links were checked by name only, so
"a -> ." then "a/a/b -> ../.." passed and a file written through them landed
above the install directory, as root. Archives now unpack through os.Root,
and a link may only point at a file beside it — which every llama.cpp
b10991 archive, CPU and CUDA, was checked to follow.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The first backup on the live control plane was written, and the control
plane reported none: /var/backups/llmgw already existed, made by an earlier
move to PostgreSQL for its own dumps, owned by postgres and 0700. MkdirAll
leaves an existing directory as it is, so the archive landed where the
llmgw user could not list it, and llmgw_backup_last_success_timestamp_seconds
said 0.
Opening that directory up would have exposed the dumps already in it, so
backups get their own, /var/backups/llmgw-server. And `llmgw server backup`
now warns when the directory it wrote to is one the control plane cannot
list, rather than leaving that to be found from an alert.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
There were no backups: docs/install.md gave a pg_dump command, and nothing
ran it. The store holds every key, person, quota and usage record, the models
added at runtime, and now the secret each engine's key is derived from; the
certificate authority's key sits in /etc/llmgw/pki beside it. Losing the disk
lost all of it.
`llmgw server backup` writes both to one archive: pg_dump's custom format, or
a VACUUM INTO snapshot for SQLite, plus /etc/llmgw. The archive holds the CA's
key, so it is 0600; the directory is listable, so the control plane — not
root — can see when the newest was written. pg_dump is given the connection
through its environment, since a DSN carries the password and a process's
arguments are readable by every user. Written beside its name and renamed,
so an interrupted backup is never the newest; the newest 14 are kept, and
nothing the command did not write is ever removed.
llmgw-backup.timer runs it nightly and catches up after downtime. The control
plane publishes the newest archive's time, and BackupTooOld fires past 26
hours — or when there has never been one. The install guide now covers
restoring, and says to copy archives off the machine, since they share its
disk.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
CI's test job has failed intermittently — two of three runs on some days —
while every run here passed, including six repetitions on one CPU under the
race detector with eight packages at once. The difference is a shared,
loaded Linux runner, and the tests most exposed to it are the ones that
sleep a fixed time and then assert:
- admission: idle admission bounded at 50ms of wall clock against a 1ms
tick, and a queued request checked for after a 50ms sleep;
- scrape: a 10ms scrape timeout, which on a slow runner reports the engine
unreachable and then the violation a second time, in a test that counts
one;
- supervisor: log lines, three lines of output and a removed state file
each checked after a fixed sleep, and an exited worker bounded at 1.5s;
- agent: a heartbeat checked for after four intervals.
Each now waits for its condition up to a deadline of seconds, and the two
wall-clock bounds are set against what they exist to distinguish — the
admission queue's tick raised to ten seconds, the exited worker bounded
below the five-second grace it must not wait out — rather than against how
fast this laptop is.
Without the CI logs this cannot be proven to be the failure; these are the
tests that fail when the machine is slow, and they no longer do.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
On the live cluster kev-0.5b — an architecture the nodes' llama.cpp did not
know — was started 155 times in eight hours, failing the same way each time,
and phi-4, which fitted on no node, logged "no placement" 250 times in two
days. Neither showed up anywhere an operator looked: the backoff never gave
up, no metric said a model was failing or unplaceable, and the one alert for
a model with nothing serving could not fire, because its series was published
only while a replica was ready.
After five failures in a row a model is parked: not placed again until its
definition changes — a new source or profile — or `llmgw ctl retry` is run.
Status, the model list and the admin API show it with its last error. The
placement warning is logged when its reason changes or every ten minutes.
A request for a parked or unplaceable model is answered at once with why,
rather than "loading; retry shortly", which a client retries forever. Each
request still clears the previous placement result and triggers a fresh
attempt, so a model that fits once room is made is not refused on a stale
answer; the router waits for that attempt, not for its full activation
window.
Metrics now cover model health (desired, ready at 0 for what should run,
consecutive start failures, parked, unplaceable), engine installs per node,
leadership, and certificate expiry read from disk at each scrape, with alerts
for each. `ctl status` counts a model against what it should have now, so an
on-demand model that was asked for and fitted nowhere no longer reads as
satisfied.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
ctl accepted only 200, so the first real `ctl restart --all` printed
"Error: 202 Accepted" over a request that had queued nine restarts, all of
which then happened. Any 2xx is now a success, and its body is read.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every fix so far became its own patch release, and the natural alternative —
v0.10.3-a — sorts before v0.10.3 by SemVer's rules: a hyphen marks a
pre-release, so every tool that orders versions would take the fix for the
older one. A release candidate of the next version, v0.10.4-rc.1, sorts
after v0.10.3 and before v0.10.4, which is what it is.
CI already builds any v* tag. Its release notes now come from a
"## Unreleased" section in CHANGELOG.md, where changes gather until the
version they belong to is released, and the release is marked as a
pre-release. Onboarding, in Go and in the shell script, counts a candidate
as a release and downloads it rather than copying the operator's binary.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
There was no way to restart a running replica. A running engine picks up an
engine key, an edited profile or a new engine version only when it starts,
so after v0.10.3 the nine replicas already running stayed without a key —
reachable from any node — with nothing to do about it short of draining
nodes, which on a small CPU fleet has nowhere to move a model to.
`llmgw ctl restart REPLICA... | --model M | --all` queues replicas to be
replaced. A replica being replaced keeps serving but stops counting toward
its model, so the reconciler places a replacement beside it; once that is
ready the old one is drained and stopped as a node drain would. Replicas of
one model go one at a time, so a model never needs the memory of every
replica twice. Where no second copy fits anywhere, the old replica is
stopped after the drain grace period and the new one starts in its place,
and the log says so, since that is the one case with an outage.
`ctl replicas` shows "(restart queued)" and "(restart replacing)" while it
happens.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Engines listen on every address their node has, with no key. Checked on the
live cluster: from a neighbouring node, a request straight to an engine's
port was answered — past the gateway's authentication, quotas and
accounting, by anyone who could reach the node.
Each replica's engine is now started with a key and refuses requests without
it. The key is derived from the replica id with an HMAC over a secret kept
in the store, rather than stored per replica: the router has to send the
right key to replicas this process never started — adopted after a restart,
or placed by the other control plane — and a function of a shared secret
gives that without a table to keep in step. A key taken from one node opens
that replica only. The secret is made on first use and copied with the rest
of the store when moving to PostgreSQL; a new one would lock the router out
of every running engine at the moment of the move.
The key reaches llama.cpp and vLLM through their environment
(LLAMA_API_KEY, VLLM_API_KEY) rather than their command line, which every
user on the node can read; SGLang only takes it as an argument. The agent's
readiness check, the router's proxy and its metrics scrape send it, the
client's own Authorization still never reaches the engine, and `ctl bench`
fetches the key of the replica it is pointed at from the admin API.
Engines started before this keep running without a key until their replica
is next started. With no store there is nowhere to keep the secret, and
engines start without one, with a warning.
Also formats two files that went in unformatted.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The admin listener has no login. Reaching it is meant to mean holding the
control plane, so it binds to loopback and an operator reaches it over an
ssh tunnel — and so does every page open in that operator's browser. The
API decoded JSON whatever the Content-Type said, so a page could POST a
text/plain model definition, no preflight needed. A model's profile env
reached an engine the agent runs as root, unfiltered: LD_PRELOAD in a
definition was code execution on every node. A page that rebound its own
name to 127.0.0.1 could also read pprof's heap, which holds API keys.
The admin listener now answers only when Host is an address or localhost,
refuses requests a browser marks as cross-site (Origin, Sec-Fetch-Site),
and takes changes only as application/json, which no page can send across
origins without a preflight nobody grants. ctl sends that on every change,
bodiless DELETEs included; curl and ctl are otherwise unaffected.
Profile env that changes what runs rather than how — the dynamic loader's
LD_*, PATH, Python's PYTHON* start-up variables, BASH_ENV and their like —
is refused when a model is defined, and dropped again by the engine adapter
before the process starts, whatever sent the profile. None of the live
catalog's models sets any.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Checked on the live cluster: from any node, the mirror's management API on
:8082 answered — list, fetch, delete — with no credentials. Unset
--api-allow meant anyone. It now means this machine only; the control plane
is let in with --api-allow, which `server init` already wrote for the
mirrors it sets up, or by mutual TLS.
A fetch went to whatever URL it was given, and the file server on :8081
then handed the result to anyone: "fetch http://169.254.169.254/..." read a
cloud's instance metadata through the mirror, and a URL on the LAN read
whatever answered there. Downloads started through the API now go to http
and https only, and are refused at dial time — after resolution, on every
redirect — when the address is loopback, private, link-local, CGNAT or
multicast. --allow-private-sources lets them through for a mirror that
really does copy from a server on its own network. `llmgw mirror add`, run
by someone on the mirror itself, is unchanged.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Two places built a path from a name somebody else supplied and checked only
for slashes.
The mirror refused "a/b" as a model name but took "..", which is a single
path element too: a fetch for model ".." wrote beside the mirror root rather
than inside it. Model, version and file names now have to be one element
that does not start with a dot — the mirror's own checksum files are dotted
— on the fetch path and the API alike.
The agent took the engine version from the control plane as given and made
a directory of it under /opt/llmgw/engines, and a RemoveAll before replacing
one. A version of "../x" would have done both outside the install root, as
root. The agent now installs only versions that look like one — letters,
digits and ._+- — whatever the control plane sent.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Once server init's screen closed, the whole run was printed again under the
frame that had just shown it. Now the password and the next steps are, the
whole run goes to /var/lib/llmgw/init.log (0600: it holds the password), and
it is printed in full only when init fails. Seen in a draft of the install
video, where it scrolled the screen away.
The spinner and the text field are written here instead of taken from
bubbles, whose packages query the terminal's colours as they load. bubbletea
does the same deliberately, so the query stays; the dependency goes.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
mirror init sets a machine up as the model mirror — the user, /srv/models,
and a unit serving files on :8081 and the management API on :8082 — and
server init --mirror points the control plane at it instead of running one.
Before, that took a unit written by hand and an edit to server.yaml.
The management API can fetch and delete models and answered anyone who
could reach it. mirror serve --api-allow limits it to the addresses given
and the machine itself; mirror init passes the control plane's. The unit
text is shared with the one server init writes for a mirror beside it.
With --no-node the starter model is still fetched and catalogued, unpinned,
so it starts on the first node added, and the init form shows the mirror in
use instead of offering to run one.
Tried on three empty VMs: the mirror answered the control plane and refused
the node, and the starter model came up on the node when it was added.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
In a terminal and without --admin, server init shows a short form: the
administrator's email, the other names nodes will use (with the addresses
it found), SQLite or PostgreSQL, and the mirror, node and starter model.
Its own lines then fill in as a list — read from the writer init already
uses, so nothing in init changed — with downloads updating in place and the
password in a box until the end. The whole run, password included, is
printed again when the screen closes, so it is in the terminal's history.
The last screen says how to add a node and to back up the authority's key.
node add takes several addresses and adds them in turn, with a summary.
Suggested fabrics carry what the adapters report them to be, so a name
given for one node is used for the next on the same fabric without asking.
Tried on empty VMs: init from the form through to the starter model
serving, and two nodes added to that cluster in one command. The first
attempt at the form lost its keystrokes to the terminal colour query in a
pseudo-terminal that never answers it; a real terminal does.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
doctor checks a node alone, and every check can pass while two nodes still
cannot reach each other. ctl fabric-test A B asks B's agent for an
ib_write_bw server and A's for the client, adapter by adapter as NCCL pairs
them, one rail at a time, and grades each against its link's rate. It goes
through the agents' streams, so there is no SSH between nodes; the agent
installs perftest the first time, and names the RoCE v2 GID itself.
The result is read from perftest's table by its heading, not a column
number. Agents now list the requests they answer beyond the usual
(Register.features), and the control plane refuses a test on one too old to
answer instead of waiting minutes for it. The forwarding dispatcher carries
the new capability, and its test says so: a missing forward would have made
the command quietly unavailable in production.
Not yet run over real InfiniBand or RoCE; this cluster has neither.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A control plane of this release asks every node for vLLM. An agent of the
last one does not know how to install it, and answered "unsupported", so a
cluster half way through its upgrade showed every node as failing. Agents
now say which engines they can install (Register.installable_engines), and
are asked for no others; an agent that says nothing is asked for llama.cpp
alone, as it always was.
doctor reports the CPU's vector extensions, and warns when AVX2 is missing,
with the fix for a VM: the test node for node add was one, on QEMU's default
CPU model, where llama.cpp runs several times slower and nothing else looks
wrong. node add shows it, since it runs doctor.
The changelog's unreleased section was two releases' worth of sections one
after the other, with two lists of fixes and the upgrade notes in the
middle. It is one entry now: largest changes first, the fixes together, and
every upgrade note at the end, with vLLM's default install and how to keep
your own.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Adding a node was deploy/bootstrap/onboard-node.sh, which is not installed
anywhere and wanted the node's name, the control plane's address and a
dozen flags. It is now in the binary, run on the control plane:
sudo llmgw node add ubuntu@10.0.0.21
The node id is the machine's hostname, and the address its agent dials is
this machine's on the route to it, checked against the server certificate.
It installs llmgw there (the release for its architecture, fetched and
checked by the node; uname rather than dpkg, so RHEL works), runs doctor
there and stops on a problem before changing anything, issues and installs
the certificate, writes agent.yaml and the unit, starts the agent, and waits
for it to join and install llama.cpp. The authority is read first, so a
missing one is said before the node is touched.
The steps are events. In a terminal they draw a checklist (bubbletea) that
asks what a script cannot: which fabric each group of active adapters is on,
grouped by the subnet manager or RoCE network they report, and whether to go
on past a problem. Elsewhere, one line per step.
Found by running it on an empty VM: ssh's "Permanently added" warning on
stderr was read as the machine's hostname, so standard error is now kept
apart; and a node added again while cordoned never counted as joined, so
joining now means a heartbeat since the restart.
llmgw node remove forgets a lost node, which otherwise stays listed until
the control plane restarts. ctl nodes no longer lists vLLM as skipped on
every CPU node. The quick install, install and onboarding docs use it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
install.md explains every choice and is long for someone who wants the
commands. quick-install.md has only those, and what to check after each:
the binary, server init, onboarding GPU nodes (with doctor first and the
script fetched from the release tag, since it is not installed anywhere),
a model from the mirror's explorer, keys and editors, and Caddy. Every
button and path it names was checked against the pages. quickstart.md is
now said to be the from-source, one-workstation path it is; README and the
about page point at the new guide.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
vLLM was installed by hand and named in each agent's configuration. The
agent now installs it: a Python 3.12 environment built by uv, which is
itself a GitHub release checked against its checksum, from PyPI resolved as
it stood a day after that vLLM was published and installed with
--require-hashes. The resolved list is kept in the environment. The
environment is built relocatable under a scratch name and renamed into place
once it runs, like an archive.
The default is 0.30.0, whose flags were checked against the ones llmgw
builds. It is built for CUDA 13, so it needs driver 580 and a Turing GPU or
newer, and 25 GB free while it installs; a node lacking one says which, and a
node with no GPU skips it without being shown as a fault. gcc comes with it
because Triton compiles its launcher at the first model load. SGLang
installs the same way when the control plane names a version. llmgw doctor
reports whether vLLM can be installed on the node.
Checked by installing 0.30.0 this way on a test node: 7.7 GB, torch 2.13 for
CUDA 13.0, importing cleanly; the node has no GPU, so vLLM was not started.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The website was a file built by hand outside the repository. It is now a
template in internal/web: every gateway serves it at /about, without a key,
linked from the landing page's footer, and the published site is rendered
from the same template.
The fonts are vendored (Latin subsets, OFL) so the page sends nobody's
browser to Google. The proxy configurations it shows are copies of
deploy/proxy, kept equal by a test. A development build says so instead of
linking to release files that do not exist.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The changelog's unreleased section had stopped at refusing unloadable
weights; it now covers the rest: telling independent fabrics apart,
their subnet managers and maps, NVLink on every build and across nodes,
the multi-node launch fixes and vLLM's new default, doctor, node load,
editors and the Responses API, the chat and mirror pages, a person's own
use, the admin pages, the new metrics, and how to upgrade.
docs/fabrics.md is new, for a site with InfiniBand or NVLink: what
llmgw does with them, what each node and fabric needs first, declaring
fabrics, checking a node with doctor, what the pages and metrics show
and which to alert on, what an engine is given, and where to look when
a replica spanning nodes will not run. It says plainly that none of it
has yet run on such hardware.
install.md, node-onboarding.md, onboarding.md, the README and the
example agent configuration no longer say the portable binary cannot
see NVLink or that a fabric's name cannot be checked, and describe the
account page and connecting an editor.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The account page makes a key for Continue, Codex or Cline and shows
what the tool needs with the key, this gateway's address and a served
model in it: Continue's assistant file, written to its own place under
~/.continue/assistants so the person's config.yaml is not touched;
Codex's provider and profile, added to ~/.codex/config.toml without
changing its default model; and for Cline, whose settings live in VS
Code's secret storage, the four fields to paste. Continue and Codex get
a one-line command whose script, served without a session, holds no
secret - the key comes from the environment, never a URL - keeps a copy
of whatever it replaces, and leaves the rest of the file as it was.
POST /v1/responses translates the Responses API, which Codex speaks, to
chat completions and back: instructions, messages, function tools and
Codex's free-form custom tools, their calls and outputs, and usage, as a
response object or as its stream of typed events. It goes down the same
path as any request - quotas, admission, routing, usage. Server-side
tools and stored responses have no equivalent and are ignored.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
While a reply streams the Send button is Stop, which aborts the request
- and with it the one to the engine - keeping what had arrived. Under
each reply: its tokens, from the engine's own count (the page's turn now
asks for include_usage), tokens a second, time to the first token and in
all. A settings fold takes a system prompt, sent before every turn, and
a temperature, which the gateway holds to 0-2 whatever the page sends.
The conversation, the prompt and the temperature are kept in the
browser's storage, so a reload puts them back - the model's turns
rendered again by the server as when they came - and "New conversation"
clears them. The gateway still keeps no conversation.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
On a phone every admin page was wider than the screen - its tables set
the width - so the browser zoomed the whole page out until nothing on it
could be read. Under 760px a table now scrolls sideways in its own box,
charts stack, form rows fold, and the menus wrap; every page measured at
390px wide now fits.
The cluster page drew a Fabrics heading and its note on a cluster with
no fabric at all.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Fetch on a repository opened in the explorer posted a form, and the
page that came back had closed the repository and lost the list's place.
The page's script now sends it asking for JSON: the outcome appears
under the button, and only the mirror's own table is fetched again,
every five seconds while anything in it is downloading. Without script
the form posts as before.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
An account's "add to" picker and its button sit on one line; suspend
and remove, the two things done least and the two drawn in red on every
row, are in a menu behind "⋯"; and the box for a new project under
every team, and for a new team under every organization, is folded
behind "+ project" and "+ team" until wanted.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The usage and organization pages drew their choices as filled green
chips, the newer pages as outlined pills. They are all the outlined
pill now.
The NVLink page's GPU squares were styled as .chip, the class those
choices already used, which shrank the usage page's filters into
1.7rem squares wherever the stylesheet loaded. They are .gpu-chip.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
An empty launch_mode meant ray: a Ray head and joins on every node, a
dependency to install, and a path never checked against a real cluster.
It now means mp - every node runs vllm serve with its rank and the
leader's address, the workers headless - which is how vLLM documents a
multi-node launch today and needs nothing but vLLM. ray is still there
for a vLLM too old for mp's flags. The example agent configuration,
doctor and the reference say so.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A GB200 NVL72 rack or an NVLink Switch system is a scale-up island: its
GPUs reach each other over NVLink whatever node they are in. Each GPU
now reports its fabric from nvidia-smi -q - state, status, ClusterUUID,
CliqueId, bandwidth - at registration and once a minute after, since a
fabric manager restarting drops every card out at once; and each node
reports nvidia-imex and its channels, without which the domain's GPUs
cannot share memory across nodes. A single machine's NVSwitch, which
reports a cluster of zeros, is no domain.
A replica spanning nodes is placed inside one domain first, on nodes
whose every card is joined and whose IMEX is up, and its shards are
told NCCL_MNNVL_ENABLE=1; InfiniBand is the fallback.
/admin/nvlink draws each domain as a rack - the NVLink Switch band on
top, light running down a backplane, trays of GPU chips that glow when
joined, pulse red when not, turn amber on degraded bandwidth - and
lists what keeps a tray out. A node's page shows its GPU-by-GPU NVLink
matrix, each card's NUMA node and nearest adapter, and its domain and
IMEX. doctor checks joining, bandwidth, IMEX and its channel; metrics
llmgw_gpu_nvlink_joined and llmgw_imex_up; /api/v1/nvlink.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The control plane asks one node on each physical fabric to run
ibnetdiscover every five minutes and parses the text itself, so a
format the parser has not met is a control plane upgrade, not one on
every node. From the map it works out the switch tiers (leaf, spine,
core, by distance from the adapters), which adapter is which node's,
which cables trained below the rate the rest of their kind run at, and
whether the rails line up: every node's Nth adapter on its own leaf.
A page per fabric (/admin/fabrics) draws it: switches in their tiers,
adapters fanned out under their leaves, cables coloured by rate and
bundled, degraded ones marching red with a marker, and the cluster's
own adapters live - down ports glowing red, and particles running each
cable as fast as its adapter is sending and receiving, from the
counters. It zooms, pans, fades what a hovered switch or adapter does
not touch, opens a panel of detail on a click, colours by rail, finds a
node, and refreshes every fifteen seconds. The data is also at
/api/v1/fabrics/{id}/topology.
Placement reads the map: a replica spanning nodes takes nodes of one
leaf together when it can, and nodes whose cable trained short last.
Static assets are named by their content, as the stylesheet already
was.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A fabric's identity was its master subnet manager, which changes on a
failover: for the minute its ports took to agree, one fabric read as
two, and a name covering it as two separate fabrics. Each agent now
also lists every subnet manager on a fabric (saquery SMIR, through one
port per master, once a minute); managers listed together are one
fabric, named by the lowest GUID among them, so the identity survives
the master changing.
The same list says what a fabric can survive. The cluster page's new
subnet manager table, /api/v1/fabrics, doctor and metrics
(llmgw_fabric_master_sm, llmgw_fabric_subnet_managers,
llmgw_fabric_problems) show each fabric's master, priority and
standbys, and say when a fabric has no standby, its ports disagree on
the master, ports are up with no subnet manager (INIT), or the master
changed in the last hour. Every change of master a port reports is
kept, and a fabric's last day of them is listed.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A fabric was what agent.yaml called it. With several independent
InfiniBand fabrics at a site, two given one name would have had a
replica spread across both, starting and then hanging on its first
collective. Each active InfiniBand port now reports the master subnet
manager it sees (sminfo, asked once a minute), which every port on one
fabric shares and ports on separate fabrics cannot; a RoCE port reports
the IP network of its RoCE v2 address. Ports also carry their node and
port GUIDs.
Placement divides a declared fabric by that physical identity and forms
a replica on one part only; a node whose adapters for one name are on
two fabrics is refused, and a node that cannot say is kept apart from
those that can. /api/v1/fabrics and the cluster page list each physical
fabric with its subnet manager, and say when one name covers two
fabrics, two names one fabric, an adapter is up but undeclared, or a
node cannot be checked. doctor reports each port's subnet manager, and
an agent on a node with adapters installs infiniband-diags when sminfo
is missing.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
For machines nobody here can log into. It reads the GPUs and their
nvidia-smi topology, NVSwitch and the fabric manager, every InfiniBand
and RoCE port with its counters and RoCE v2 GID, the RDMA and
GPUDirect kernel modules, the agent's declared fabrics against the
adapters, the management interface NCCL will bootstrap over, the
agent unit's memlock limit and the engines installed - and says what
is wrong and how to fix it, most serious first, with the NCCL
environment a shard on each fabric would be given. --json writes every
fact it concluded from, to send back. It exits non-zero when something
fails.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The agent's defaults said eth0 before normalize looked for the interface
holding the management address, so it never looked: a node on ens3 or
bond0 had NCCL and Gloo bootstrap over an interface it does not have. It
is now found, with eth0 only as the last resort.
The agent unit sets LimitMEMLOCK=infinity. RDMA pins what it registers,
and under systemd's 8 MB default an engine without the capability that
lets root ignore the limit fails its first collective over InfiniBand.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
GPUs were chosen by growing one card's NVLink peers, which took cards
each joined to that one but not to each other, and knew nothing of the
fabric. Every set of the right size is now scored - a node's handful of
cards makes that cheap - by NVLink pairs within it, then for a
multi-node shard by how near each card is to one of the fabric's
adapters (PIX before SYS), then by NUMA nodes spanned, with the lowest
indices breaking ties. A node without topology still gets the lowest
indices.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
vLLM's mp launch gave workers no rank and the leader's address as their
own VLLM_HOST_IP; every node now runs serve with --nnodes, --node-rank,
--master-addr and --master-port, and the workers --headless, as vLLM
documents it.
NCCL_SOCKET_IFNAME and GLOO_SOCKET_IFNAME were the control plane's one
interface name for every node; the agent already worked out its own -
the one holding its management address - and now uses it.
On RoCE, NCCL's default GID index is a v1 link-local entry on most hosts,
which does not route. The agent finds the RoCE v2 entry holding an IPv4
address and passes it as NCCL_IB_GID_INDEX; InfiniBand gets none.
A worker whose only process needs the leader - a Ray join - started
without waiting for it, since the wait was honoured for sidecars alone;
it found no head, exited and failed the replica. It now waits.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A port's state was read once, at registration: one that went down later
left its node on the fabric, offered multi-node work that would hang on
its first collective, until somebody restarted the agent. The agent now
re-reads every IB and RoCE port on each heartbeat - state, physical
state, rate and the error and data counters - and when the fabrics it
can carry traffic on change, says so once; the control plane replaces
the node's fabrics, which is what placement reads.
The ports are exported as llmgw_ib_port_active (with state and
phys_state, so an unplugged cable reads differently from a missing
subnet manager), llmgw_ib_port_rate_gbps and llmgw_ib_port_counter_total,
and a node's page lists them with the fabric each carries and any error
counter that is not zero.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The release binary is built without cgo, so without NVML, and NVML was
the only source of NVLink peers: on every NVLink machine the portable
agent reported none, and tensor-parallel groups were chosen by GPU index.
Discovery now reads `nvidia-smi topo -m` after either path, for the
NVLink peers (kept from NVML where it gave them), each card's NUMA node,
and how far each network adapter is from it - PIX, PXB, PHB, NODE or
SYS - which travels to the control plane as nic_paths for placement to
use. Both the NICn-with-legend layout of newer drivers and the
adapter-named columns of older ones are read.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Grouped by name, the four qwens in the mirror were four groups of one
and the table looked unchanged. A family is the letters a name starts
with - qwen2.5-coder-7b and qwen3-8b are qwen, llama-3.1-8b is llama -
and a family of two or more models gets a heading with how many and how
much disk; a family of one stays a plain row.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each model is a heading - how many versions, how much disk, and whether
one is downloading, failed or in use - with its versions beneath it.
Sorted by state alone, versions of one model were scattered down the
table around unrelated downloads. A model with a download or a failure
still comes first.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A bytes axis read "20.0BB" - compact's billion with a B after it - and
a quiet node's CPU axis read 0% all the way up. Sizes are now binary like
every other on these pages, a percent under ten keeps a decimal, and a
window of an hour is "the last hour" rather than "1h0m0s".
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The nodes are CPU machines, and nothing measured how busy one was: the
heartbeat carried GPU memory and disk, and llmgw_node_memory_bytes is the
total. The agent now reads /proc between heartbeats - CPU across all
cores, load averages, memory used and available, and each engine's
process group's cores and resident memory - and sends it with the
heartbeat. Nothing new runs on the nodes and Prometheus needs no new
target: the control plane exports it with the rest of the fleet, as
llmgw_node_cpu_utilization_percent, llmgw_node_cpus, llmgw_node_load,
llmgw_node_memory_used_bytes, llmgw_node_memory_available_bytes,
llmgw_shard_cpu_cores and llmgw_shard_memory_bytes. A measurement more
than a minute old is left out rather than repeated.
The cluster page's node table shows CPU and RAM as bars, and each node
links to a page of its own: current CPU, RAM, load and disk, the engines
on it with their cores and memory, and an hour of CPU, RAM, load,
per-model cores and memory and queue depth from Prometheus. It refreshes
in place.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Capacity, nodes, models and replicas come first, then a rule, then
traffic, people, what is being run and usage. The people and per-model
bar charts, which repeated the tables under them, are gone.
The page swaps its content every few seconds with htmx instead of
reloading, so a reader keeps their place, and it stops while the tab is
hidden. Without script the meta refresh still reloads it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The account page (was "Your keys") shows the last thirty days of the
person's own use - requests, tokens, spend and tokens per day - and, for
each place they can spend, what they used there and a bar for every
budget that applies: their own, a per-member one, or a project's, team's
or organization's shared with others, named as such. Bars turn amber past
four-fifths and red when spent. It sits under the keys now, above the
password form.
Spend is read from the usage table for the pool each limit counts
against, from the start of its period.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A reply streams as text as before; once it ends, the server renders it
with the model card's rules - raw HTML dropped, dangerous links emptied,
images as links - and the page places that. Headings, lists, tables and
code blocks read as they were written.
Maths outside code is set aside before the markdown, so its underscores
and asterisks are not taken for emphasis, and marked for the page to
typeset. KaTeX no longer searches the reply for delimiters, which took
"$5 and $10" for a formula. A reply that cannot be rendered stays
plain text, with the old search.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The cluster page's usage table showed session:usr-... and proj-... ids;
it now shows the person's email, whether it came through a key or the
chat page, and the project's path or "personal". Times read "3 hours
ago" rather than "3h7m25s ago", everywhere.
The usage page's legend gave each line's latest slice, which was 0
beside a table of hundreds; it gives the window's total, and the table's
numbers are right-aligned like everywhere else.
The PostgreSQL placeholder rewriting skips line comments. An apostrophe
in one opened a string that swallowed every placeholder after it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Clicking a repository in the explorer shows its card and files in a row
beneath it, fetched from /admin/mirror/repo, and a second click closes
it; the list stays where it was. The link still leads to the
repository's own page without script or with a modifier key.
A repository's files now take one line each, and an importance matrix
published beside the quantisations is marked as not a model rather than
offered for fetching.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Opening a repository shows its model card, rendered from markdown, with
its base model, license and languages above it; the listing names each
repository's base model and license under its name. The card is a
stranger's text: raw HTML and dangerous links are dropped, images are
shown as links rather than loaded, and relative links point into the
repository.
Adds github.com/yuin/goldmark for the markdown.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The mirror page lists Hugging Face's trending models without anything
typed, with views for what is new this month and what is most
downloaded, filters for task and size, and each repository's parameter
count, architecture, context length and age from one registry call.
"Only what these nodes can run" is on by default and leaves out
architectures the nodes' llama.cpp cannot load, and adapters, saying how
many it hid. Listings are remembered for ten minutes.
The stylesheet is now linked by a hash of its content. It is cached for
a day, so under a fixed name a deploy's new styles reached people a day
after its markup.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A message to a model that was not running waited 45 seconds and came
back as "the model is loading; retry shortly". The page now asks for
the model with /chat/wake, shows where it stands from
/chat/model-status - waiting for a node, downloading with bytes and a
percentage, starting the engine - and sends the message when it answers.
Also passes the nodes' llama.cpp build to the pages, which the mirror's
checks for loadable weights needed and were never given.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
kev-0.5b declared architecture "ggmlc", which llama.cpp does not know.
It was fetched, catalogued and placed, and failed on every node — while
the answer was in the first kilobytes of the file.
internal/gguf reads a GGUF header (v1-v3) up to general.architecture,
over HTTP with a Range request, and judges it against the architectures
llama.cpp b10991 lists in LLM_ARCH_NAMES (152, copied from the source at
that tag); a vision projector is not a model on its own. An unknown
build is not refused, only not vouched for, and a header that cannot be
fetched does not refuse anything.
Checked in three places: the registry listing (from the architecture
Hugging Face reports with expand[]=gguf, so one request per repository)
shows it and withholds the fetch button; a fetch reads the header and
refuses unless "fetch anyway"; and the model service refuses to add
such a model, which covers the serve form and ctl model add.
Tried on real files: the mirror's qwen2.5-1.5b (qwen2) and bge-m3
(bert), and Qwen2.5-0.5B on Hugging Face through its redirect.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The run mode sat after a tag repeating the model's name, so each selector started where the name ended and long names wrapped the row. The column is now Runs: one fixed-width selector per model, with the catalog's name above it only when it differs from the mirror's. Sizes no longer wrap.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every mirrored version was in use on the cluster, so the serve link — hidden for a version already served — appeared nowhere, and changing a model between always on and on demand needed the CLI. Each model using a file now shows how it runs and offers always on, on demand (15 minutes idle) or stopped, through the model service's Scale; models declared in server.yaml show their mode and where to change it. The serve form asks the same question.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
With every organization removed, /org answered an administrator with a 404 while the menu kept offering it. An administrator now gets a page saying there are none and where to make one, and the menu offers Organizations only when there is one to show (to a manager, one they manage).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A member of an organization can be its manager (users join --org X
--role manager, or the star on the people page). /org shows a manager
their organization: usage over time by team, project or person, its
teams and projects with limits and 30-day spend, and its members, each
once with every place they belong. Read-only; administrators see every
organization. Only what was spent in the organization's projects is
shown — UsageBucketsIn filters by where a request was spent.
Found on the way:
- Usage by org, team and project, and the quotas page's spend, joined
through api_keys and so dropped every chat-page request, which has no
key row. They now read usage.project_id; keys no longer move between
projects, so the two agree.
- /org was not exempt from API-key auth and answered browsers with a
JSON error. A router test now drives the session pages through the
middleware; it fails without the fix.
Membership queries, org-scoped usage and the reworked joins pass the
PostgreSQL store tests on 16.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A person was in exactly one project, chosen at approval, and could not
use the gateway before that. People belong to several teams and to
none, so:
- memberships: a person is a member of any number of organizations,
teams and projects, or of none; a member of a team or organization
may use every project in it. users join/leave, and the people page
adds and takes out from either side. Approval no longer needs a
project.
- personal keys: a key may have an owner and no project. /profile asks
what a new key spends against; keys create --for does the same.
auth.personal sets default limits for personal use or turns it off.
- limits for people: --scope user holds a person across every key,
project and the chat page; --per-member on an org, team or project
gives each member their own share there. Resolution folds the tree,
each per-member limit for the key's owner, and the owner's own, each
in its own pool; the quotas page shows and sets both.
- the chat page asks what to spend against and checks it per message.
- usage records who spent it, so a person's budget survives a restart,
personal use and chat included; restore no longer drops rows with no
project.
Leaving a project revokes the keys made there unless another membership
still reaches it. A personal key keeps no threads or notes, which are
per project, and is refused rather than dropped into the anonymous pool.
Migration 0010 turns each person's project into a membership, adds
usage.user_id and backfills it, and makes api_keys.project_id nullable.
SQLite cannot drop NOT NULL, so the table is rebuilt; that needs foreign
keys off, which only works outside a transaction, so migrations marked
for it run on one connection with foreign_key_check before commit.
Tested on a populated SQLite store with claims, and on a copy of the
cluster's PostgreSQL store: 5 people became members, 17 keys and 3
claims intact, 145 of 577 usage rows and all 69 chat rows attributed.
The PostgreSQL store tests, including new membership cases, pass on 16.
users move is gone; join and leave replace it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A branch with no limit of its own read "—" in every column while its
organization held it to sixty requests a minute. Each cell now shows
the strictest value from above, muted and marked with an arrow, naming
its source on hover and in the form's placeholders. A larger value set
under a stricter one says what caps it, for rate and token limits as
well as budgets; only budgets were checked before.
The table follows the people page: indented tree, tinted organizations,
spend with a meter that turns red past 90%, a form that opens and saves
without reloading (htmx, falling back to plain links), and the
explanations folded under "How limits combine".
Rows are built from the store's Tree, so a project sits under its team
by id. Matching the start of its path put projects under the wrong team
when two organizations shared a name.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The ranked bars said "8825 ms" beside a table that said "8.8 s". One formatter, humanLatency, now serves both, with a test.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- Numeric column headers are right-aligned like their values; the
stylesheet only aligned the cells, so "REQUESTS" sat nowhere near its
numbers in every table that has them.
- Ranked bars are an HTML grid instead of a fixed-width SVG: the value
had 96 units and "34.2k tok 114 req" did not fit, so the figure a
reader came for was the one cut off. Labels ellipsise in CSS and give
their full text on hover.
- Mean latency reads as 5 ms, 8.8 s or 1m 51s rather than 110695 ms.
- The "What the cluster is doing" charts sat between the Models heading
and the models table; they now come before it.
- A fleet without GPUs shows its CPU capacity — nodes ready, memory,
disk free for weights — instead of "No GPUs have been reported", and
a node with no GPUs shows "—" rather than 0/0.
- The profile's forms are a form's width rather than the page's, and
the quotas page's price warning is a boxed banner rather than a line
of red text.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The profile page rendered after a password change carries the CSRF
token set on that same response, but the menu's sign-out form was left
without one — and it is the first form on the page. The e2e test that
signs in, changes the password and uses the next form caught it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each page wrote its own nav: the same links in a different order, some
missing, the current page sometimes listed and sometimes not. A shared
"nav" template now draws it everywhere — Cluster, People, Quotas, Usage,
Mirror for an administrator, then Chat, Your keys and Sign out — with
the current page marked by aria-current. Pages rendered outside the
account pages (the dashboard, node logs) get it through the request
context their admin wrapper sets. The admin listener's own page keeps
its metrics/pprof/api links.
The chat and key pages said "project proj-3kq7…"; they now show the
project's path, and the key page no longer repeats the email.
A test renders several admin pages and requires the same menu on each.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
server init now also:
- runs a model mirror beside the control plane (/srv/models, bytes on
:8081 at an address nodes reach, management API on loopback so it
needs no certificate) and adds it to server.yaml;
- makes the machine a node of its own control plane: issues its agent
certificate, writes agent.yaml labelled as the control plane, starts
llmgw-agent, and waits for it to register and install llama.cpp;
- checks for NVIDIA GPUs with nvidia-smi, and reports a card on the PCI
bus whose driver does not answer rather than ignoring it;
- fetches Qwen2.5 0.5B (Q4_K_M) into the mirror and adds it to the
catalog pinned to this node: a llama.cpp profile per GPU class found,
with a CPU profile as the fallback, then waits for it to serve.
--no-mirror, --no-node and --no-model leave these out. Every step looks
before it acts, so a rerun finishes what an interrupted one started.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each held file took two table rows — seven columns and its address
beneath — so one sharded model filled the screen. A version is now one
line: model and version, the file or "N files", total size, and the
catalog models that use it. The address is behind a "copy address"
button (a set's is its directory, trailing slash included) and the
checksum is in the file's tooltip.
The model service reports which models name each source, so a file in
use shows who uses it and is not offered for serving a second time;
removing one warns that the catalog uses it. Downloads and failures
stay one line per file, since each has its own progress or retry. The
header counts versions, not files, and the nav calls /admin/users
"people" everywhere.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Nothing in the tree could be renamed, moved or removed, and an approved
person stayed in the project they were first put in — from the web or
the CLI. Now:
- /admin/users is the tree, each project opening onto its people and
keys, with approve, move, suspend, reinstate and remove for people,
revoke for keys, and add, rename, move and remove at every level,
plus a search over every account. Forms post and get the page back,
so it works with no script; htmx 2.0.11 (0BSD, embedded, checked
against npm's integrity hash) swaps only the changed part, keeping
open projects open.
- llmgw org tree/add/rename/move/remove, users move/delete, and
keys list --all.
- Projects, teams and organizations are named by id, path
(Acme/platform/gateway) or unique name; users by id or email. With a
single project --project can be left out, instead of defaulting to
"default", which a tree built from the web does not have.
Moving a person moves the keys they made; operator keys stay with the
project. Only empty things are removed, and never a project that had a
key, since usage history refers to it.
Also: keys init no longer reparents an existing team or project;
reinstating an account with no project is refused (a rejected signup
used to come back active and fail at its first key); rejecting a
signup removes it; keys limits --budget takes whole units like the
quotas page.
Checked in headless Chrome: a move from an open project updates the
page in place, keeps the project open, and the search still filters.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
server init --postgres failed on Rocky with postgresql-setup not found.
The PostgreSQL project's packages (PGDG) provide "postgresql-server",
so dnf can satisfy that name with postgresql16-server — whose unit is
postgresql-16.service, whose tools are in /usr/pgsql-16/bin and whose
setup script is postgresql-16-setup. The code assumed the distribution's
names for all three.
The server unit is now picked from systemd's postgresql* unit files, its
data directory read from the unit's own PGDATA, and initdb run directly
as postgres when that directory is empty, with password logins over TCP.
psql, initdb and the restart use what was found. The unit names and
PGDATA values in the tests are the ones Rocky's and PGDG's packages ship.
Not yet handled: a server initialised with ident for TCP logins, one on a
port other than 5432, and several PGDG versions where the highest is not
the one on 5432.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Installs PostgreSQL with the machine's own package manager when it has
none (initdb with password auth on the RHEL family, whose default is
ident), creates the llmgw role and database with a generated password
kept in /etc/llmgw/secrets/store-dsn, and checks the address connects.
On a machine already on SQLite it stops the gateway, copies the store
so accounts and keys survive, points auth.store at the new address and
starts it again. A move that fails starts the gateway on SQLite again.
server.yaml is edited as text so its comments survive, and kept beside
it. A PostgreSQL init did not install is not resized: it may serve
something else. It refuses to move one PostgreSQL store into another.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A configuration with no static endpoints and no models was refused as
having nothing to serve. That is every new cluster: under systemd the
unit restarted every two seconds and nothing ever listened on 8080.
With no endpoints a router-only process has nothing it could serve, so
such a configuration is now a control plane with an empty catalog.
Static endpoints without profiled models remain router-only.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
install.md said the client port belongs behind a reverse proxy and gave
nothing to put there. deploy/proxy/ has both, attached to every release.
What they get right is what a first attempt gets wrong: streamed replies
are not buffered, the read timeout outlasts a model's silence before its
first token (nginx's 60s default is a 504), the body limit matches the
router's, and X-Forwarded-Proto and -Host are set by the proxy rather
than passed through — llmgw marks session cookies Secure from the first
and builds the landing page's base URL from the second.
The Caddyfile was validated and run against a streaming upstream: events
arrive as sent, and forged forwarded headers are replaced. nginx uses
`listen ... http2` because `http2 on;` needs 1.25.1 and Ubuntu 22.04 and
Debian 12 ship older.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
`llmgw server init` does what install.md's steps 2 and 3 asked of the
operator: the llmgw user and directories, a certificate authority, the
control plane's certificate for every address the machine has, a
configuration, the default project, an administrator with a generated
password, and the systemd unit — then starts it and waits for /health.
Every step looks before it acts, so a second run changes nothing, and
--host later reissues the certificate with the old names kept. It never
overwrites server.yaml or an authority, and leaves alone certificates
issued by an authority kept on another machine.
The configuration install.md showed could not start: an empty catalog
without control.enabled is refused as having nothing to serve. The one
init writes sets it, and a test holds that.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Seeing what a node did meant ssh through a jump host, to a machine whose
address nobody remembers, to read a journal only root may open. When an
engine install failed on a new node today, the reason was one line — and it
was four steps away from the page that said the install had failed.
Each node's row on the cluster page now links to its logs: the agent's own,
and every engine running there, chosen from the list the node itself sends.
The page can follow, refreshing every five seconds, and is on both the admin
listener and /admin. `llmgw ctl logs NODE [--source SHARD] [-n N]` reads the
same.
They travel over the stream the agent already holds, so they need nothing
the control plane does not already have: a request goes down, the lines come
back, and a reply that does not come in ten seconds is an agent too old to
answer. The agent keeps its last 2000 lines in memory as well as writing them
wherever it did before; the engines' last few hundred lines were already kept
and sent only when one crashed.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Found by building a node from nothing. Its agent tried to install libgomp1
while cloud-init and unattended-upgrades were running apt, and failed with
"Unable to lock directory /var/lib/apt/lists/": DPkg::Lock::Timeout covers
dpkg's lock, not the one `apt-get update` takes on its lists. The node was
correctly kept out of placement and said why, and then sat idle for the ten
minutes until the next pass.
A package command that fails only on a held lock is now retried for up to
ten minutes, reported as "waiting for another apt-get to finish" rather than
as a failure. A pass that does fail is retried after a minute, backing off to
ten.
The weights directory is created when the agent starts rather than by the
first download, so a new node reports its disk from its first heartbeat.
Onboarding takes several control plane addresses, comma separated, so a new
node gets the same failover list as the rest.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The llama.cpp build is published from Ubuntu but is not Ubuntu's in any way
that matters: it needs glibc 2.34 and GCC 12's libstdc++, which Debian 12
and RHEL 10 have and RHEL 9 does not. Refusing every system not called
"ubuntu" turned away the first two for their name. The build is now tried
everywhere, and ldd decides.
What ldd says is made readable. "/lib64/libstdc++.so.6: version
`GLIBCXX_3.4.30' not found" is a system too old for the build, and is
reported as "libstdc++.so.6 is too old (needs GLIBCXX_3.4.30)", once —
rather than as a missing library that is plainly installed.
Packages are named per family — libgomp1 for Debian's, libgomp for Red
Hat's — and installed with whichever of apt-get, dnf, yum or zypper the
machine has, decided from ID_LIKE rather than a list of distribution names.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A node ran whatever llama-server somebody had once unpacked on it. Onboarding
checked an engine only when told its path, installed none, and a node with
nothing on it registered as ready. The one time that went wrong it went wrong
invisibly: the binary was there, executable, and could not start for want of
libgomp1, while the node reported healthy.
The control plane now says which engines every node has, at which version —
`engines:` in server.yaml, defaulting to the llama.cpp build this release was
checked against — and each agent converges on it. What only the node can
know, the node works out: the build for its processor and driver (CUDA with
its runtime when an NVIDIA driver is loaded, CPU otherwise), the OS packages
that build needs from its own package sources, and whether the result runs.
Every archive is checked against the sha256 GitHub publishes for it and
unpacked refusing any path or link that leaves the install directory. ldd
names a missing library outright; running `--version` proves the rest. Only
then is a directory marked installed, and it is checked again every ten
minutes, because an unrelated upgrade can remove a library.
An engine is reported only once installed and seen to run, or when the
agent's configuration names it on purpose — never because something happened
to be on PATH. The agent now says when it has probed, so an empty list means
"has none" rather than "too old to say": a node rebuilt from nothing is no
longer a candidate for every model at the moment it can run none of them.
Placement's refusal carries the node's own words, "installing: downloading
llama-b10991-bin-ubuntu-x64.tar.gz (16 MB)", and so do `ctl nodes` and the
cluster page, along with a warning for a CPU without AVX2 — the VM CPU type
that makes everything five times slower and nothing look wrong.
Agents also report their OS, architecture, vector extensions and NVIDIA
driver. Onboarding installs llmgw itself from the CI release for the node's
architecture, checked against the release's checksums, rather than copying
whatever binary the operator's machine had.
vLLM and SGLang are not installed yet: they are Python environments of
several gigabytes, not an archive, and still have to be named in the agent's
configuration.
Also fixes a race on the weights fingerprint, which two heartbeat loops touch
during a failover.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every release on Gitea was a list of binaries and nothing else. Gitea makes
an empty release for each pushed tag, and the workflow only ever attached
files to it, so what had changed was written in CHANGELOG.md and nowhere a
release was looked at.
A job after the build now copies the tag's section of the changelog onto its
release, and fails if there is none — a version nobody described should say
so where it is seen. Run by hand, it syncs every version the changelog has,
which is how the releases made before this get their notes. The changelog
stays the one place notes are written; this copies it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every agent file on the cluster named its own node, its own address, the
path to llama-server and two weight directories — different on every node,
kept true by hand, and not true: node05 carried "ram: 32g" beside a measured
31 GB, looked for weights in an empty directory while fetching them into
another, and the other four never looked in theirs at all, so four nodes
reported no disk. Growing a node's RAM would have gone unnoticed, because
memory was read once, at registration.
The agent now finds these out. Its id is the hostname, which is also the
certificate's common name. Its address is the one on the route to the
control plane, which the kernel picks for a UDP dial without sending
anything. llama-server is looked for on PATH and then where release archives
are unpacked. A missing nvidia-smi is a node without NVIDIA cards rather
than a failed discovery — which also cut discovery short before InfiniBand
on every CPU node not told to skip GPUs. Fetched weights land in
/var/lib/llmgw-weights by default, and wherever they land is always looked
in.
Memory is read on every heartbeat, and installed engines are probed again
each minute and sent when they change, so a VM given more RAM or a node
given vLLM becomes a candidate without anyone restarting its agent. The
control plane logs both changes, since "the node now qualifies" is the
answer to why a model placed an hour after it was asked for.
What is written still wins. The agent file shrinks to the control plane's
addresses and the TLS files, the same on every node; onboarding writes only
the overrides it was given. What cannot be found stays configuration: which
fabric an adapter is on, because OpenSM's default prefix is the same on all
of them.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The picker listed what had an endpoint, and a model at min 0 has none until
somebody asks for it. On this page a message is the asking, so an on-demand
model could never be started from here: mistral-nemo-12b-instruct sat in the
catalog, served by nothing, and absent from the one place someone went to
try it.
Such a model is now offered, marked "not loaded", with a note saying the
first message fetches the weights and starts the engine, and that until it
is up the reply is "the model is loading". A model stopped outright — max 0 —
is still left out, because nothing would load it and offering it would only
produce an error.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A report nobody can reach from the page they land on is a report nobody
reads. It sits beside accounts, on the client port only — the admin
listener's copy of this page serves pprof and has no session to authorize
anybody.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Usage could be read as a total — per key, per user, per model, per group —
and the cluster's own charts covered the fleet's hardware. What nobody could
see was when a group used it. "Research doubled its traffic three weeks ago"
is not a question a column of totals can be asked, and it is the question
somebody asks before a quota conversation.
/admin/usage draws requests, tokens and spend over a chosen window, grouped
by any level of the tenancy tree or by person. Person is not a level of that
tree — one person holds keys in several projects — so the grouping is its
own type rather than store.Scope with a fourth value nothing else could use.
Three things learned from the charts already here. Buckets are gap-filled
per group, because a line drawn straight across an absence claims traffic
that did not happen, and because the chart sums its tail by timestamp and
can only do that if every series shares a clock. Series are ranked by total
rather than by peak — a counter's question is "who used the most", and
ranking by peak would colour the team that made one burst over the team that
ran steadily all week. And the spend chart is replaced by a sentence when no
model has a price, rather than drawn flat along the bottom.
The bucket query groups by ordinal and was planned against the cluster's
PostgreSQL before being believed, with a test alongside the others that run
there: three reports have shipped working on SQLite and failing on it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The page imposes its own ceiling on a turn, so a playground request cannot
occupy a replica for ten minutes. It sent that ceiling regardless of the
model, and a model declaring a smaller max_output then answered every turn
with "max_tokens is 2048, but this model allows at most 1024" — the gateway
refusing its own page, over a number the person never chose and has no way
to change.
The ceiling is the page's to impose on itself and the model's limit is not
negotiable, so it sends the smaller of the two.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
An endpoint advertises every name a client may address it by — the model's
own and all of its aliases — because that is what the router matches a
request against. The picker took those strings as they came, so one model
appeared as several: "coder", "coder-7b" and "qwen2.5-coder-7b-instruct"
were three choices, and choosing between them is a question with no answer.
Each name is now resolved through the catalog and the model listed once,
under its catalog name — the one that also appears in the model list, the
usage table and the logs.
/v1/models still lists the aliases separately, which is deliberate: there a
client is discovering which ids it may send, and every alias is one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The price table is written against the catalog. The name reaching the
accountant was the one the caller typed, so every request made by an alias
missed the table and fell to the default price — the same request charged
two different rates depending on which of a model's names the client
happened to be given.
With no default price that is zero, which is silent. With one it is a wrong
number, which is worse: a budget spent at the rate of whatever the default
says rather than at the model's own.
The router already resolves the alias, two lines above, to decide whether
the key is allowed the model at all. It now passes the resolved name to the
accountant as well, in the ordinary path and in the budget fallback.
The test charges a real request through an aliased name and compares the
figure with arithmetic it states itself, because the first version of it
asserted "more than the fallback" and passed with the bug in place.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Read start to finish as somebody standing up a new cluster, it sent them
down a path we no longer take.
The mirror was absent from all 821 lines, and the page still said that
distributing weights was "a known gap, not an oversight" — so the reader
would have gone away to solve, by hand, the problem the binary in front of
them solves. It is step 4 now, before the nodes, because where a node
fetches from is the difference between minutes and a day, and on a customer
network the nodes often have no route out at all.
Two of the listed gaps had been closed and were still listed: a fetch does
report its progress to the control plane, and there is an HA pattern. A
gap somebody believes in is worse than one they discover.
Added what this week made obvious was missing. Upgrading, including the
zero-byte copy that ssh -n silently produces and the size check that is the
only thing between that and a fleet on a truncated binary. Backing up, in
the order that matters, and the fact that API keys are hashed and do not
come back. Pricing in the first configuration, because a budget without a
price is displayed, counted against nothing, and never binds. And scale,
bench and the source check, none of which were written down anywhere.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The stale-profile flag could not fire where it matters. The control plane
keeps no record of a running replica across its own restart — it rebuilds
them from what the agents report — and adoption left the fingerprint empty,
so every replica read as unknown afterwards. That is precisely the moment
the catalog is most likely to have been edited: change the configuration,
restart the server, and nothing would ever say the fleet disagrees with it.
The fingerprint now travels with StartShard, the agent keeps it beside the
pool and fabric it already echoes back, and adoption takes the replica's
version from the agent rather than from the catalog it is about to be
compared against. An agent too old to report it sends nothing, which still
reads as unknown rather than as matching.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two defects, both of the same kind: something the CLI knew and the rest of
the system did not.
A profile can be edited while a replica is running against the old one. The
reconciler deliberately leaves it alone — churning a replica that is serving
costs more than the imperfection — but there was no way to see it, so the
cluster could disagree with its own configuration indefinitely and read as
healthy. Profile now carries a fingerprint, a replica records the one it was
placed with, and both the API and `ctl replicas` say "stale profile" when
they differ. An empty fingerprint is not stale: it means a replica adopted
from an older agent, which is unknown rather than wrong.
Adding a model from the admin page skipped the source check `ctl model add`
has done all along, so a typo or a mirror that is down became a replica
failing on a node minutes later, with an error naming the node. The check
moves out of the CLI into a package both use.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A team was given a budget of a million, the quota page showed it, and after
sixty requests it read zero percent used. Not because the counting was wrong:
cost comes from a price table, no model had a price, and every request
therefore cost nothing. The number was a limit in appearance and bound on
nothing.
An empty price table is the ordinary case — self-hosted models have no
per-token price, which the schema already says. Showing somebody a budget
without mentioning it is not.
The page says so, and says it precisely: rate and token limits are unaffected,
because they count requests and tokens rather than money. A warning that read
as "no quota works here" would be worse than the silence it replaces — those
two were enforcing perfectly well, as twenty served and ten refused had just
shown.
AnythingPriced is on the accountant rather than the price table, because what a
caller is asking is whether the budgets it is about to display mean anything,
and the accountant is what would enforce them. A model priced at zero is not
priced; a fallback price prices everything.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The mirror would happily hold a four-shard model that nothing could serve. A
catalog entry names one source, the agent fetched one URL, and so a version of
several files had no way to be described — which is every safetensors model,
and every GGUF too large to ship whole.
A source ending in a slash names a version rather than a file. The agent fetches
every file in it into one directory and hands the engine the directory: vLLM
reads the shards and the config out of the folder, and llama.cpp is given the
first part and finds the rest beside it.
The trailing slash is the whole signal. A flag beside the source would be a
second place for the two to disagree, and a URL already distinguishes a
directory from a file everywhere else.
The mirror answers a version with a manifest instead of a file listing. A plain
file server gives HTML there, and something that has to know what a version
consists of would be left parsing it. It is on the byte listener, not the
management API: fetching weights is the one thing every node does, and making
it depend on the administrative port would undo the split that keeps that port
closeable.
Built from the directory rather than a file written beside it. A manifest on
disk is a second record of what is there, and the two disagree the first time
somebody copies a file in by hand — which the mirror explicitly allows.
The whole set or nothing. A directory with three of four shards loads as a
broken model rather than failing to load, so a completion marker is written
last and an interrupted fetch starts the missing files again rather than being
found and served. Files already there at the right size are not fetched twice.
Progress is reported under the version's address, not each file's: the shards
waiting on it asked for the version, and progress under a file's own URL
reaches nobody.
A filename from the manifest is refused if it contains a separator. The mirror
is a place an operator copies files into by hand, so what it reports is not
automatically safe to join onto a path.
An HTML listing where a manifest was expected says so. "invalid character '<'"
sends somebody looking at the wrong thing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Where to put them turned out not to be a question about capacity. Measured on
this cluster: 2,359 series at 41 samples a second, about 7 MB a day, so a year
of history costs roughly 3 GB against 900 GB free. Retention was the package
default of fifteen days, which answers none of the questions this data is for —
what capacity looked like last quarter, whether it was always this slow, how
much the fleet grew.
A year, with a size cap as well. The time limit is the intent; the cap is what
stops Prometheus becoming the thing that fills the disk if somebody adds a
fleet and the series count grows tenfold.
The part that matters is what shares that disk. The model mirror and the
metrics database are on one filesystem, so a mirror that fills it stops
Prometheus writing — monitoring goes quiet exactly when the disk is the
problem, and the alert that would have said so is the one that died.
Nothing was watching it. node_exporter arrives with the Prometheus package and
was already running on that machine, unscraped, while llmgw's own disk metrics
cover only its nodes — and the mirror is not one. It is a target now, and the
alert fires at fifteen percent rather than five, because what is being
protected is the ability to be told.
No remote write and no second tier. At two thousand series that is
infrastructure for a problem this cluster does not have.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The page is served twice — on the admin listener, and at /admin behind a
session — and the history source reached only the second. The first version
built it after the admin listener's copy was already registered and passed that
one an empty NavExtras, so the page an operator reaches by SSH tunnel had no
charts and no explanation.
Built before either is registered, and handed to both.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The dashboard charted what the cluster had been billed for — requests, tokens,
spend, all from llmgw's own store — and nothing about whether it was healthy.
Saturation, queue depth, cache occupancy and replica counts were numbers in a
table at the instant the page loaded, with no way to see that one of them had
been climbing for an hour.
Four line charts now, read back from wherever the metrics were scraped to.
Read back rather than kept twice. llmgw publishes these already; what a gauge
cannot do is remember, and a ring buffer in this process would be a rival copy
of the same numbers with its own retention, lost on every restart and
disagreeing with the one an operator has on screen elsewhere. monitoring.prometheus
is empty by default and the panels are then absent rather than empty.
One axis per chart, always. Two measures of different scale belong in two
charts: a second y-axis lets any two lines be made to cross wherever the author
likes and the reader cannot tell it happened.
Four colours, handed out in fixed order, never cycled. A fifth series is summed
into "other" and drawn in grey — dropping it would make a chart that does not
add up, and inventing a hue for it would put two lines on the page that nobody
can tell apart. The fold is by timestamp, not by position: the tail's series do
not share sample times, and adding them by index slides points sideways onto
each other's clock.
A legend for every chart with more than one line, because colour alone is not
an encoding a reader can rely on.
A flat zero is drawn along the bottom rather than left blank: nothing happening
is a real answer, and an empty frame reads as a broken chart. A single sample
is not a line and says so instead of putting a dot in the corner.
One failed query fails all four panels. Three drawn and one blank invites
reading the blank one as "nothing is happening there".
NaN is dropped rather than plotted. Prometheus sends it as a string for a
sample it has no value for, Go's ParseFloat accepts it happily, and one NaN
coordinate is an SVG path that does not draw at all — taking the rest of the
line with it.
Axis maxima are rounded to a number a person would pick. An axis topping out at
0.8347 makes every gridline unreadable, and the shape is what a chart is for.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Prometheus, Alertmanager and Grafana, with every rule written against a series
this system actually publishes. Nothing here is aimed at a metric that might
exist.
One scrape target for the whole fleet. Agents publish nothing of their own —
the control plane gathers what they report and republishes it — which is worth
writing down before somebody goes looking for an exporter to install on each
node.
It scrapes metrics_listen, not the admin listener. That one is loopback because
it also serves pprof, and pprof can dump the heap where API keys sit in
plaintext. Metrics are the one thing on it something else genuinely needs, so
they get a port of their own rather than becoming a reason to open the other.
The README shows how to confirm the new port carries nothing else; it answers
404 to both pprof and the admin API.
Nothing is sent anywhere. Alertmanager groups and holds and has no receiver:
where an alert should go is a decision with credentials attached, and a
monitoring system that quietly starts mailing people is worse than one that is
visibly not wired up. critical already routes separately so it can be pointed
somewhere else without touching the rest.
The rules encode what the metrics mean rather than round numbers. A saturation
of 1 is the healthy limit measured by ctl bench, not the point of failure, so
the alert waits five minutes above it. A thirty-second p99 first token is the
threshold because this is CPU inference and would be absurd on a GPU. A model
that is deliberately stopped publishes no series at all, so "nothing serving"
cannot fire for one nobody asked for — which is what lets an on-demand
catalogue sit quiet.
Inhibitions, because a node that is down explains every replica on it and
reporting both buries the cause.
The dashboard is provisioned from a file rather than clicked together, so it
survives a reinstall and can be read in a diff.
Checked by firing one: cordoning a node moved llmgw_node_up to 0, the rule went
pending and then firing after its two minutes, the alert arrived at Alertmanager
with the node in its labels, and uncordoning resolved it. An alerting system
nobody has seen fire is an assumption.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Placement compared a profile's memory requirement against the node's total, so
a 31 GB machine accepted two models that each declared 15: each was asked about
on its own, and nothing recorded that the first had claimed anything. GPUs were
never handed out twice this way — they are tracked as claimed indices — and
memory was, because the allocation carried no figure for it.
An allocation records what its profile asked for, and eligibility subtracts
what a node already owes before comparing. The rejection names both numbers,
because "needs 15" beside a 31 GB machine reads as a bug in the scheduler
rather than a full node.
Declared, not measured. A profile that says 15 is treated as holding 15 whether
the process has touched that much yet or not — the alternative is deciding
against what a model has grown into so far, which is wrong every time one is
still loading.
Zero is nobody said, and claims nothing. Counting it would let one undeclared
model make a whole node look full.
Adoption after a restart re-reads the figure from the catalog rather than from
the agent: what a profile asked for is not something a running process reports,
and a control plane that forgot every claim on restart would put a second large
model onto a node already full of the first.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A gateway serving nine models offered nine names and nothing to choose between
them. /v1/models is the only view most clients ever have, and it listed ids.
Models carry a description now: written when the model is added, shown on the
dashboard and sent to clients in /v1/models.
Written rather than derived. A registry offers a pipeline tag and a wall of
README, and neither answers the question somebody picking from a list is
asking — which of these should I send this request to. "General chat, strong
Swedish; slower than the 8B" is not in any metadata field.
It is optional, and absent means absent rather than an empty string in the
response: a model nobody described should read as undescribed, not as described
with nothing.
On the dashboard it sits under its row rather than in a column. A sentence in a
table cell either wraps every other column into a stripe or gets cut off.
Co-Authored-By: Claude Opus 5 (1M context) <moreply@anthropic.com>
Serving a model from the mirror page left "Needs RAM (GB)" optional and empty.
Empty writes no constraint, and a CPU profile has no card to be short of, so
every ready node became a candidate — and a 4.68 GB model went onto a 7.9 GB
machine and left it 3.5 GB for a KV pool. Placement was not wrong; it was never
told what the model needed.
The mirror knows the size of the weights, so the field arrives filled in:
weights plus half again for the KV pool, plus two for the engine. The half is
there because the cache grows with the same things the weights do, layers and
heads. It is an estimate and the field is still editable — what it is not is
blank.
A 1.12 GB model asks for 3 GB, so the small nodes stay usable. Pinning every
model to the one large machine would be a different way of getting placement
wrong.
An unknown size fills in nothing rather than the floor of two, which would
exclude no node while reading as though somebody had measured it.
And when the estimate is larger than any node in the fleet, the page says so,
with the largest node's memory. Better read there than discovered by a node
being OOM-killed twenty minutes into a load.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A node's registry holds an adapter for every engine whether or not its binary
exists, the agent never said which were installed, and placement carried an
Engine field it had never once read. So a vLLM replica was as happily placed on
a node with no vLLM: the node registered healthy, the shard failed minutes
later with "executable file not found" in the engine's own words, and the
reconciler tried again.
That is how fifteen gigabytes of safetensors reached this cluster's mirror
before anybody noticed it had no card and no vLLM to run them with.
Agents now probe what is installed and report it — the protocol has carried
engine_versions since it was written and nothing ever filled it in. Placement
filters on the key set, `ctl nodes` has a column, and the mirror page's serve
warning is a fact instead of a guess about cards, which would have waved
through a fleet of eight-GPU nodes that never had vLLM.
Probing resolves binaries and runs none of them. The first version asked each
engine for --version, which is fine for an engine and not fine for the wrapper
scripts people put in front of one: a script that shifts past flags it does not
recognise and then execs the server will start an inference server, at agent
startup, on every node. The agent's own test suite installs exactly such a
script and caught it. So presence is a path lookup, and the version is left to
`ctl engine-check`, which is run by somebody who meant to run the engine.
SGLang is the exception and has to execute, because a module has no path to
look up — but the command is llmgw's own one-line import rather than the
engine's launcher. Its exit code is the answer: python3 is on almost every
machine, and treating a failed import as "installed" reported SGLang on every
node in a fleet that had none, which is the failure being fixed, inverted.
Probed once, not per reconnect, and bounded by the context that registration
waits on: an agent that cannot reach the control plane is worse than one
reporting an incomplete list.
A node that reports no engines at all is not excluded anywhere — by placement,
or by the page. That is what every agent older than this reports, and emptying
a fleet on a control plane upgrade is a worse failure than the one being fixed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Fetching a model put it on the mirror and left the operator to hand-write a
catalogue entry and run `ctl model add` on the control plane — for a source the
page was already displaying, with an engine decided by the file extension it
had just shown them.
A present row now offers to serve it. The form opens on that row with the name,
the mirror's own URL and the engine filled in, and asks for the two things
nobody can infer: the context pool and how many replicas.
What it refuses is the point.
A version of several files is refused with the reason. A catalogue entry names
one source, so an entry for a sharded model is an entry for one shard, which
loads as a broken model rather than failing to load. Nothing said so before —
the mirror happily holds a four-shard model that no catalogue entry can
describe.
Safetensors on a fleet with no GPU is warned about before anything is added.
This is the failure we hit today: fifteen gigabytes of shards arrived, and only
then did looking at a node show no card, no vLLM and no SGLang. The control
plane is not told which engines a node has installed, so the check is on cards,
which it does know — a warning rather than a refusal, because a card can be
added tomorrow.
A name already declared in configuration is refused here too, matching what the
model service would say anyway, but before the form is filled in rather than
after.
Without a store there is no serve action at all: a model added at runtime would
not survive a restart, and offering it then refusing is worse than not offering.
The mirror page's dependencies became a struct. Three are optional interfaces
and two are the same shape, which at a call site is a swap nobody notices until
a page shows the wrong thing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The explorer listed every weight file with its own button. For a repository of
alternatives that is right — q4_k_m or q8_0, you want one — and for a sharded
model it is fifteen clicks to fetch one thing, with nothing stopping you from
losing count and leaving a version with fourteen of its parts.
Files are grouped into sets now. A run of "-00002-of-00004" is one row and one
button; everything else stays its own row.
Not "fetch all the weights", which would be wrong half the time: two safetensors
files are halves of one thing and useless apart, two GGUF files are usually
alternatives and fetching both spends gigabytes on a copy nobody asked for. The
set is the unit that is whole.
A repository can hold both the split and the joined copy of the same
quantisation — the real Qwen coder repos do — so a shard set keeps the -of-
notation in its name rather than collapsing onto the name of the whole file
beside it.
A set publishes no single checksum. Several files have several, and showing one
of them invites comparing it against the wrong thing.
The form posts every source in the set, and a failure part way says how many
already started: a set that fails on its fourth shard has three arriving, and
reading that as "nothing happened" sends somebody to fetch them all again.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Mirroring anything meant knowing a repository existed, knowing which of its
thirty files were the weights, and assembling a resolve URL by hand. The mirror
page had a text box for that URL and no way to find one.
It now searches Hugging Face and lists a repository's weight files with their
sizes and published checksums, each with a button that fills the fetch form.
Only weight files. A repository is mostly a README and a tokenizer, and listing
those buries the four entries somebody came for.
The size shown is the LFS size, not the plain one. For an LFS-stored file —
every weights file — the plain field is the size of the pointer, so the other
one reports a gigabyte model as 133 bytes, and that is the number somebody
would then plan disk against.
The source URL is pinned to the commit the listing came from rather than to a
branch, so what the mirror fetches is what was on screen. A repository
re-uploaded under the same branch otherwise leaves two different sets of
weights with identical provenance and no way to tell them apart. The version
defaults to that revision for the same reason.
The registry's own failures are reported on their own line. It is somebody
else's server on the other side of the internet, and it being down says nothing
about the mirror — which is what the page is for and has to keep working
regardless. With no registry configured there is no search box at all, rather
than one that can only time out.
Gating decodes from a bool or a string, because the field is documented as the
first and is the second in practice: strict decoding fails the whole response,
so one gated model would hide the other nineteen in a page of results.
And a bug this made obvious: downloads were keyed by model and version, so the
second shard of a sharded model was refused as "already fetching" until the
first had finished. A fifteen-shard model had to be pulled one file at a time,
each waiting on the last. Keyed by file now — which is what the single-flight
was ever protecting, since the race is two writers on one .part. A version made
of several files reports as the worst of them: it is not present until every
shard is, and one shard failing makes the version unusable however many others
arrived.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Resolution took the most restrictive value along the chain and enforced only
that one. For a single caller's ceiling that is right: a team cannot raise what
its organization set. For a pool it is not, because two levels are two pools.
An organization given 1000 with four teams given 500 each spent 2000. Each team
filled its own counter, the organization's value lost to the tighter one and
was never looked at again. The number was written, displayed, reported against
— and enforced by nothing.
Every level that sets a limit is now checked against its own counter, and
settled onto all of them, so an organization's pool fills from its teams'
spending rather than only from requests that somehow named it. A team still
stops before its organization when its own pool runs out first, which is what a
sub-allocation is for.
Effective carries the chain as well as the collapsed values. The collapsed ones
are still what a header or a profile page should show — they are the ceiling a
single caller faces — they are just not enough to enforce with.
Rate limits charge as they check, deliberately, so that two concurrent requests
cannot both see room. With several levels that means a refusal at one has
already taken from the levels before it, so those are handed back. Without the
refund a team sitting at its organization's ceiling would quietly eat its own
rate limit with requests that were never served.
Budgets are checked before any of that, and only read, so a refusal on a budget
charges nothing at all.
Status reports the fullest pool rather than the smallest number. A team inside
an organization with the same budget is bounded by whichever is nearer its
ceiling, and that is the organization as soon as a second team spends anything;
a header showing the team's would disagree with the refusal after it.
A cluster default joins the chain as a key-level entry. Left out, it would be
reported as the ceiling and then never checked.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The tree was rendered and then, underneath it, a form with a dropdown asking
which of those rows you meant. The tree already said. Picking a branch, reading
its current numbers, scrolling past the tree and selecting the same branch from
a list is three steps to do one thing, and two of them are a chance to pick the
wrong row.
A quota belongs to a branch, so the form opens on it: each row links to
?edit=scope:id, that row renders its form inline, pre-filled with what is set
there now and naming the branch in full.
By link rather than by disclosure on every row: a collapsed form per branch
would put a form in the markup for every group on a page whose job is to be
read as a tree. One is open at a time, and the rest keep their link.
Clearing moved into that form as a second submit, so a row that is not being
edited has one control and not two.
Saving closes it. Re-rendering the form over the row it just changed hides the
result and invites submitting the same thing twice.
The note about a quota being shared came out of the footer: the open form says
it, at the point where somebody is about to do it.
The dropdown's parsing went with it rather than being left unreachable.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A quota could be enforced and reported, and set nowhere but a shell. Nothing in
the browser showed one at all — not what a team may spend, not what it has
spent, not that it had no limit. In practice that means quotas nobody sets.
/admin/quotas lists the whole tree, whether or not a group has a limit: the one
without a quota is the one most likely to need one, so it cannot be the row the
page leaves out. Spend over the last thirty days sits beside the budget, with
the fraction used.
Setting a quota here is a change of position. It was kept to the shell because
raising a budget is a financial act and a stolen admin session should not
perform one. That argument holds for exactly one row on this page: resolution
takes the most restrictive value along the chain, so raising a team's or a
project's limit cannot exceed what the organization allows. Only an
organization's own limit is bounded by nothing. Every change is logged with who
made it and what it was before.
Per-key limits stay on the CLI. A key belongs to a person and is already
bounded by its project; letting an admin session reach individual keys would
put every caller's throughput one click away, and the page refuses a key scope
rather than rendering it and failing later.
A limit an ancestor already caps lower is marked as written-but-unreachable.
Resolution means a project given more than its team is not a bigger budget, it
is a number nobody will reach — and without saying so the page would show it as
though it applied, with the first sign being a refusal that disagrees with the
screen.
The group is chosen, not typed. Ids became random in the previous commit so two
organizations could both have a "platform" team, which also made them
impossible to remember and a typo impossible to notice: it would write a quota
onto nothing and report success.
A blank field is no limit of that kind, not zero — zero is a limit of nothing,
which refuses every request. Clearing removes the row rather than writing nulls
across it: both resolve the same, only one tells a reader nobody has set a limit
here.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
An account is approved into a project, a project belongs to a team, a team to
an organization — and until now the only way to make any of them was `llmgw
keys init` on the control plane. So the first thing to do after installing was
leave the interface you were already in, and an approval form with an empty
dropdown gave no hint why.
The accounts page now creates all three.
Creating a container is not granting access: nobody can use one until an
account is approved into it, which is the step already on this page. Quotas stay
on the CLI. Making somewhere to put people is structural; raising what they may
spend is financial, and a stolen admin session should be able to do the first
and not the second.
Ids are random now instead of slugs of the name, which fixes a bug that only
appears when you have a second organization — which is exactly what this
feature invites. Team ids were slugs in one flat namespace and PutTeam upserts
on id, so creating "platform" under a second organization moved the first one's
team into it. Nothing failed. The team changed owner.
Create* is separate from Put* rather than replacing it. The upserts exist so
`keys init` can be re-run while provisioning a machine; a person filling in a
form means something else by a name that already exists, and gets told.
A duplicate is refused among siblings only. Two companies both having a
"platform" team is ordinary, and refusing the second would be the flat namespace
again wearing a better error message.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Quotas can be written on an organization, a team or a project, and nothing
could show what any of them had used: usage came back per key, per user and per
model only. A quota whose consumption cannot be seen is one nobody reviews
until the day it refuses somebody — and since the previous commit made a group
quota a shared pool, it is now a number several people are spending against
without being able to look at it.
`llmgw ctl usage --by team` sums every key beneath each group, with the budget
that level sets beside it.
Beside it, and only when that level sets one. A team bounded by its
organization's budget has none of its own, and printing the organization's
against each of its teams would read as though each had that much — the exact
misreading this work exists to remove.
Joined through api_keys rather than read from usage.project_id: that column
records where the key was when the request was served, while the quota is
written against the tree. They agree until a key is moved, and a report that
disagrees with the limit shown beside it is the useless one.
Three existing reports were fixed on the way, all failing on PostgreSQL and all
passing in tests, because the tests run on SQLite:
column "k.name" must appear in the GROUP BY clause (UsageSince)
column "us.email" must appear in the GROUP BY clause (UsageByUser)
column "usage.ts" must appear in the GROUP BY clause (UsageBuckets)
SQLite picks an arbitrary row for an ungrouped column and carries on; PostgreSQL
refuses. The first two were selecting names they had not grouped. The third
grouped "ts / ?" while selecting "(ts / ?) * ?", which PostgreSQL matches as
text and so saw a bare ts — it is grouped by ordinal now, the same expression by
construction. The dashboard's usage charts have been empty on the cluster this
whole time, with the reason only in the log.
All four queries were run against the cluster's own PostgreSQL before this was
committed, because passing on SQLite is what they were already doing.
The grouping level picks from a fixed table of levels rather than being
interpolated, so an unknown one is an error and not a fragment of SQL.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two things that were only ever confusing, and one of them stopped a command
working on the cluster.
The store type was named SQLite and kept the name when PostgreSQL was added, so
OpenPostgres returned a *SQLite and all seventy-odd methods on the production
store were spelled after the database it was not running on. Nothing broke;
people did. Earlier today I added three methods to *SQLite for a table that on
this cluster only ever lives in PostgreSQL, and the name did not once look
wrong. It is a field, which is all it ever was.
`llmgw users` and `llmgw keys` defaulted --store to a SQLite path. On a
deployment whose store is PostgreSQL every one of them failed with "unable to
open database file" while the right answer sat in auth.store in the config file
beside them; the conclusion an operator draws is that the tool is broken, and
they are not far off. They now ask the server's own configuration before
falling back, and the error names which of the three sources chose the address
— the failure was never the path, it was which one picked it.
The single-machine default still applies last, so a deployment with no config
file keeps working with no flags.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The tenancy tree handed limits down and never pooled what was spent against
them. A limit written on a team was inherited by every key beneath it, and then
enforced with the key as the counter — so a project given a budget of 1000 and
ten keys permitted ten thousand, and a team given 60 RPM got 60 per member.
That is "each of you may spend this much", which is a useful thing to be able
to say. It is not what somebody writing a budget on a team means, and nothing
in the system said which one you were getting.
Resolution now records which scope set each value that won, and the counter is
keyed by that scope. The value and the counter come from the same place, so a
project tightening its team's limit is enforced against the project's own pool
rather than eating into a quota another project never agreed to.
Each limit is owned separately, because they resolve separately: a team budget
with a per-key rate limit — the group's money, the individual's throughput — is
the ordinary arrangement and has to stay expressible.
Cluster defaults are owned by the key. Read as a shared pool, one default would
become a single cluster-wide rate limit that every key in every organization
competes for: not a default, an outage.
Restoring after a restart had to change with it. Per-key totals rebuilt a
counter named after the key while the check read one named after the team, so
the control plane came back believing the team had spent nothing — a budget
that reset on every restart, for exactly the scopes where it matters most.
SpendSinceByScope returns each key's place in the tree and the accountant folds
it into every scope that might own a budget.
Counters are namespaced by level throughout. A bare key id in one path and a
"key:"-prefixed one in another are two counters for the same key, which is how
the restore was silently doing nothing.
Still missing, and worth naming: no usage report rolls up by project, team or
org, so a shared quota can now be enforced but not yet shown.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`ctl bench` has always been able to work out what one replica can serve — it
ramps concurrency until the SLO breaks and reports the last level that held.
It printed that and nothing kept it, so the one number saying what a replica is
worth lived in whoever ran it's scrollback. Everything downstream then had to
guess: admission applies a single global default for healthy queue depth, and
sizing a model against real demand had nothing to size against.
`ctl bench --save` records it; `ctl capacity` reads it back.
Keyed by model and profile. Capacity belongs to the hardware and the engine
arguments as much as to the weights, so the same model on CPU and on an H100
are two numbers, and one key would overwrite the other on every run.
The profile is looked up from the replica rather than asked for, because the
control plane already knows which profile answers on which endpoint and a
measurement filed under the wrong one is worse than none. A target that is not
a replica — the gateway, usually — is refused with that said: benchmarking the
whole path is worth doing and says nothing about the hardware.
The conclusion is stored, not the ramp. Every level and its latencies are worth
reading once while judging whether the run was sound; keeping them would grow
without bound as a model is re-benchmarked.
Re-measuring replaces. A second run is a better answer to the same question,
not another data point, and keeping both would leave a reader choosing between
numbers with no basis for the choice.
The SLO and the date are stored with it. Capacity means nothing without the
target it was measured against, and hardware and engine builds change — a
measurement is an observation with a date on it, not a fact about the model.
A failed save is reported and not fatal: the run has already cost the GPU time.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The count is scaling.min, and until now the only way to change it was to edit
server.yaml on the host and restart — or, for a model added at runtime, to
remove it and add it back, which means having its whole definition to hand.
`llmgw ctl scale <model> --min N` instead. Only the scaling block moves:
everything else is read back from what was stored and written out untouched, so
changing a number cannot quietly become a redefinition.
The catalog is rebuilt and validated before anything is written, as in Add. A
max below a min is refused with the stored definition untouched — persisting a
rejected change would come back on the next restart as a catalog that does not
build.
The answer reports the scaling in force rather than the request. The catalog
fills in a max that was left out, and a caller echoing its own request would
print a number the cluster is not using.
A model declared in configuration is refused, for the reason Remove already
refuses one: the YAML is read on every start, so the new count would hold until
the next restart and then silently revert. A cluster disagreeing with its own
config file is worse than an error saying where to edit.
The help says what --max does, because it is not what it looks like: a model
with a minimum runs exactly that many and nothing scales it up under load, and
an on-demand model loads one replica however high its max.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Control plane mode was inferred from the config file alone: a model with
profiles meant "control plane", nothing meant "router only". So a deployment
that manages its catalog at runtime could not start at all — the file declares
no models, the process refuses for want of anything to serve, and the stored
models it would have loaded are never read. A cluster unmanageable at runtime
because it is managed at runtime.
control.enabled says so outright. AgentListen cannot carry the signal: it has a
default, so every deployment would silently become a control plane.
The inference stays, so nothing already working has to add a flag.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
/admin/mirror, behind the same admin session as the other admin pages. What is
held, what is arriving with its byte count, what failed and why; start a fetch,
retry a failed one, remove a version.
Not mounted at all without a mirror configured. The right page for a deployment
with no mirror is no page, rather than an empty one pointing at nothing.
It reloads itself only while something is downloading. A reload at any other
moment takes with it whatever is half-typed into the fetch form, and there is
nothing to watch when nothing is moving.
A failed row offers Retry, not Remove. A fetch that failed left nothing on disk,
so the remove could only answer "the mirror holds nothing for this" — a button
whose single outcome is a confusing error. Starting the same download again is
what somebody looking at that row wants, and the mirror allows it: only a
download still running is refused.
An unreachable mirror is said outright instead of rendering as an empty table.
The two look nearly identical and mean opposite things: one is a mirror waiting
to be filled, the other is a machine to go and look at.
Dashboard takes NavExtras rather than a second bool. Two adjacent booleans at a
call site get swapped silently and are discovered by somebody clicking the
wrong link.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
All() preferred the disk entry wholesale, and a directory listing has no
checksum in it. So the sha256 a fetch computed was reported while the bytes
were still arriving and vanished the moment the file landed — it disappeared
exactly when it became verifiable, which is the one moment it is for.
It is the field an operator compares against what upstream published. A mirror
is the single place in this system where a corrupted file is copied to every
node instead of being caught once.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The mirror has had a management API since it was built, and the only thing
driving it was curl on the mirror's own machine. This is the other half: a
client, so the operations belong to whatever has a reason to perform them.
The mirror's refusals travel as their own sentence rather than a status code.
Putting the reason in the body was deliberate on the server side, and a client
that kept only the number would undo it — "the mirror said 400" sends somebody
to the wrong machine, "the source URL answered 404" sends them to the right one.
NextProtos is cleared for the same reason the mirror's server clears it. That
configuration is shaped for the agent stream, which pins h2; a transport
carrying a pinned protocol it is not set up to speak negotiates h2 and then
talks HTTP/1.1 over it, and fails with an error mentioning neither ALPN nor
the mirror.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The page described nginx and a curl loop, which was accurate when the mirror
was a directory somebody kept by hand. It is a subcommand now, so it says so:
the two listeners and why they are two, the API and what each endpoint answers,
and the certificates coming from the authority the fleet already has.
nginx stays, further down. It serves a directory perfectly well and anyone
already running it should not be told to install something to replace it.
The systemd unit spells out --base in a comment because it is the flag that
goes wrong quietly: it ends up in catalog entries, and a wrong one makes
`mirror unused` call files unreferenced while they are in use.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The test started a sidecar and an engine, each a shell appending its name to a
file, and demanded the sidecar's line come first. The supervisor spawns the
sidecar first, but spawning is not sequencing: without WaitForLeaderPort
nothing holds the engine back, so two shells race and the scheduler picks the
winner. It failed under full-suite load, and lately on its own.
The wait that is a real guarantee already has its own test, so the fix is to
check what this one is actually for — that both processes run under one shard,
that the shard's pid is the engine's rather than the sidecar's, and that their
output is distinguishable in a single log tail. Renamed to say so.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The mirror holds the weights, so the mirror does the downloading. A control
plane that had to pull the bytes itself and push them here would carry every
model twice over the network, through the one machine least able to spare the
bandwidth.
Two listeners, and the split is the point. Weights stay on plain HTTP where
every node can reach them: that path is eight nodes pulling forty gigabytes,
and putting a client certificate in front of it would mean re-certifying every
agent's fetcher in order to change how models are administered. Management gets
its own port, which can be given the fleet's CA — the same tlsconf the agent
stream uses, so this is a configuration rather than a new idea.
A fetch answers 202, not 200. The bytes are not here yet, and a caller that
read it as "done" would point a catalog at a file still arriving, which every
node would then copy. Progress is asked for rather than waited on, because no
client between here and the operator will hold a connection open for an hour.
A failed fetch is remembered with its reason. Forgetting it would leave the
control plane waiting on a model that is never coming, with nothing to say why.
Status falls back to reading the disk, so a restart does not report a model
that is plainly there as unknown.
Removal takes the whole version: a model split across files is useless with one
of them missing, and half a version on disk looks present and will not load. It
refuses while a download is running, because the fetch would carry on writing
and the file would reappear, looking like the delete had silently failed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The mirror was a directory and a discipline written down in a README. This
makes the three operations that matter into commands, on the mirror host,
where `llmgw keys` already sets the precedent of operator tooling that works
on files rather than through the API.
The gateway still never writes to the mirror. That property was chosen
deliberately — a mirror writable by the thing reading from it is a way to
replace a model with one every node then fetches without complaint — and a
person running a command does not take it away. `serve` states it rather than
inheriting it, so a handler added later cannot quietly make the mirror
writable.
`add` downloads to a temporary name and renames, for the reason the agent's
fetcher does: a mirror is copied to every node, so a partial file here is a
partial file everywhere. It verifies the length the server promised, reports a
checksum — the mirror is the one place a corrupt file reaches every node
rather than being caught once — and refuses to replace an existing version,
which is the whole point of the layout.
`unused` refuses to answer when it cannot read the catalog. An empty catalog
and an unreachable one produce the same list, everything, and the obvious next
step after that list is deletion. It compares by path rather than by spelling,
because a base typed with a different scheme or a trailing slash would
otherwise report a file in use as unused.
`remove` checks the catalog first and says to retire the entry before deleting
the file, because a node that has to refetch a deleted file cannot, and that
failure arrives later, on the node, looking like something else.
Verified end to end: Range requests answer 206, which is what makes an
interrupted fetch on a node resume; writes answer 405; a traversal attempt
answers 404.
Note: internal/agent/supervisor's TestSidecarRunsBeforeEngine fails under full
suite load on main as well as here — it asserts that one shell's echo lands
before another's, which the supervisor does not promise. Everything else
passes. Left alone rather than fixed in a commit about mirrors.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
"Set up a local mirror" was advice. This is the nginx config, the directory
layout and the reason the layout has a date in it — a repository that changes
upstream changes what you are running, and the first sign is a prompt that
worked last week answering differently this week.
It says what the mirror is not, because both mistakes are easy. It is not a
caching proxy: nothing populates it on demand, deliberately, since a mirror
that reaches the internet at the least convenient moment fails in a way that
looks like llmgw being broken. And it is not needed at all if you already run
a shared filesystem — put the models there, use a path, and nothing is fetched.
It also says when a mirror stops being enough, and that llmgw is positioned to
distribute between nodes because it already knows which of them hold which
weights. That is not built, and the page says so rather than implying it is
coming.
Every behaviour it relies on was checked against the source: an http:// source
is fetched and a path is not, a killed download resumes, a partial file is
never visible under the model's name, and a shared filesystem is not counted
once per node.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Neither had a page. "We have GPUs, now what" was left to the reader, and what
llmgw stores was scattered across four documents and a changelog — which is
the same as not being written down when somebody is doing a security review.
docs/data.md describes the behaviour and leaves the legal mapping to whoever
holds the obligation. Every claim in it was checked against the source, and it
names where to look for each: usage rows carry no text field, keys are stored
hashed, conversations are off until a named project turns them on and cannot be
turned on without a retention window.
It also states the egress exactly, which took reading every outbound call to
say honestly: nothing on the serving path leaves the cluster, and the two
things that can reach outside — fetching weights, and ctl plan reading a model's
architecture — are both started by an operator. Weight fetching points wherever
you tell it, including an internal mirror.
docs/choosing-models.md is the opinion llmgw itself does not have. Three jobs
rather than one model, with the third — embeddings — being the one that gets
forgotten and the one that decides whether the deployment can answer questions
about your company or only questions in general.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The choice is unchanged and now better argued: a deployment does not announce
which language its notes are in, so any stemmer picked here is the wrong one
somewhere — and applying the wrong one does not fail loudly, it quietly stops
finding things.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The last piece of HA, and it is a deployment pattern rather than code: llmgw
picks one leader and agents follow it, but client traffic arrives in front of
both and llmgw is not in front of anything.
Writing the pattern found a bug in the pattern. The first draft said to
health-check `/health` and claimed a standby still routes to replicas the
leader placed, so sending it traffic was fine. It does not. A standby's
reconciler places nothing and it holds no agent streams to adopt from, so it
publishes an empty endpoint list — a load balancer following that advice would
have sent half the traffic to something that answers every request with 503.
So there is now a `/ready` that means "this process has somewhere to send a
request": 503 on a standby, 503 on a leader whose replicas have all failed,
200 when there is something to serve. `/health` keeps meaning the process is
up, which is a different and less useful question. Both are open, because a
load balancer has no key.
There is a gap during failover where neither passes — three seconds when the
leader stops cleanly, up to the lease TTL when it does not — and the document
says so rather than implying otherwise. Traffic cannot go to a control plane
that has nothing to route to.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The agent has always logged a percentage as it fetches weights, with a comment
saying a replica pulling a hundred gigabytes and one that is wedged look
identical from the control plane. They still did: the log is on the node, and
the node is not where anyone is looking when a replica has been starting for
ten minutes.
Progress now rides on the shard status and appears under the replica in `ctl
replicas` — a line rather than a column, because most replicas are not
downloading anything and a column would be empty for all of them.
It is deliberately not a state transition. Treating each chunk as one would
re-mark the shard as starting on every report and reset the timer that decides
when a slow start has become a failed one, so an hour-long download would keep
a doomed replica alive indefinitely. It arrives on its own method for that
reason, and a real transition clears it — otherwise a finished download sits in
the listing at ninety-nine percent under a shard that has been serving for an
hour.
Watchers are keyed by source rather than by shard, because a fetch is
single-flight: eight shards starting together produce one download and all
eight are waiting on it.
No percentage is invented when the server sent no length. One that stops moving
is worse than a byte count that keeps rising, so the field is absent and the
listing says "so far".
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
There was no reference. Four narrative documents explain how to install and
operate llmgw, and the only way to learn that `lease_ttl` exists was to read
the one that happened to mention it — or the struct.
Written by hand it would have been wrong within a fortnight. Six settings
arrived in the last two weeks alone: store, servers, lease_ttl, the two
retention windows, min_disk_gb. A reference nobody maintains is worse than
none, because it is the same list with the setting you came to look up quietly
missing from it.
So it is derived from the structs, with the field comments as the descriptions
— which means the only way for it to be wrong is for the code to be wrong, and
it makes the comments worth writing twice over. 1163 lines, 24 blocks, 87 keys.
Two tests keep that true after everyone has forgotten this exists. One
regenerates and compares, so adding a setting without regenerating is a
failure rather than an omission — checked by adding one and watching it fail.
The other walks the tree for koanf tags and fails on a configuration file
that is in neither the source list nor the short list of structs an operator
never writes, because a list is a thing to forget.
The generator finds the repository from its own path rather than the working
directory, so `go generate ./...` produces the same document from anywhere.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Five nodes fit in a table. Fifty do not, and this is built for people who
already have racks of GPUs — so the flags are how a question gets asked:
which nodes on this fabric, which of this GPU class, which are not ready,
which replicas have a shard on this machine.
Filtered here rather than in the API, so --json is narrowed the same way the
table is. A server-side filter would be two implementations of one predicate,
and the JSON one is the one nobody would notice had drifted.
An empty result says which case it is. A filter that matched nothing and an
empty fleet call for opposite next steps — one is a cluster to look at, the
other is a word to correct — and printing "no nodes" for both makes a mistyped
fabric name read as an outage.
A state matches on prefix because states are few and distinct and somebody
scanning a fleet types quickly. A pool matches either half, because it is
fabric/class and both are worth narrowing by. A replica is found from any of
its nodes, since one spanning two machines is on both.
--label takes a bare key as "carries this at all", which is how you ask for
every node with a role without caring which. The flag's help said so before
the parser did; the test caught them disagreeing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The store stops being a file on one machine, which is what stood between this
and more than one control plane.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Three seconds to move leadership, six for every agent to arrive, and every
replica adopted rather than restarted. Also what it did not do: the public
address returned 502 while the standby served correctly on its own port,
because nothing in front of the two moved.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The second half of leadership. A standby that refuses to place but still holds
an agent's stream is worse than useless: commands travel down the stream the
agent opened, so the leader's placements fail with "no agent connected" for a
node that is plainly connected.
So a standby refuses the stream, before reading the opening message, and says
why. And an agent takes a list rather than an address, trying each in turn — a
standby refusing, a stopped control plane refusing the connection and a crashed
one never answering are the same thing from a node, and the answer to all three
is somewhere else.
`server` and `servers` are both accepted, with `server` first when both are
given: every agent configuration written before there could be a second control
plane names one address, and an operator adding a standby expects the one
already working to stay preferred. With a single address the behaviour is
exactly what it was, including the backoff — which is skipped only while there
is another address that has not just been tried, because the delay exists to
stop hammering one control plane and the next one has not been hammered.
Removed a check that asked whether one field was empty, which is how the first
run of these tests failed: the list was populated and `server` was not, and an
agent configured entirely correctly was refused by its own validation.
The tests run a real agent against a real service over h2c. The first attempt
used a plain HTTP server, and the refusal arrived as an HTTP/2 framing error
that said nothing about why — which is exactly what an operator would have seen.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
These TRUNCATE every table before each test. Pointed at a deployment that is a
silent and complete data loss — every key, every user, every conversation.
Not hypothetical. I did it: an SSH tunnel to the production database had been
opened minutes earlier for a perfectly good reason, and the same DSN was then
handed to the test suite. Twenty-one users, nine API keys, eighteen sessions,
four hundred and sixty-one usage records and three runtime-added models, gone
between one command and the next. Recovered only because the pre-migration
SQLite file was still there.
The guard is the database name. A test database has to say so, because a check
that could be satisfied by accident would not have stopped this one — the DSN
was entirely correct for what it had been opened for, and catastrophic for
what it was then used for.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The first thing HA needs that a shared store does not give. Two control planes
against one store both reconcile: both see a model wanting two replicas and
none running, and both place. The model ends up with twice what it asked for,
each replica holding GPUs the other's inventory believes are free — and nothing
crashes to say so.
A lease in a table rather than a lock held by a connection. PostgreSQL has
advisory locks and SQLite has nothing like them, and this store keeps one set
of queries for both; a row with an expiry works everywhere, and it survives a
leader being powered off rather than asked to stop, which is the case that
matters. Taking and renewing are the same statement, so the two cannot disagree
about what a held lease looks like, and it is one statement, so two processes
racing cannot both believe they won.
The owner id is random rather than the hostname. A hostname would let a
restarted process match its own stale lease and believe it had been leading all
along — including through the gap that made it restart, during which somebody
else may have taken over and started placing.
A store that cannot be reached is not evidence that somebody else is leading,
so an unreachable store keeps the claim. Two leaders is worse than one, but no
leader means nothing is placed and nothing is evicted, and a brief database
blip should not stop a cluster healing itself.
Unset means leader. Every deployment that existed before this, every
single-control-plane install and every test acts unconditionally, with nothing
to configure.
The same seven assertions run against both backends, which is how the test
harness was caught inheriting a lease between subtests — it passed on SQLite,
where each subtest gets a fresh file, and failed only on PostgreSQL.
What this does not yet do is move agents. A standby holding an agent's stream
is one the leader cannot command, so failover is degraded rather than clean
until agents know more than one address.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The section leads with why SQLite is still the default rather than with how to
leave it: one file, no server, and a request path that barely touches it. Two
things change that, and both are recent — conversations and notes read and
write on every turn, and a second control plane is impossible while the store
is a file on one machine.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Keys are stored hashed, so a migration that drops them cannot recreate them:
every person and every script needs a new one issued, and the old ones cannot
even be looked up to say which they were. `llmgw keys move-store` copies
instead.
Found by running it against the real cluster, twice, after the gateway had
already been stopped.
The first order came from the order the migrations create tables in, and
users is created in a later migration than api_keys even though api_keys
references it — so the copy died three tables in. The order is now checked by
a test that derives the foreign keys from the migrations rather than restating
them, so a new reference fails there instead of during somebody's migration.
Checked by putting the wrong order back and watching it fail.
Then key_claims. Its sealed column is emptied the moment a claim is used, so
every claimed row holds an empty blob in a NOT NULL column. An empty blob
scans as a typed nil slice and a genuine NULL as an untyped one — a
distinction this side can see and pgx cannot, since it encodes any nil slice
as NULL. Restored before the insert, which is the difference between a copy
that completes and one that stops four tables in.
One transaction per table rather than one for the whole copy: a single
transaction holding every row of a large usage table is how a migration runs
out of memory on the machine least able to spare it. A failure therefore
leaves a partial copy, which is why the destination must have been empty —
starting again is dropping and recreating it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A path is still SQLite and a postgres:// URL is PostgreSQL, so every existing
configuration keeps working untouched and moving is one line changed rather
than a new setting to find. An address of the form @/path/to/file reads the
real one out of that file, which is how a DSN carrying a password stays out of
a configuration that gets copied, diffed and pasted into issues.
The driver is pgx, which is pure Go. That is not a preference: the whole binary
is built with CGO_ENABLED=0 so it runs on a node whose libc is older than the
builder's, and a cgo driver would take that away from every deployment,
including the ones still on SQLite.
Tested against a real PostgreSQL, which found two things nothing else could.
The schema_migrations table is created in Go rather than in a migration file,
so it escaped the port and still said INTEGER — 64 bits in SQLite, 32 in
PostgreSQL, and a millisecond timestamp does not fit, so the very first
migration failed to record that it had run. And the memory migration uses an
FTS5 virtual table with three triggers keeping it in step, which has no
counterpart: PostgreSQL derives a tsvector from the row, so there is nothing
to keep in step and no trigger to forget.
That leaves exactly one query that is not the same sentence in both dialects —
the note search, MATCH against @@ — and it is written out twice rather than
hidden behind a shared spelling that would mean something subtly different in
each. Everything else is one set of queries.
The text search uses the 'simple' configuration rather than English. These are
an agent's own notes, in whatever language it chose, and stemming English
rules over Turkish would quietly stop finding things.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Groundwork for PostgreSQL, with no PostgreSQL in it yet: the store still opens
SQLite and behaves exactly as it did.
One set of queries rather than one per database. The SQL here is almost
entirely portable — ON CONFLICT upserts work in both, and the only construct
that genuinely differs is the placeholder — so two copies of every query would
buy nothing and drift, one query at a time, discovered in production.
The rewrite is a wrapper around the handle rather than a change at each call
site. There are sixty of those and sixty chances to forget one; this way the
only route to the driver goes through the rewrite. Making it a wrapper found
six places that reached past it, including all three manual transactions and
two prepared statements — every one of which would have worked on SQLite and
failed on PostgreSQL, on a path nobody exercises before deploying.
The rewriter skips quoted sections. No query here contains a question mark in
a literal today; this is so that adding one is a test failure rather than a
statement that silently binds the wrong argument.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The gap list said a fetch "reports no progress". It does — the agent logs a
percentage as it downloads, with a comment saying a replica pulling a hundred
gigabytes and one that is wedged look identical from the control plane.
That comment describes the real gap, which is narrower and more useful than
the one written down: the progress exists, it just never leaves the node, so
the operator looking at ctl or the dashboard still cannot tell those two
apart.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Seventeen commits since v0.6.0, and not a phase of the roadmap — which is why
the table leaves the phase column empty rather than inventing one. What landed
is what a cluster needs to be lived with rather than demonstrated: somewhere
to keep a conversation, somewhere to keep a note, and an honest answer about
the disk underneath both.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Found by answering "where can we see this live". `ctl weights` was offering a
file for deletion that I had deleted two days earlier, and the node's disk
figure was the one it had when its agent started. Discover() is called at
registration and nowhere else, so both were frozen.
Disk is a fact about a moment, not about a machine. Sampled once, the figure
is wrong on exactly the node worth watching — the one busy filling up — and it
stays wrong until that agent restarts. The heartbeat already refreshes GPU
memory for the same reason, with a comment about noticing a foreign process
"after registration"; disk belonged beside it and I did not put it there.
The stale list cut both ways. A model fetched after registration was invisible
until a restart, so placement could not prefer the node that had just spent ten
minutes downloading it, and the disk check could not see the space it used.
Storage goes in every heartbeat, being a few numbers per filesystem. The file
list goes only when a fingerprint of paths and sizes changes, with a flag
saying the list is meant — without that, an empty list would be ambiguous
between "this node now holds nothing" and "nothing to report", and the second
reading would empty a node's inventory five seconds after it registered.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two of the four items were already built. A model falls through to its
fallbacks on no capacity as well as on a spent budget, and the per-request
routing profile header exists — it is ignored under authentication on purpose,
because a header that picks a routing profile is a header that promotes a
client into the top priority band.
Checked against the code rather than against memory, which is how both got
onto the list in the first place.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The section opens on the default rather than on the feature, because the
default is the part someone deploying this needs to be sure of: nothing anyone
types is stored until a project asks, and the command to ask is an act rather
than a setting.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
SetThreadRetention existed and nothing called it, so the feature was
unreachable: a project could never opt in and every request carrying a thread
was refused.
A command rather than a configuration key, because it is a decision about data
rather than about deployment. Someone has to choose to keep what a team types,
and choosing it should look like an act rather than like a default nobody
read.
--retain-hours is required. Without a number the command would create exactly
the thing the design avoids: a conversation store with no expiry, holding
everything a team ever typed because nobody chose one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two halves of keeping the promise that storing makes.
DELETE /v1/threads/{id} is a client endpoint, not an admin one: the person who
said the thing should be able to unsay it, and making them ask an operator
would mean the operator reads it first. Deleting a thread in another project
answers exactly as deleting one that never existed does — two answers would
let a caller map which ids are real in projects they cannot read, and this one
would do it by trying to destroy them.
And something now runs the pruner. Without it "we keep conversations for 24
hours" was a comment rather than a promise: the rows would accumulate and the
number a project chose would describe nothing. Hourly, and once at startup,
because a gateway that has been down for a week should not wait another hour
before honouring a window that expired while it was off.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The request path half. A caller sets x-llmgw-thread and sends only the new
turn; llmgw puts the stored history in front of it, sends that to whichever
model is serving, and stores the exchange afterwards.
A header rather than a body field because some SDKs drop fields they do not
know and all of them let a caller set a header.
Capturing the reply is opt-in and bounded. This is the hottest path in the
system, so a request without a thread accumulates nothing at all, and one with
a thread stops accumulating at a cap rather than growing to it — a gateway
that kept every reply in memory would be one long answer away from trouble. A
reply is truncated on a rune boundary rather than dropped: a shortened turn
replays as a shortened turn, while nothing at all replays as the assistant
having said nothing, and invalid UTF-8 would break the next turn's body for
good rather than for one turn.
The streamed path sets the captured reply on every exit, not only the one that
looks like the end. A stream can also finish with the client gone or the
engine stalled, and what the assistant said up to that point is still what it
said.
A thread on a project that keeps no conversations is refused. The e2e test
asserted that and got a cheerful 200: the request was served, nothing was
stored, and the caller would have discovered it only when the model had no
idea what they were referring to — which reads as the model being broken. It
now asks whether storage is on before doing anything else.
History is spliced after the model is finally settled, because a budget
fallback may have changed it and trimming is measured against the context of
the model that will actually serve.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The part that makes a thread work across models: the history goes in front of
what the caller just said, because a client using a thread sends the new turn
rather than the conversation again.
Trimming has three rules, and each is there because the alternative fails
quietly. The oldest stored turns go first, so a conversation that outgrew the
model now serving it stays usable — after a budget fallback that model was not
the caller's choice, and a permanently unusable thread would be a strange way
to punish them for running out of budget. The caller's own messages are never
dropped: answering something they did not ask is worse than refusing. And a
system message is kept wherever it sits, because losing one silently changes
what the model is for while it carries on answering.
A model that declares no context length has nothing trimmed. Guessing a limit
for it would drop turns for no reason and the caller would never learn why the
model forgot.
The body is rebuilt through a map so every field the client sent survives.
Forwarding fields llmgw does not understand is how engine-specific options
reach an engine at all, and a thread must not be what starts dropping them.
An empty history returns the body byte for byte, so a request without a thread
takes exactly the path it took before this existed.
Not yet wired to the request path.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The storage half of continuing a thread on a different model than it started
on. Nothing uses it yet.
Until now llmgw stored no message content anywhere: usage rows carry token
counts, costs and latencies, never text, and even its own chat page keeps the
conversation in a JavaScript variable that a refresh discards. That property
is worth more than the feature, so it is the default and stays the default.
A project stores nothing until someone turns it on for that project.
Retention is a number of hours rather than a boolean, because a thread store
with no expiry is a transcript of everything a team ever typed, kept forever
because nobody chose a number. Turning it on means saying for how long. The
window measures idleness rather than age — expiring a conversation someone is
still having would be a strange way to keep that promise.
Threads are scoped to a project and the scope is part of every lookup, not a
check afterwards. Reading a thread from another project and reading one that
does not exist return the same error: two errors would let a caller learn
which ids are real in projects they cannot read. Writing across projects is
refused for the same reason in the other direction — append, then read your
own reply back.
Deleting a thread takes its messages with it, which the test checks by
counting rows rather than by trusting the cascade. Someone has to be able to
say "forget that", and a retention window is no answer to a message sent by
mistake.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The context check ran against the model the caller asked for. The budget
fallback then swapped in a different one and nothing looked again, so a prompt
sized for a 32k model could be sent to a 512-token one. Measured before the
fix: 10002 prompt tokens delivered to a model declaring 512.
The caller sees that as a truncated reply or an engine error naming a model
they never chose, which is a worse outcome than the refusal the fallback
exists to avoid — it fails later and explains less.
Candidates that cannot hold the prompt are now skipped, and a fallback further
down the list that can hold it is used instead. When none fit, the request is
refused as it was before the fallback existed.
The history is deliberately not trimmed to make it fit. Dropping a caller's
messages to stay under a limit they never asked about would be a silent loss,
inside a request that was already being quietly moved to another model.
The tests drive the real request path, because a unit test on the fallback
would have asserted the same wrong thing the code did. Both were checked
against the unfixed code and fail there.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A command that deletes data needs its limits written down more than its
syntax. The section says plainly that llmgw only takes back files it fetched
itself, and then says the uncomfortable half out loud: that this means it will
not reclaim the storage most likely to be filling your disk, because weights
you placed yourself are yours to remove.
Also records the dash in the DISK column as "not asked" rather than "no
space", which is the reading that matters when a node will not schedule.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`ctl weights --collect` answered "this control plane cannot ask nodes to
delete files" against a real cluster, with the whole suite green.
The controller is handed a dispatcherRef rather than the agent service —
an indirection that exists to break a construction cycle, and one that
satisfies the dispatcher interface by forwarding each method by hand. The
controller reaches past that interface by type assertion to find capabilities
beyond scheduling, and an assertion against dispatcherRef fails for anything
nobody remembered to forward.
Nothing could have caught it below the wiring. The dispatchers the tests use
do not implement the capability either, so its absence there is the expected
case and looks identical to this bug. The test therefore asserts on the
production types themselves: that the forwarding dispatcher carries the
capability, and that the service it forwards to has one to carry. Checked by
deleting the method and watching it fail.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two halves of the same problem: nothing stopped a model being placed on a node
that could not hold it, and nothing ever removed the weights of a model that
had been deleted from the catalog.
Placement now has a disk check, shaped exactly like the memory one it sits
beside, with one addition: it is skipped on a node that already holds the
weights. A node holding the file needs no room for it, and without that the
node best placed to run a model — because it already has it — is the one most
likely to be rejected, since holding the file is what used the space.
Reclaiming is the dangerous half, so the whole design is about what it will
not do. llmgw deletes only files it fetched itself, and that is a calculation
rather than a heuristic: the cache name is a hash of the source, so the set of
wanted files can be computed from the catalog exactly. Anything without that
name was put there by an operator — scp, a shared mount, a copy by hand — and
is none of this system's business however unreferenced it looks. A model that
took a week to copy must not be removable by a command someone ran to tidy up.
The node checks every path again before unlinking it, and that duplication is
the point. The control plane knows the catalog; the node knows its disk. A
control plane that is wrong, newer than the agent, or tampered with must not be
able to talk a node into deleting something, so the agent refuses anything
outside its own cache directory or without the cache name — including a path
that tries to traverse out of it.
Listing is the default and deleting takes --collect. The summary says "asked to
remove", because each node decides, and one that refuses a path it cannot prove
it fetched is this working rather than failing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Disk was the one resource nothing reported. `statfs` was already being called
on every model directory, but only to ask whether the filesystem was a network
one — the free space it returns in the same struct was thrown away. So the
only way to find out why a node was filling up was to ssh to it, which is what
this system exists to avoid.
Reported per filesystem rather than per configured directory. Free space
belongs to the disk, not the path: two model directories on one disk are one
pool of bytes, and counting them separately would let the control plane add
24 GB to 24 GB and place a 40 GB model on a node that has 24.
Shared filesystems are summed separately from local ones and left out of a
node's own total. Space on a filer belongs to everyone mounting it, so twenty
nodes would otherwise each report the same free terabyte, and a scheduler
would place twenty models into room for one.
Nothing reported means nothing is known, everywhere: an agent too old to send
storage and a node with no model directories both send nothing, and both must
read as unknown rather than as a full disk. `ctl nodes` and the dashboard show
a dash, the metric is absent rather than zero, and a missing directory — which
is what an unmounted filer looks like — is skipped rather than reported as
having no space.
Shown as free out of total, because the first is useless without the second:
8 GB free is comfortable on a 10 TB filer and nearly full on a 10 GB root
partition, and only the second stops a model being placed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
It had one verdict for a failed attempt — "This is the engine, not the model.
Choosing a bigger model will not help" — and that verdict was wrong, in a way
that sends someone off rebuilding llama.cpp when the fix is a one-line catalog
change.
Measured on the cluster. The engine renders the tool prompt correctly: the
<tools> block and the instruction to wrap the call in <tool_call> are both in
the prompt, which I confirmed through the engine's own /apply-template. The
coder models simply do not use that wrapper — qwen2.5-coder-3b puts the call
in a markdown fence, qwen2.5-coder-7b invents <function-call> — so the parser,
which looks for <tool_call>, has nothing to find and leaves it in the content.
qwen2.5-3b-instruct, same engine, same build, same flags, returns a populated
tool_calls array. So it was the model all along, and the advice against trying
another one was the most expensive sentence in the output.
A failed attempt is now two verdicts, told apart by whether the reply used the
wrapper its template asked for. Wrapped and unparsed is genuinely the engine's
parser. Unwrapped is the model, and the fix is one trained for the format
rather than a bigger one from the same family.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
It said "there is no distribution mechanism yet" and then, two paragraphs
later, documented the distribution mechanism. The known-gaps list — the part
someone reads at 2am — still said "no cache, no fetch, no progress". Fetching
landed in this cycle; both sentences were written before it and left standing.
node-onboarding.md carried the same claim.
What is actually still missing is narrower and worth naming precisely: a fetch
reports no progress, so a first start on a large model looks like a hang, and
every node fetches its own copy, so eight nodes on a new model are eight
downloads from one origin.
Also: the dashboard now shows who is using the cluster and what they are
running, which the sentence describing it did not mention; and one sentence
about idle timeouts had lost the words that made it a sentence.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-18 06:04:00 +02:00
1190 changed files with 227343 additions and 3129 deletions
| v0.9.0 | — | Managing people and the tenancy tree | released |
| v0.10.0 | — | InfiniBand fabrics and NVLink, doctor, node add, engines installed by the nodes | released |
| v0.11.0 | — | High availability from setup, the setup and console screens, usage sums and per-tenant alerts | released |
| v0.12.0 | — | The cluster's settings in the store, in force when changed; the firewall switch; wipe | released |
| v0.13.0 | — | The whole system stopped and started from one command; Qwen3Guard judges, on a GPU and CPUs together | released |
| v0.14.0 | — | Knowledge bases: documents from people's own machines, kept apart per tenant, answered from | released (v0.14.1: adapters, the tree over the API, the firewall on every machine; v0.14.2: pictures from ComfyUI; v0.14.3: 3D models as jobs; v0.14.4: any ComfyUI workflow as a job, presets, licences; v0.14.5: picture and 3D models that add themselves) |
| v1.0.0 | 5 | What phase 5 has left: precise KV routing proven on vLLM, the block cost picker, defragmentation across NVLink domains, prefill/decode disaggregation | planned |
As of v0.6.0 a cluster runs unattended: agents report their hardware, the
control plane places replicas — including ones larger than a single machine —
models load when someone asks for them and stop when nobody does, and keys,
budgets and rate limits decide who may use any of it.
Before v1.0.0: routing profiles selectable per request, spillover to a fallback
pool, tracing, and reference documentation. One thing is worth saying plainly:
Since v0.11.0 (each release's detail is in [CHANGELOG.md](CHANGELOG.md)):
- **Picture and 3D models that add themselves** (v0.14.5). Where a node's
card fits flux-schnell or triposr and the catalog makes nothing of that
kind, the leader adds it on demand, so a new cluster's portal has its
Images and 3D pages; the cluster page and the console show each preset
and add it by hand, and `presets.auto` turns it off. Stable Fast 3D's
colours can be brought to the picture's (`albedo_match`), and every mesh
carries its whole licence.
- **Any ComfyUI workflow, and models by name** (v0.14.4). `POST
/v1/workflows/runs` runs a ComfyUI model's workflow as a job, its inputs
(text, numbers, choices, images, masks, audio, video, files) declared in
the catalog and listed in `/v1/models`. `llmgw ctl model add triposr` or
`flux-schnell` adds a tested model with no YAML; a profile may name the
card it needs instead of a class, and ComfyUI now installs on V100s. A
licence carries what a machine can match, and a person can leave
licences out on the profile page so no key of theirs reaches such a
model. A 3D job waiting for the one card says what holds it and for
about how long.
- **3D models as jobs** (v0.14.3). `POST /v1/3d/generations` answers with a
job, asked after until its mesh is ready to download. ComfyUI makes it on
the cluster's own cards, TripoSR being the tested model. The job says how
far along it is, the model's licence is written into the GLB, and
`per_mesh` prices it. People make them at `/3d` on the portal; `llmgw ctl
3d` shows every owner's jobs. A model profile may also load weights of
its own, and an admission band may wait longer than the rest, for bulk
work.
- **Pictures from ComfyUI** (v0.14.2). `POST /v1/images/generations` is made
by ComfyUI, behind a gate that demands the replica's key, as an engine the
agents install from its source checked against a digest. `per_image`
prices it, and people make pictures at `/images` on the portal.
- **Adapters, and the tenancy tree over the API** (v0.14.1). `llmgw ctl
kb dataset` and `llmgw ctl train` make a LoRA adapter from a collection
on a node's spare GPU memory, kept on the mirror and seen only by its
owner. Organizations, teams, projects, keys and memberships are in the
admin API and the Terraform provider. The agents' settings are edited
with the cluster's, on the portal too, and the firewall setting reaches
every machine.
- **Knowledge bases** (v0.14.0). People put a company's documents in from
their own machines with `llmgw-kb`, a separate download for Windows,
macOS and Linux that reads PDF, Office, OpenDocument, mail and many more
formats and sends text and vectors, never the files. Collections are kept
apart per tenant, down to row-level security in PostgreSQL. A chat
request, or a part of the tree by default, is answered from them with
its sources: [docs/knowledge.md](docs/knowledge.md).
- **One command for the whole system** (v0.13.0): `llmgw stop`, `start` and
`restart`. Qwen3Guard is the guardrails' judge, on a GPU and on CPUs
together.
- **Settings that apply when changed** (v0.12.0). The cluster's
configuration lives in the store, the same on every control plane and in
force within seconds (`llmgw ctl config`, the console's Settings). The
firewall can be switched on and off for the whole cluster.
- **Security hardening** (v0.11.2): signed releases only, a token on the
admin API, engines not run as root, TLS for the agent port, etcd and
Patroni, a firewall for llmgw's internal ports, encrypted backups,
bounded connections, sign-in throttling, a Content-Security-Policy.
- **Guardrails** (v0.11.2). Policies check requests and answers for PII,
secrets, prompt injection, deny patterns, and with a judge model; they
observe or enforce, and attach to any part of the tenancy tree:
[docs/guardrails.md](docs/guardrails.md).
- **Administration** (v0.11.2):
- an audit log of every change (`ctl audit`);
- staff roles (viewer, operator, billing, admin);
- single sign-on through OpenID Connect and provisioning through SCIM
2.0;
- key lifetime limits, and rotation with overlap (`keys ttl`, `keys
rotate`);
- usage export as CSV (`ctl usage export`).
- **Notifications** (v0.11.2): webhooks, signed or Slack-shaped, for a
model parked, a node lost, a budget spent and new weights published, and
Alertmanager receiver templates.
- **Traffic and models** (v0.11.2):
- request tags, canaries of new weights with automatic rollback, and
quality suites gating updates;
- a response cache, kept and replayed requests, and batches;
- Anthropic's Messages API, and more of OpenAI's (structured output,
images, rerank, audio, moderations);
- scaling by the clock, and SLO-aware admission;
- TensorRT-LLM, MIG, and models imported from Ollama;
- routing by what each vLLM replica's cache holds, and LoRA adapters
loaded on demand;
- other llmgw clusters as providers, and placement that knows racks,
PDUs and power caps.
- **Operations** (v0.11.2):
- `server restore`, off-machine backup copies, and secrets in Vault;
- forgetting a person, and keeping a project's requests in the cluster;
- offline installs from a bundle, and a Helm chart;
- tracing from the control plane to the engines;
- the admin API in OpenAPI, with a Terraform provider;
- the portal in English and Turkish, with its docs at /docs.
Every one of these is in the console as well as on the command line:
[docs/console.md](docs/console.md).
What is still ahead is in [BACKLOG.md](BACKLOG.md). Two things that look like
gaps are not: a model falls through to its fallbacks when there is no
capacity as well as when a budget is spent, and routing profiles are chosen
by the key rather than by a request header on purpose — a header would let
any client promote itself into the top priority band.
One thing is worth saying plainly:
multi-node replicas are proven on a simulated cluster and on real hardware
without a fabric. The InfiniBand, NVML and NCCL paths are written against the
formats and failure modes those tools actually have, but they have not yet run
on a machine that has them.
without a fabric. The InfiniBand, NVLink and NCCL paths — telling fabrics apart
by their subnet managers, mapping them, finding NVLink domains, launching across
nodes — are written against the formats and failure modes those tools actually
have and tested against them, but they have not yet run on a machine that has
them. On such a machine, run `llmgw doctor --json` first and keep the report;
[docs/fabrics.md](docs/fabrics.md) says what to look at after that.
## Documentation
Every page is also in the binary: `llmgw docs`, or **Manual** in the console.
**Installing**
- [docs/quick-install.md](docs/quick-install.md) — from empty servers to a
first request, in six steps: start here
- [docs/install.md](docs/install.md) — the same with every option explained,
and running the cluster after: monitoring, notifications, single sign-on,
the firewall, upgrades, backups
- [docs/node-onboarding.md](docs/node-onboarding.md) — adding nodes to a
cluster that exists
- [docs/kubernetes.md](docs/kubernetes.md) — the control plane on
Kubernetes with the Helm chart, and why GPU nodes stay outside it
- [docs/fabrics.md](docs/fabrics.md) — InfiniBand fabrics and NVLink: what
llmgw needs of them, checking a node with `llmgw doctor`, and what to look at
when a replica spanning nodes will not run
**Running it**
- [docs/console.md](docs/console.md) — the console that bare `llmgw` opens:
every view, and the command each of its actions runs
{{- fail "image.repository: the image built from deploy/container/Dockerfile, pushed where the cluster can pull it (the project publishes binaries, not an image)" -}}
"description":"A clock held low under load is the card throttled; see the next panel for why.",
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
},
"gridPos":{
"x":12,
"y":45,
"w":12,
"h":7
},
"id":20,
"targets":[
{
"expr":"llmgw_gpu_sm_clock_mhz",
"legendFormat":"{{node}} gpu {{gpu}}",
"refId":"A"
}
],
"fieldConfig":{
"defaults":{
"unit":"rotmhz"
},
"overrides":[]
},
"options":{
"legend":{
"displayMode":"list",
"placement":"bottom"
},
"tooltip":{
"mode":"multi"
}
}
},
{
"type":"timeseries",
"title":"What holds GPUs back",
"description":"Cards throttled now, by reason, and uncorrected ECC errors since the driver loaded.",
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
},
"gridPos":{
"x":0,
"y":52,
"w":24,
"h":7
},
"id":21,
"targets":[
{
"expr":"sum by (reason) (llmgw_gpu_throttled)",
"legendFormat":"throttled: {{reason}}",
"refId":"A"
},
{
"expr":"sum by (node, gpu) (llmgw_gpu_ecc_errors)",
"legendFormat":"ECC errors {{node}} gpu {{gpu}}",
"refId":"B"
}
],
"fieldConfig":{
"defaults":{
"unit":"short"
},
"overrides":[]
},
"options":{
"legend":{
"displayMode":"list",
"placement":"bottom"
},
"tooltip":{
"mode":"multi"
}
}
},
{
"type":"timeseries",
"title":"Output tokens/s per GPU, by model",
"description":"What each model gets out of each card it holds.",
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
},
"gridPos":{
"x":0,
"y":59,
"w":12,
"h":7
},
"id":22,
"targets":[
{
"expr":"sum by (model) (rate(llmgw_tokens_total{direction=\"completion\"}[5m])) / count by (model) (llmgw_gpu_assigned)",
"legendFormat":"{{model}}",
"refId":"A"
}
],
"fieldConfig":{
"defaults":{
"unit":"short"
},
"overrides":[]
},
"options":{
"legend":{
"displayMode":"list",
"placement":"bottom"
},
"tooltip":{
"mode":"multi"
}
}
},
{
"type":"timeseries",
"title":"Output tokens per watt-hour, by model",
"description":"Tokens for the energy the model's cards draw.",
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
},
"gridPos":{
"x":12,
"y":59,
"w":12,
"h":7
},
"id":23,
"targets":[
{
"expr":"sum by (model) (rate(llmgw_tokens_total{direction=\"completion\"}[5m])) * 3600 / sum by (model) (llmgw_gpu_power_watts * on (node, gpu) group_left (model) llmgw_gpu_assigned)",
"legendFormat":"{{model}}",
"refId":"A"
}
],
"fieldConfig":{
"defaults":{
"unit":"short"
},
"overrides":[]
},
"options":{
"legend":{
"displayMode":"list",
"placement":"bottom"
},
"tooltip":{
"mode":"multi"
}
}
},
{
"type":"timeseries",
"title":"GPU wear",
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
},
"id":24,
"description":"What wears cards out: XID errors in the last 15 minutes, memory rows remapped, PCIe retries an hour. Anything above zero here is worth a look on that card's node page.",
"gridPos":{
"x":0,
"y":66,
"w":24,
"h":7
},
"targets":[
{
"expr":"increase(llmgw_gpu_xid_errors[15m]) > 0",
"legendFormat":"XID {{node}} gpu {{gpu}}",
"refId":"A"
},
{
"expr":"sum by (node, gpu) (llmgw_gpu_remapped_rows) > 0",
"description":"Requests, tokens, failures, refusals and first-token time per tenant \u2014 the level router.tenant_label names \u2014 and every budget's share spent.",
"description":"What people wait for before anything appears.",
"gridPos":{
"x":18,
"y":0,
"w":6,
"h":4
},
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
},
"targets":[
{
"refId":"A",
"expr":"histogram_quantile(0.95, sum by (le) (rate(llmgw_tenant_ttft_seconds_bucket{tenant=~\"$tenant\"}[5m])))",
"legendFormat":"",
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
}
}
],
"fieldConfig":{
"defaults":{
"unit":"s",
"thresholds":{
"mode":"absolute",
"steps":[
{
"color":"green",
"value":null
},
{
"color":"orange",
"value":3
},
{
"color":"red",
"value":10
}
]
}
},
"overrides":[]
},
"options":{
"reduceOptions":{
"calcs":[
"lastNotNull"
]
},
"textMode":"auto",
"colorMode":"value"
}
},
{
"id":5,
"title":"Requests by tenant",
"type":"timeseries",
"description":"Stacked, so the top edge is the whole of it.",
"gridPos":{
"x":0,
"y":4,
"w":12,
"h":9
},
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
},
"targets":[
{
"refId":"A",
"expr":"sum by (tenant) (rate(llmgw_tenant_requests_total{tenant=~\"$tenant\"}[5m]))",
"legendFormat":"{{tenant}}",
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
}
}
],
"fieldConfig":{
"defaults":{
"unit":"reqps",
"custom":{
"stacking":{
"mode":"normal"
},
"fillOpacity":30
}
},
"overrides":[]
},
"options":{
"legend":{
"displayMode":"table",
"placement":"right",
"calcs":[
"mean",
"max"
]
},
"tooltip":{
"mode":"multi",
"sort":"desc"
}
}
},
{
"id":6,
"title":"Tokens by tenant",
"type":"timeseries",
"description":"Prompt and completion together: what was consumed.",
"gridPos":{
"x":12,
"y":4,
"w":12,
"h":9
},
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
},
"targets":[
{
"refId":"A",
"expr":"sum by (tenant) (rate(llmgw_tenant_tokens_total{tenant=~\"$tenant\"}[5m]))",
"legendFormat":"{{tenant}}",
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
}
}
],
"fieldConfig":{
"defaults":{
"unit":"short",
"custom":{
"stacking":{
"mode":"normal"
},
"fillOpacity":30
}
},
"overrides":[]
},
"options":{
"legend":{
"displayMode":"table",
"placement":"right",
"calcs":[
"mean",
"max"
]
},
"tooltip":{
"mode":"multi",
"sort":"desc"
}
}
},
{
"id":7,
"title":"Refused by a limit, by tenant",
"type":"timeseries",
"description":"The share of each tenant's requests a limit turned away.",
"gridPos":{
"x":0,
"y":13,
"w":12,
"h":8
},
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
},
"targets":[
{
"refId":"A",
"expr":"sum by (tenant) (rate(llmgw_tenant_requests_total{tenant=~\"$tenant\",outcome=\"limited\"}[5m])) / sum by (tenant) (rate(llmgw_tenant_requests_total{tenant=~\"$tenant\"}[5m]))",
"legendFormat":"{{tenant}}",
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
}
}
],
"fieldConfig":{
"defaults":{
"unit":"percentunit"
},
"overrides":[]
},
"options":{
"legend":{
"displayMode":"table",
"placement":"right",
"calcs":[
"mean",
"max"
]
},
"tooltip":{
"mode":"multi",
"sort":"desc"
}
}
},
{
"id":8,
"title":"Failing, by tenant",
"type":"timeseries",
"description":"One tenant alone failing is usually what it sends; all of them together is a model.",
"gridPos":{
"x":12,
"y":13,
"w":12,
"h":8
},
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
},
"targets":[
{
"refId":"A",
"expr":"sum by (tenant) (rate(llmgw_tenant_requests_total{tenant=~\"$tenant\",outcome=\"error\"}[5m])) / sum by (tenant) (rate(llmgw_tenant_requests_total{tenant=~\"$tenant\"}[5m]))",
"legendFormat":"{{tenant}}",
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
}
}
],
"fieldConfig":{
"defaults":{
"unit":"percentunit"
},
"overrides":[]
},
"options":{
"legend":{
"displayMode":"table",
"placement":"right",
"calcs":[
"mean",
"max"
]
},
"tooltip":{
"mode":"multi",
"sort":"desc"
}
}
},
{
"id":9,
"title":"First token p95, by tenant",
"type":"timeseries",
"description":"",
"gridPos":{
"x":0,
"y":21,
"w":12,
"h":8
},
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
},
"targets":[
{
"refId":"A",
"expr":"histogram_quantile(0.95, sum by (tenant, le) (rate(llmgw_tenant_ttft_seconds_bucket{tenant=~\"$tenant\"}[10m])))",
"legendFormat":"{{tenant}}",
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
}
}
],
"fieldConfig":{
"defaults":{
"unit":"s"
},
"overrides":[]
},
"options":{
"legend":{
"displayMode":"table",
"placement":"right",
"calcs":[
"mean",
"max"
]
},
"tooltip":{
"mode":"multi",
"sort":"desc"
}
}
},
{
"id":10,
"title":"Budgets spent this period",
"type":"bargauge",
"description":"Every organization, team, project and person with a budget. Requests are refused at 100%.",
"gridPos":{
"x":12,
"y":21,
"w":12,
"h":8
},
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
},
"targets":[
{
"refId":"A",
"expr":"sort_desc(llmgw_budget_spent_ratio)",
"legendFormat":"{{scope}} {{name}} ({{period}})",
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
}
}
],
"fieldConfig":{
"defaults":{
"unit":"percentunit",
"min":0,
"max":1,
"thresholds":{
"mode":"absolute",
"steps":[
{
"color":"green",
"value":null
},
{
"color":"orange",
"value":0.8
},
{
"color":"red",
"value":1
}
]
}
},
"overrides":[]
},
"options":{
"displayMode":"gradient",
"orientation":"horizontal",
"reduceOptions":{
"calcs":[
"lastNotNull"
]
},
"showUnfilled":true
}
},
{
"id":11,
"title":"Who used most, last 24 hours",
"type":"table",
"description":"Spend in money is on the people page's usage tab, which reads it from the store.",
"gridPos":{
"x":0,
"y":29,
"w":24,
"h":8
},
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
},
"targets":[
{
"refId":"A",
"expr":"sort_desc(sum by (tenant) (increase(llmgw_tenant_requests_total{tenant=~\"$tenant\"}[24h])))",
"legendFormat":"requests",
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
},
"format":"table",
"instant":true
},
{
"refId":"B",
"expr":"sum by (tenant) (increase(llmgw_tenant_tokens_total{tenant=~\"$tenant\"}[24h]))",
"legendFormat":"tokens",
"datasource":{
"type":"prometheus",
"uid":"llmgw-prom"
},
"format":"table",
"instant":true
},
{
"refId":"C",
"expr":"sum by (tenant) (increase(llmgw_tenant_requests_total{tenant=~\"$tenant\",outcome=\"limited\"}[24h]))",
A model's licence is in `/v1/models` (`license`: its `id`, `name`, whether
it allows commercial use and the regions it grants nothing in). On your
profile page, under **Licences**, you can leave licences out: one by one,
every licence that forbids commercial use, or every licence that grants
nothing in a region you name (`EU` for the union). The rules cover models
added later too. From then on those models are not listed to your keys and
a request for one is refused with 403 and the code `licence_excluded`,
naming the choice that left it out — on every key you own and on the
portal's pages, whatever a key's own allowlist says. A change reaches keys
already in use within a minute.
### Headers llmgw sends back
| Header | Says |
|---|---|
| `x-request-id` | The request's id in the gateway's logs and usage records — quote it when asking an operator about a request. |
| `x-llmgw-served-model` | The model that answered, when it was not the one asked for: a spent budget can move a request to a cheaper model. |
| `x-llmgw-queue-ms` | How long the request waited for room before it was sent. |
| `x-llmgw-thread-trimmed` | How many stored turns were left out to fit the model. |
## When something is refused
Errors come back in the shape of the API that was called — OpenAI's
`{"error": {"message": …, "type": …}}` or Anthropic's — so each SDK raises
its usual exception with the gateway's message in it. The ones worth knowing:
| Status | Usually means |
|---|---|
| 400 | Besides a malformed request: a prompt too long for the model, or a guardrail policy refused it (the code `content_policy_violation`). |
| 401 | No key, a mistyped one, or one that was revoked or has expired. |
| 403 | The key may not use that model; with the code `licence_excluded`, its licence is one you left out on your profile page. |
| 429 | A rate limit or budget of the key, its project or its team was reached. `Retry-After` says when to try again, and is absent when waiting will not help — a spent monthly budget. |
| 503 | No model of that name is served (the message lists the ones that are), the model is still loading, or every replica is busy and the queue is full. The SDKs retry by themselves; a name that is not served will not start to be. |
| 504 | Every replica was busy for as long as the key's band may wait. `Retry-After` says when to send it again. |
## Priority
With keys on, the key decides the request's priority band and routing
profile, set by an operator; the `x-llmgw-band` and `x-llmgw-profile` headers
are ignored, so a client cannot promote itself. On a gateway without keys
For a site whose GPU nodes are joined by InfiniBand or RoCE, by NVLink, or both
— several independent fabrics, NVL72 racks, eight-GPU HGX boards. It says what
llmgw does with them, what it needs from you before it can, how to check a node
in one command, and where to look when a replica spanning nodes will not run.
Read [install.md](install.md) first; this is the part of it that only matters
once there is a fabric.
> **Where this stands.** Every piece here was written against the real formats
> of `sminfo`, `saquery`, `ibnetdiscover`, `nvidia-smi topo -m` and
> `nvidia-smi -q`, and tested against them — but none of it has yet run on a
> site with the hardware. Run `llmgw doctor --json` on two or three nodes first
> and keep the reports: if a parser meets a format it does not know, that report
> is what fixes it, and the fix is a control plane upgrade, not one per node.
## What llmgw does with them
**Two kinds of interconnect.** InfiniBand or RoCE joins nodes (scale-out).
NVLink joins GPUs — inside one machine on an HGX board, or across the nodes of a
rack on an NVL72 or an NVLink Switch system (scale-up). A replica spanning nodes
is placed, in order of preference:
1. inside one **multi-node NVLink domain**, if any has enough usable nodes:
tensor parallelism between its nodes then runs over NVLink;
2. on one **physical InfiniBand or RoCE fabric**, preferring nodes under one
leaf switch and putting nodes with a degraded cable last.
It is never split across two fabrics or two NVLink domains. Within a node, the
GPUs a shard gets are the set with the most NVLink pairs, then the nearest to
the fabric's adapters, then on one NUMA node.
**A fabric has a name and an identity.** The name is what `agent.yaml` says; the
identity is what the adapters report — on InfiniBand the subnet managers every
port on one fabric sees, on RoCE the IP network of its RoCE v2 address. The name
is how you address a fabric; the identity is how llmgw catches the name being
wrong: two independent fabrics given one name are placed as two, and the
cluster page says so.
## Before a node joins
On every GPU node:
| What | Why | Check |
|---|---|---|
| NVIDIA driver | everything | `nvidia-smi` |
| `nvidia-fabricmanager`, on an NVSwitch machine | without it CUDA does not start on those GPUs | `systemctl is-active nvidia-fabricmanager` |
| `rdma-core`, `ib_uverbs` loaded | user-space RDMA at all | `ls /sys/class/infiniband`, `lsmod \| grep ib_uverbs` |
| `nvidia_peermem` loaded | GPUDirect RDMA; without it NCCL copies every transfer through host memory | `lsmod \| grep nvidia_peermem` |
| `infiniband-diags` | tells llmgw which fabric a port is on; the agent installs it if missing | `which sminfo` |
On an NVL72 or NVLink Switch system, also:
| What | Why | Check |
|---|---|---|
| `nvidia-imex` running, every node of the domain in `/etc/nvidia-imex/nodes_config.cfg` | GPUs in different nodes share memory through it | `systemctl is-active nvidia-imex` |
| an IMEX channel (`options nvidia NVreg_CreateImexChannel0=1`) | an engine opens one | `ls /dev/nvidia-caps-imex-channels` |
| every GPU joined to its domain | the fabric manager does this | `nvidia-smi -q \| grep -A6 Fabric` shows `State: Completed` |
On each InfiniBand fabric:
- **A subnet manager, and a standby.** Without one, ports come up to `INIT` and
stay there; without a standby, the fabric keeps its routes if the master
stops but cannot recover from any change after. llmgw watches this but does
not run one.
- **Its own name** in `agent.yaml` on every node attached to it. Optionally, a
distinct `subnet_prefix` in each subnet manager's configuration, which makes
GIDs unique across fabrics; llmgw does not need it.
## Declaring the fabrics
In `/etc/llmgw/agent.yaml`, the adapters on each fabric, by name:
| `NCCL_IB_HCA` | that fabric's adapters on this node, and no others |
| `NCCL_IB_DISABLE` | `0` (`1` for a single-node shard) |
| `NCCL_SOCKET_IFNAME`, `GLOO_SOCKET_IFNAME` | the node's management interface: `mgmt_interface`, or the one holding its management address |
| `NCCL_IB_GID_INDEX` | on RoCE only: the RoCE v2 entry with an IPv4 address |
| `NCCL_MNNVL_ENABLE`, `NCCL_CUMEM_ENABLE` | `1`, for a replica inside a multi-node NVLink domain |
A profile's `env` overrides any of them. vLLM launches a multi-node replica
with its own multiprocessing (`engines.vllm.launch_mode: mp`, the default):
every node runs `vllm serve` with its rank and the leader's address, the
workers headless. `launch_mode: ray` is there for a vLLM too old for that.
## When a replica spanning nodes will not run
| It | Look at |
|---|---|
| is not placed | `llmgw ctl explain-placement <model>`: a fabric split by identity, a node whose IMEX is down, a node refused because its adapters are cabled to two fabrics |
| starts and hangs on its first collective | `llmgw ctl fabric-test` between two of its nodes; the cluster page's fabric problems; `doctor` on each node — a port down, the wrong interface, RoCE without a v2 GID |
| runs, slowly | `nvidia_peermem`; the map for a degraded cable; the node page's NVLink matrix for a group that crossed PCIe |
| worked yesterday | the subnet manager table for a failover; `link_downed` on the node's ports |
- **Feeds.** For a domain with two feeds the cap is the smaller feed's rating
times its derating (typically 80%), not the sum.
- **Already over.** A domain found over its cap, because the estimate was low
or the cap was lowered, gets no new placements, and an alert. It is not
emptied automatically; shedding power is the facility's decision, and
stopping inference to do it is ours only if asked (see shedding below).
### Temperature and cooling
Thresholds are per domain with defaults, and each condition needs to hold for a
while before it changes, in both directions, so a single reading neither
starts nor ends anything:
| Condition | Enters when | Placement |
|---|---|---|
| `ok` | below the warm threshold for 5 min | normal |
| `warm` | inlet above the warm threshold for 2 min, GPU temperatures near throttling, or any GPU thermally throttled | allowed, scored down: other domains first |
| `hot` | inlet above the hot threshold for 2 min, a cooling unit in alarm, or a sustained share of the domain's GPUs thermally throttled | no new placements, as a cordon would, but automatic and reversible |
| `unknown` | no fresh reading | as `ok`, and reported as unknown |
**Shedding** (optional, off by default): a domain `hot` for longer than its
shed delay has load moved off it using step 2's machinery. The lowest class goes
first, and replicas above their model's minimum go before any at it. A critical
minimum is never touched, and moves are rate limited to one replica per domain
per interval. The replica is placed elsewhere first where there is room, so
shedding moves work rather than dropping it. GPU power capping (`nvidia-smi
-pl`) is a gentler alternative worth measuring against shedding: less
throughput, no restart.
### Spreading
Replicas of one model are spread across racks, and across PDUs, the way they
are spread across nodes today. This is a preference, dropped when it is the only
thing standing in the way, as node spreading already is.
### Scoring
Placement today filters, then takes the first candidate in a fixed order
(fewest free GPUs for GPU profiles, fewest replicas for CPU ones, nodes holding
the weights first). Facility signals are better as a score than as more
filters. Candidates that pass the hard limits are ordered by:
1. condition: `ok` before `warm`;
2. power headroom left in the rack after this replica, more first, so load
levels across racks;
3. the existing order.
Once measurements include power, step 4's profile preference can use tokens per
watt instead of tokens per GPU where both are known.
## Where the data comes from
A `FacilitySource` interface, implemented several ways, chosen in
`server.yaml`:
```go
type FacilitySource interface {
// Domains is what the source knows: budgets, thresholds, members.
Domains(ctx) ([]Domain, error)
// Readings is the latest value of each signal, per domain.
Readings(ctx) ([]Reading, error)
}
```
- **Static file.** Domains, caps and thresholds, written by hand or generated
from a DCIM export. Enough for power budgets with no live data at all.
- **Prometheus.** DCIM, PDU (SNMP exporter) and BMC (Redfish exporter)
readings usually end up in Prometheus. The source runs configured instant
queries and maps their labels onto domains:
```yaml
facility:
source: prometheus
prometheus: http://monitoring:9090
signals:
inlet_temp_c:
query: max by (rack) (dcim_rack_inlet_temperature_celsius)
domain_label: rack
stale_after: 2m
power_watts:
query: sum by (pdu) (pdu_outlet_power_watts)
domain_label: pdu
cooling_alarm:
query: max by (row) (crac_alarm)
domain_label: row
```
- **Agents (built, step 0).** What each node reads about itself, every ten
seconds at most, sent with its heartbeat:
- **GPUs, from nvidia-smi:** temperature, memory (HBM) temperature where the
driver reports it, power draw, the enforced power limit, and what is
holding the clocks down now — `sw_power_cap`, `hw_slowdown`, `sw_thermal`,
`hw_thermal`, `power_brake`.
- **CPUs, from hwmon** (what lm_sensors reads, without it installed): the
package temperature, or the hottest core; and the board's inlet or ambient
sensor where it has one (`SYSTIN`, `Inlet`, `Ambient`).
- **CPU power, from RAPL:** the packages' energy counters over time,
subdomains not counted twice, counter wrap handled.
A machine without a sensor reports nothing for it, never a zero; a VM has
none of them. Published as `llmgw_gpu_temperature_celsius`,
The control plane runs in Kubernetes, from the Helm chart in
`deploy/helm/llmgw`. The GPU nodes do not: they stay ordinary machines,
joined with `llmgw node add` as on any other cluster. This page says why,
and how the two meet.
## What goes where
| Part | Where | Why |
|---|---|---|
| Control plane: the API on 8080, the agents' port 7443, the scheduler | A StatefulSet, from the chart | A Go binary with a store and three ports: what Kubernetes runs well |
| Store | A file on a PersistentVolumeClaim (SQLite), or your PostgreSQL | Either, as on a host |
| GPU and CPU nodes | Outside, as hosts | See below |
| Model mirror | Outside, on a machine with a large disk | Every node pulls tens of gigabytes from it over plain HTTP; a pod with a volume that size is a host in all but name |
### Why the nodes stay outside
The agent on a node is not a service that happens to use a GPU. It installs
the engines on the machine — the llama.cpp build for its driver, vLLM in a
Python environment — checks them against their publisher's checksums and
runs them; it chooses which GPUs each replica gets, by index and by NVLink
topology; it starts engines as processes that outlive its own restarts, so a
node upgrades while it serves; it writes the machine's firewall; and it
replaces its own binary when the cluster upgrades. All of that is root on
the host, which is exactly what a container is built to prevent.
Run as a DaemonSet it would need to be privileged, with the host's network,
process namespace and filesystem, and the NVIDIA device plugin would have to
be kept off its GPUs — or the two would hand the same card to two owners.
What is left is the host agent with a container's indirection around it, and
none of the scheduling, isolation or restart guarantees that make a pod
worth having. So the chart does not ship one, and this is the
recommendation:
- **Keep GPU machines out of Kubernetes**, and add them with `node add`.
- **If they are already Kubernetes workers**, dedicate them: taint them so
no other pod lands there, keep the NVIDIA device plugin off them (most
installs select nodes by label — leave the label off), and run the llmgw
agent on the host, where `node add` puts it. llmgw then owns the GPUs, and
Kubernetes schedules nothing that would want them.
## The image
The project publishes binaries, not an image. Build one from the repository
and push it where the cluster pulls from — at an offline site, the internal
From empty servers to a model answering requests, in six steps. This page
has only the commands and what to check after each; [install.md](install.md)
explains every choice.
## What you need
| | |
|---|---|
| **A control plane** | One machine, or three for high availability (the store replicated, either one able to take over). Linux, x86-64 or ARM64: Debian 12, Ubuntu 22.04+ or RHEL 10. No GPU needed; 2 GB RAM, 20 GB disk each — 8 GB RAM when it is also a node running the starter model and the two judges. |
| **GPU nodes** (optional to start) | The same Linux, with the NVIDIA driver installed. vLLM needs driver 525 or newer and a Turing GPU or newer; a V100 gets an older vLLM. |
| **root** | on the control plane. On the other two of a three-machine setup, `sudo` (its password is asked for if it needs one); on nodes, `sudo` without a password. |
| **SSH** | from the control plane to the other machines. The setup screens take a key, the ssh agent or a password; `node add` uses ssh keys. A machine nothing may log into joins with a code instead (`llmgw join`). |
| **Network** | nodes reach the control plane on 7443 and 8081; users reach it on 8080. The machines need internet access to download engines and models. |
You do not install llama.cpp, vLLM, Python or CUDA. The nodes do that
Any OpenAI client works. Only the base URL changes:
`http://<control-plane>:8080/v1`.
**An editor:** the account page connects Continue, Codex or Cline. Pick the
tool and a model, and it gives a one-line command that writes the
configuration.
**Other people's accounts:** they sign up at `/signup`, and you approve them on
the people page (`/admin/users`).
## 6. Put it behind HTTPS
Port 8080 is plain HTTP. Before users outside a trusted network use it, put a
reverse proxy in front with a real certificate. With Caddy, which gets the
certificate by itself:
```
sudo apt-get install -y caddy # Debian and Ubuntu
curl -fsSL $U/Caddyfile | sed 's/llm.example.com/YOUR.DOMAIN/' | sudo tee /etc/caddy/Caddyfile
sudo systemctl reload caddy
```
Keep 7443 (nodes) and 8081 (mirror) on the internal network. They never go
through the proxy. There is an nginx configuration too:
[deploy/proxy/](../deploy/proxy/).
## When something does not work
| What you see | What to do |
|---|---|
| `node add` stops at `join` | `sudo journalctl -u llmgw-agent -n 50` on the node. Usually it cannot reach port 7443 on the control plane. |
| `node add` says the address is not in the certificate | Run `sudo llmgw server init --host <that address>`, then `node add` again. |
| A node says it cannot run vLLM | The reason is on `/admin`: the driver is older than 580, the GPU is older than Turing, or there is less than 25 GB free. |
| A model shows `not placed` | `llmgw ctl explain-placement <model>` names each node and why it was not used. |
| A reply stops after 60 seconds behind nginx | Use the nginx configuration from the release. Its timeouts are set for slow models. |
| Anything else about a node | `sudo llmgw doctor` on it. `--json` writes everything it read, to send to whoever helps you. |
## Next
- [install.md](install.md): the same steps with every option explained, and
running the cluster after: monitoring, notifications, single sign-on,
backups, upgrades.
- [console.md](console.md): everything the console does.
- [fabrics.md](fabrics.md): InfiniBand and NVLink.
- [onboarding.md](onboarding.md): accounts, projects, budgets and keys.
{"NVRM version: NVIDIA UNIX x86_64 Kernel Module 570.86.15 Thu Jan 23 21:30:38 UTC 2025\nGCC version: gcc version 12.2.0\n","570.86.15"},
{"NVRM version: NVIDIA UNIX Open Kernel Module for x86_64 580.82.07 Release Build (dvs-builder@U16-I3-B03-4-3) Wed Aug 27 00:51:12 UTC 2025\n","580.82.07"},
{"NVRM version: NVIDIA UNIX Open Kernel Module for aarch64 580.95.05 Release Build\n","580.95.05"},
if_,err:=ParseSMIR("ibwarn: can't open UMAD port");err==nil{
t.Error("an error should not parse")
}
}
Loaded 100 of 1190 files, more files were not shown because too many files have changed in this diff.
Show more
Reference in new issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
