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>
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>
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>
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>
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>
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>
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 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>
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>
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>
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>
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 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>
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>
Three places where the system took the infrastructure's word for something
and would have failed silently on real hardware.
A declared fabric adapter that is not on the node. Fabric membership has to
be configuration — OpenSM's default subnet prefix is identical on every
fabric, so a matching prefix proves nothing — but taking it entirely on
trust was a different mistake. A typo, or a card that enumerated
differently after a reboot, produced NCCL_IB_HCA naming an adapter that
does not exist.
An adapter with no ACTIVE port. Physically linked with no subnet manager
sits in INIT: it looks healthy to anything counting adapters, carries no
data, survives a reboot, and looks like a software problem. Discovery had
recorded port state all along and placement only ever asked whether a
fabric had any adapters listed.
Both end the same way — the replica starts, passes its health checks, and
hangs on its first all-reduce, which is the failure the fabric invariant
exists to prevent arriving through a door nobody was watching. The agent
now compares declared against discovered, where both are visible for the
only time, and does not advertise a fabric that cannot carry traffic. The
reason is logged on the machine that knows it, because the operator who
typed the adapter name is the only person who can fix it. The fabric's rate
is the slowest active link, since a collective runs at the speed of its
slowest.
And NVLink topology, which no published binary could see. NVML is its only
source, the bindings need cgo, and CGO_ENABLED=0 is deliberate so the
portable binary runs on an older libc. The cost was silent: tensor groups
fall back to the lowest free indices, which is right on an SXM board with
NVSwitch and wrong on a PCIe box where bridges pair specific cards — a
group straddling two of them runs several times slower and nothing says so.
Releases now also publish an nvml build for those nodes, and
llmgw_gpu_nvlink_peers makes the difference visible: a flat zero across a
fleet means the topology is unknown rather than absent.
A test that declared mlx5_0 and mlx5_1 on a node whose sysfs held neither
had been passing, because nothing compared the two.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The router published twenty-two metric families about requests, endpoints,
admission and scoring — everything about traffic and nothing about the
machines carrying it. There was no way to graph GPU usage, watch a node go
unhealthy, see a fabric fill up, or answer "was the cluster full at the
time", which is the first question anyone asks after an incident. The
agents exported nothing at all.
Exported by the control plane rather than by each agent. It already holds
every number: hardware arrives at registration and state on each heartbeat,
so this is a map read away. One scrape target for a fleet of hundreds, no
Prometheus configuration per node, no firewall rule per machine, and
nothing new running on the nodes. On the same /metrics as the traffic
figures, because correlating "the cluster was full" with "requests were
queueing" should not need two systems.
A collector rather than gauges kept current by a ticker: the numbers are
derived from state that already exists, so computing them when asked is
simpler than maintaining a copy that can drift, and nothing runs between
scrapes.
GPU utilisation is new, and is the number people actually watch. Memory
used says a model is loaded; this says whether it is doing anything, and a
fleet at ninety percent occupancy and two percent utilisation is a fleet
holding models nobody is asking for.
Adding it caught two things. utilization.gpu goes last in the nvidia-smi
query and is optional, because a driver too old to know it returns a
shorter row — requiring it made an llmgw upgrade silently stop discovering
GPUs on an older node, which is far worse than not knowing how busy a card
is. And I gave the new protobuf field number 14, which ecc_errors already
had; buf lint is why that was a moment rather than a corrupted wire format.
Absent is kept distinct from zero throughout, for the same reason as
everywhere else here: an unreported utilisation shown as a flat zero reads
as an idle fleet rather than a missing measurement, and a replica nothing
has reported would be the most idle thing in the cluster.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Storage will be different at every installation — Lustre, NFS over RDMA, an
image exported with SRP or NVMe-oF, local NVMe, an rsync — so llmgw observes
it rather than managing it. An agent told which directories to look in
reports what it can see; how the files got there is nobody's business here,
and a node told nothing reports nothing and behaves exactly as every node
did before this existed.
That fact buys two things. Placement prefers a node that already holds the
model over one that must first move tens of gigabytes, as a preference and
never a filter: fetching them is a valid answer, and refusing to place would
turn a slow start into no start at all. And ctl status shows what each
fabric holds.
Per fabric because fabrics are not connected to one another. Weights in
ib-a are not weights in ib-b — the only path between them is the management
network — so a model seeded into one fabric leaves the rest of the cluster
unable to serve it without a slow copy, and nothing said so before now. One
shared filer counts the same as twenty local copies, because "can this
fabric serve it" is the question being asked.
Three distinctions the tests hold to, because each is a way to get this
quietly wrong. A node that was never asked to look must rank with one that
looked and has nothing, or enabling the feature on a single node would
drain every other. An empty directory is not weights, because a
half-finished download looks exactly like one and would send placement to a
node whose engine then fails to load. And local storage is never mistaken
for shared: a filer read as local costs an unnecessary copy, while local
read as shared would have the cluster believe a model is everywhere when it
is on one node.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Deploying the memory-aware placement showed node01 to node04 reporting 7G
and node05 reporting nothing, on a machine whose /proc/meminfo plainly says
32 GB. node05 has no nvidia-smi, so GPU discovery returns an error, Discover
returns with it, and the memory read — which has nothing to do with cards —
never happened.
That is the node whose memory matters most, because a CPU profile is what a
GPU-less node will be asked to run, and min_memory_gb would have excluded it
from everything.
Read first now, before any path that can return early. The test fails
against the old order.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Placement had no idea how much memory a node has or a model needs. A CPU
profile has no card to be free, so eligibleNodes accepted every ready node:
Qwen2.5-Coder 14B, whose working set is about 24 GB, would have been placed
just as happily on one of the 8 GB nodes, which reports itself perfectly
healthy right up until the OOM killer arrives.
node_selector was yesterday's stopgap and it is the operator restating by
hand something the system can know. Agents now report MemTotal, profiles
declare min_memory_gb, and placement filters on it, naming what was short —
"has 8 GB of memory, needs 24" is actionable where "no candidates were
considered" is not. The label selector stays, because "this node is for
coding models" is a real thing to want to say that memory does not express.
Two deliberate non-exclusions. A profile with no requirement fits anywhere,
which is every profile that exists today. And a node that reports no memory
is still usable: an agent older than the field says nothing, and letting a
control plane upgrade quietly empty half a fleet is a worse failure than the
one this prevents.
ctl nodes grows a MEM column showing "-" rather than 0 for a node that did
not say, because an old agent is not a small machine.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
One bidirectional connect-go stream per node, carrying registration, heartbeats
and shard status up, and start, stop and drain down. The agent dials out: GPU
nodes sit behind firewalls far more often than a control plane does, and an
outbound connection needs no inbound rule on every node in the fleet.
- registration carries the whole hardware inventory, so there is no separate
discovery call and a reconnecting agent refreshes what the scheduler knows.
The fabric labels travel with it; they cannot be inferred from the wire, so
losing them in transit would silently disable multi-node placement.
- agentsvc is both the stream handler and the reconciler's dispatcher, because
it holds the open streams and is therefore the only thing that can deliver a
command to a node.
- the agent validates a profile before launching anything, since an engine that
refuses a profile does so after minutes of loading, then supervises the
process and reports every transition with the engine's output tail on failure.
- reconnects forever with jittered backoff. A control plane restart must not
require touching every GPU node, and a fleet must not reconnect in lockstep
and knock it over again. Engines keep serving throughout, because the router
talks to them directly.
llmgw agent is real now, and llmgw server runs a control plane whenever any
model declares a profile. The static endpoint list stays for router-only runs.
Verified with real binaries as separate processes: the agent registers, a
replica is placed, an engine launches through the vLLM adapter's actual argument
construction, streaming traffic is served, and killing the agent moves the node
through unreachable to lost with placement explaining why nothing can run.
config.Model and catalog.Model were the same thing described twice, so the
router now reads catalog entries directly and the validation rules live in one
place.
Still missing: mTLS. The stream is plaintext HTTP/2 and must stay on a trusted
network until that lands.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
NVML is the only source of NVLink topology, and that topology is what keeps a
tensor-parallel group off the slow path inside a node. It needs cgo, so it sits
behind the nvml tag and the default build keeps using nvidia-smi.
Discover() picks the backend and falls back to nvidia-smi when NVML fails,
logging the fallback rather than performing it silently: on a node that is meant
to have working NVML, a quiet downgrade hides a broken driver installation and
costs the scheduler its topology.
CI now compiles the tagged build. That immediately paid for itself: ErrNoNVML
was declared only in the untagged file, so the tests did not compile under
-tags nvml. It is now declared where both builds can see it.
Not verified: the NVML path has never run against a driver, because there is no
GPU machine available. The NVLink domains it reports should be treated as
unverified until someone runs `llmgw agent inventory` on a real node.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Everything here works without GPUs and is covered by tests. The control plane
connection is not built yet, so llmgw agent itself is still a stub and the phase
exit criterion is not met.
- discovery: IB from /sys/class/infiniband, parsing the formats sysfs actually
uses and telling native InfiniBand from RoCE; GPUs via nvidia-smi, the
cgo-free path, with health flags for uncorrected ECC and memory already held
by another process. A node without IB is not an error, it is a single-node
island.
- engines: vLLM and SGLang command and environment construction, kept free of
process handling so it stays testable. The NCCL environment is the part the
plan marks critical: NCCL_IB_HCA lists only the chosen fabric's adapters,
because leaving it open lets NCCL pick an adapter on a fabric where the peer
is unreachable, and the job then hangs instead of failing.
- validation before launch: no mixed GPU types in a tensor-parallel group, the
node/GPU/TP arithmetic has to add up, and fp8 is refused below compute
capability 8.9 rather than silently falling back to a slower path.
- supervisor: process groups, because engines fork workers that hold GPU memory
and outlive a signal sent only to the leader; a bounded log tail, since engine
output is the only diagnosis when loading fails; Starting separate from
Loading, because a 70B model legitimately takes minutes; and orphan adoption
through pid files fingerprinted with the command line, since pids get reused
and adopting an unrelated process is worse than leaking the original.
- readiness is two questions: /health says the process is up, /v1/models says
the weights loaded. Routing on /health alone sends traffic into a loading model.
- llmgw agent inventory reports what the scheduler would see on a node.
One plan correction: GPU class sizes come from the driver in GiB, which does not
always match the marketing name. An L40S is sold as 48GB and reports 44.99 GiB,
so its class is l40s-45g, not l40s-48g. Consistency matters more than matching
the box, because that number decides whether a model fits. Plan §2 and §5.1
updated.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
