The mirror page's form began every GGUF model at 4096 tokens, the
console's at 4096 or the plan's figure. Both now start at what the model
was trained for, cut to the most whose KV cache fits beside the weights
where the form places it; the mirror page's slots are as many of that, up
to four, as the room left holds. Plan carries the trained context.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each agent measures one core's memory rate beside the whole machine's; the
estimate takes a replica's threads at about one core's rate each, up to the
machine's, calibrated by llama-bench on the PoC. The serve form shows it
beside a card's, warns when slow, and asks for the CPU on purpose when a
reply would take minutes; ctl model plan, the console and serve jobs say
it too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
llmgw ctl model gpus MODEL N [--profile P] [--in-place],
POST /api/v1/models/{name}/gpus and the console's GPUs…: how many cards of
one node each replica uses. Tensor parallel for vLLM, SGLang and
TensorRT-LLM, held to the model's KV heads (modelplan.SplitsHeads, now one
rule for the planner and this); the layers split for llama.cpp, its old
--tensor-split dropped. Refused for a count no node has, for cards that
would not hold the weights and the context, and for engines on one card.
The replicas start again beside the old ones; with no room beside them the
command says so, and --in-place stops each first.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ctl model context MODEL TOKENS [--slots N], POST /api/v1/models/{name}/context
and the console's Context length…: the catalog, the engine's arguments and
the KV memory change together, refused past the trained context or what the
cards hold, then the replicas restart one by one.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
From review. A MIG instance is given its share of the card's bandwidth,
not the whole card's. NVLink between a layout's cards means all of them
joined to each other, the largest such set on one ready node, not a count
of cards with any peer: two bridged pairs are not four on NVLink. GH200 is
matched before H200. An FP8 checkpoint is listed for cards of 8.0 and up,
which vLLM runs with Marlin; 8.9 is the agent's bound for --quantization
fp8. NVFP4 repositories tagged fp4 are sized at four bits.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
For each layout modelplan.Options offers here — one card, tensor parallel
across a node's cards, llama.cpp's layer split, a pipeline across nodes on
a fabric — modelplan.GPUTokensPerSecond gives one reply's speed: the bytes
a token reads over each card's bandwidth, by its model name (a table, SXM
and PCIe apart), plus two all-reduces a layer for tensor parallel (8 us on
NVLink, 25 us and up on PCIe), turns between cards for a layer split, and
a hop per node over InfiniBand or RoCE for a pipeline. The explorer and a
repository's files show the fastest, with the layout; mixed classes are
judged class by class, FP8 on 8.9 cards only.
Calibrated on cuda01's Quadro RTX 6000 (672 GB/s): qwen3-vl-8b Q8, 8.7 GB,
62.1 tok/s measured, 59 estimated; Qwen3-Coder-30B-A3B Q4_K_M, 18.6 GB,
144 measured, which says a model of experts reads 1.6 times its active
share (ActiveShare now does, on the CPU too). Tensor parallel and fabrics
are not measured here.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each agent measures its node's memory read rate when it starts: every CPU
sums a share of a buffer past any cache, a tenth of free memory at most,
the best of four passes, in a fraction of a second; reported in Host
(mem_bandwidth_gbs) and NodeView. A token on a CPU reads every weight it
uses once, so one reply's speed is about that rate over those bytes:
modelplan.CPUTokensPerSecond, 0.68 of the measured rate plus 5 ms a token,
calibrated on the PoC against llama.cpp's predicted_per_second (node01
24.9 GB/s, qwen2.5-1.5b: 14.5 measured, 14.7 estimated; node05 41.9 GB/s,
qwen2.5-14b: 3.3 and 3.2). A model of experts named …-A3B reads its active
part (ActiveShare).
The explorer shows it as CPU tok/s, at Q4_K_M on the fastest node that
holds the model; a repository's page gives it for each GGUF by its size.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A mixture-of-experts model on more than one card is served with vLLM's
--enable-expert-parallel; config.json's experts are read by any of their
names, and a model that also reads pictures is sized by its text_config.
Tensor parallelism past the KV heads is allowed when it is a multiple of
them (Qwen3-235B, four KV heads, runs on eight cards), which the plan
had refused. Several nodes are offered in any number, not only powers of
two. Each way says what it needs of a card, part by part, and of all its
cards; the serve form checks a context longer than planned still fits.
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>
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>
Reading the tool's own output against the real fleet: it chose q8_0 as the
best fit and emitted a vLLM profile with no quantisation. q8_0 is a GGUF
quantisation that only llama.cpp serves, and from a file rather than a
repository — so following that entry loads full-precision weights, which
the same output had just listed as not fitting one card. The planner would
have caused the out-of-memory it exists to prevent, from a recommendation
that looked right.
Choosing a precision and choosing an engine turn out to be one decision, so
a Precision now carries the engine that serves it, the quantisation flag
where one is needed, and whether it lives in a file. A GGUF fit prints a
path and says to download it; a vLLM fit prints the repository and passes
--quantization so the engine does not quietly load bf16.
The arithmetic was right and the recommendation was wrong, which is the
failure mode worth having tests for: everything reported a number, and the
numbers agreed with each other.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Running the planner against the real fleet produced twenty-six rows for a
32B, of which thirteen were duplicates: two fabrics of identical cards, and
every single-node fit listed once for each.
A replica on one node does not span anything, so which fabric its node
happens to be attached to changes nothing about whether the model fits. A
multi-node fit is still listed per fabric, because there the fabric is the
constraint and which one has room is the thing the operator needs told.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
llmgw ctl model plan Qwen/Qwen2.5-Coder-32B-Instruct
The arithmetic this replaces is the last thing between having GPUs and
serving a model on them. Choosing tp, gpu_class and a context that will
actually fit means knowing the layer count, the KV head count, what the
weights weigh at a chosen quantisation and how much of a card the runtime
keeps — and getting it wrong is not an error message, it is an
out-of-memory twenty minutes into a load. I computed those by hand for
every model on this cluster.
Both halves are read rather than retyped: the architecture from the model's
own config.json and parameter count, the hardware from the running fleet.
Attention heads in place of KV heads is the classic slip and is worth a
factor of five on a model with grouped-query attention, so it is not a
number anyone should be entering.
Every candidate is listed, not one answer. With cards to spare you want the
full-precision fit; with one card free you want to know a quantised one
exists. Choosing between those is not the planner's job.
The tests are calibrated against files actually downloaded here — llama.cpp
reported 14.62 GiB and 8.37 GiB for the 14B at q8_0 and q4_K_M, and the
model reproduces both — and they caught two errors in my own arithmetic.
The overhead was charged per rank from the whole model's weights rather
than each rank's share, and it reserved a tenth of a card that
usableFraction had already reserved. Between them, Llama-3.3-70B at bf16
across two 80 GB cards came out as not fitting, which is a configuration
people run every day. Slack reserved in two places, each unaware of the
other, reads as caution and behaves as a bug.
Pools are grouped by class and fabric together, because a replica may never
span two fabrics: eight free cards split across two of them are not eight
cards for a model that needs eight, and merging them would plan something
that starts, passes its health checks and hangs on its first all-reduce.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two things a cluster with many nodes and many GPUs needs that were missing,
and they are the same thing seen from two distances.
Idle tracking. The control plane knew what was placed and whether it was
healthy, never whether anyone wanted it — so on a fleet with more models
than cards there was no way to answer the first question anyone asks, which
is what can I stop. The router is the only thing that knows, so it now
reports it: a timestamp written in Acquire, where a request already meets an
endpoint under a lock already held, and collected once a tick as a map. That
shape is the point. Per-request notification would be a firehose carrying
one bit — the newest timestamp wins — and this costs the same on four
replicas as on four hundred.
It travels through an interface in the direction opposite EndpointSink, for
the same reason that one exists: the control plane decides what should run,
the router knows whether anyone wants it, and neither should need to be in
the other's process to say so.
Unknown is kept distinct from idle throughout. A router that has just
restarted knows nothing, and a zero timestamp would be the oldest idle
replica in every comparison — so one restart would nominate an entire fleet
for eviction. Absent from the map, "-" in the table, negative in the JSON,
and IdleFor returns a second value rather than a duration anything could
act on by mistake.
ctl status. A per-node listing is right for four nodes and wrong for two
hundred: how full am I, what is broken, what is holding cards nobody is
using are aggregates, and reading them off a long table is a chore, which
means it gets skipped. Capacity per GPU class with usage bars, free whole
nodes per fabric — the number that decides whether a multi-node replica can
land, since free cards on the wrong fabric are not capacity for a model that
spans nodes — replica and model counts, models that cannot be placed, and
what has gone unasked-for. Derived from the same state the tables show, so
it cannot disagree with them.
Unhealthy GPUs count as neither used nor free, because counting them free is
how a cluster looks like it has room it does not.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
