Sample GPU utilisation and memory every heartbeat, and chart them

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>
Choose each engine's build from the node's hardware, for every GPU generation
2026-10-01 15:42:17 +02:00 · 2026-10-01 14:35:58 +02:00 · 2026-10-01 14:20:11 +02:00 · 2026-10-01 14:04:45 +02:00 · 2026-10-01 12:39:14 +02:00 · 2026-10-01 12:21:04 +02:00
129 changed files with 7783 additions and 480 deletions
@@ -93,6 +93,24 @@ jobs:
          ./dist/llmgw-${version}-linux-amd64 --version
          cat dist/checksums.txt

+      # Signed when the repository has a signing key, so a machine that trusts
+      # it installs nothing the server alone could have put there. Without the
+      # secret nothing is signed and releases are checked on their checksums,
+      # as before; a build that trusts a key refuses those, so the secret has
+      # to be in place before trusted.pub is filled in.
+      - name: Sign the checksums
+        env:
+          RELEASE_SIGNING_KEY: ${{ secrets.RELEASE_SIGNING_KEY }}
+        run: |
+          set -euo pipefail
+          version="${GITHUB_REF_NAME:-$(git describe --tags --always)}"
+          if [ -z "${RELEASE_SIGNING_KEY:-}" ]; then
+            echo "no RELEASE_SIGNING_KEY secret; releases are not signed"
+            exit 0
+          fi
+          ./dist/llmgw-${version}-linux-amd64 release sign dist/checksums.txt
+          echo "signed: dist/checksums.txt.sig"
+
      - name: Attach to the release
        env:
          TOKEN: ${{ secrets.GITEA_TOKEN }}
@@ -20,6 +20,345 @@ is only released once its **exit criterion** is met, not when its code lands.
 Until v1.0.0 the minor version is the breaking one: configuration formats and Go
 APIs change between minors without a deprecation period.

+## Unreleased
+
+From the field report in issue #2, after a deployment on GPU nodes.
+
+### GPU metrics that move, and charts of them
+
+A GPU's utilisation and memory in use were read once, when the agent
+started, and published from then on: a card busy since looked as idle as it
+was at registration, and one idle since as busy. They are now sampled with
+the temperatures and power every heartbeat (ten seconds at most), with the
+memory bandwidth in use, the SM clock and uncorrected ECC errors besides.
+An agent too old to send them still has its registration figures published,
+and a figure the driver does not give is left out rather than shown as zero.
+
+- **New series**: `llmgw_gpu_memory_utilization_percent`,
+  `llmgw_gpu_sm_clock_mhz`; `llmgw_gpu_utilization_percent`,
+  `llmgw_gpu_memory_used_bytes` and `llmgw_gpu_ecc_errors` now from the
+  sample.
+- **The node page** shows each card's utilisation, memory, clock and ECC
+  errors in its table, and charts of utilisation, memory, temperature and
+  power, one line per card.
+- **The cluster page** charts the fleet's GPUs by class: average utilisation
+  and memory in use.
+- **Grafana** (`deploy/monitoring/grafana/llmgw-overview.json`) has a GPU row:
+  utilisation, memory, memory bandwidth, temperature, power against the
+  limit, clock, and what holds cards back.
+- **Alerts**: `GPUThermalThrottling` (ten minutes slowed by heat),
+  `GPUUncorrectedECCErrors`, `GPUHot` (above 87 °C for ten minutes).
+
+### A GPU placement no longer runs on the CPU unnoticed
+
+llama.cpp placed on a GPU, with its CUDA backend unable to load (cuBLAS not
+where the loader looks), came up on the CPU, answered at CPU speed, and was
+reported ready while its GPU sat reserved and idle. Three checks now stop it:
+
+- **Installing** the CUDA build asks it which devices it sees
+  (`llama-server --list-devices`). None, on a node with GPUs, is a failed
+  install, with what `ldd` says `libggml-cuda.so` is missing. The binary's
+  own `ldd` could not say it: the backend is opened at run time.
+- **After the upgrade**, engines the agent finds already running are held to
+  the same check when it adopts them: one serving a GPU placement on the CPU
+  is stopped, and its replacement starts through the check. The installed
+  CUDA build is checked again within ten minutes, and installed afresh if it
+  sees no GPU.
+- **Starting** a shard on GPUs is ready only when its process holds memory on
+  each of them (`nvidia-smi --query-compute-apps`, its process group
+  included). One that does not is stopped and reported failed: "placed on GPU
+  0 but computing on the CPU", with the devices the engine lists. This covers
+  a hand-installed binary (`engines.llamacpp.bin`), which nothing else checks.
+- **`llmgw ctl engine-check llamacpp`**, run on a GPU node, now also fails
+  when the engine sees no GPU, and otherwise prints the devices it computes on.
+
+Shards are also given `CUDA_DEVICE_ORDER=PCI_BUS_ID`: placement numbers GPUs
+as nvidia-smi does, by PCI bus, and CUDA's own order is fastest first — on a
+node with two kinds of GPU, a shard could get a different card from the one it
+was given.
+
+### A model of many files, mirrored whole for vLLM
+
+vLLM loads a directory — shards, config.json, the tokenizer — and does not
+read GGUF. Mirroring one could go wrong several ways: the explorer offered the
+shards alone, a node could be handed a version still downloading or missing a
+shard, and a repository's page given as a model's source was fetched as one
+file of HTML.
+
+- **The explorer fetches a safetensors set with what the engine reads beside
+  it**: config.json, generation_config.json, the tokenizer files, the shard
+  index, the quantisation config. Never code: a `.py` would be the engine
+  running what the repository ships, which is to be chosen on purpose.
+- **A version is offered to nodes only whole.** The mirror refuses its
+  manifest (409, with what is missing) while a file is still downloading, or
+  while a shard its `model.safetensors.index.json` names, or a part of a
+  split (`-00002-of-00004`), is absent. A node reports "not complete on the
+  mirror yet" and is tried again.
+- **Checksums go with the files.** The mirror keeps each file's sha256 beside
+  it and the manifest carries it; a node checks every file of a set against
+  it before using the set, and fetches one that differs again.
+- **`llmgw mirror add` takes several URLs** into one version.
+- **A Hugging Face repository page given as a source** is read as the
+  repository id for a model vLLM or SGLang runs, which the engine downloads
+  itself (the node then needs to reach Hugging Face), and refused for one
+  llama.cpp runs, which needs a file.
+
+With a vLLM profile of `tp: 4`, placement reserves four GPUs of one node for
+such a model, as before.
+
+### Signing up with an address that has an account
+
+On PostgreSQL, signing up with an address already registered showed
+"something went wrong; try again": the store recognised a duplicate by the
+word "UNIQUE", which is SQLite's, and PostgreSQL says it in lower case under
+code 23505. It now knows both, and the address is answered as on SQLite —
+with the same page as a new sign-up, so the page cannot be used to find out
+who has an account. That page now also says what to do if the address had
+one: sign in, or ask an operator to reset the password.
+
+An error the page does not explain says the account was not created and
+gives a reference (`err-…`), logged beside the error, for an operator to find
+it; the database's own words stay in the log.
+
+### An adapter no fabric uses does not keep a node out
+
+`llmgw doctor` failed a node for a RoCE port with no RoCE v2 GID entry, or for
+`ib_uverbs` not being loaded, whether or not any fabric used them; on a node
+added for single-node inference, with no fabric given, that refused it for
+RDMA nobody had asked for.
+
+- Those are failures only for ports a declared fabric uses, and only when a
+  fabric is declared. Otherwise they are warnings saying nothing uses them yet.
+- Which fabric an undeclared InfiniBand port is on, and whether it has a
+  subnet manager LID, are notes rather than warnings; `node add` still warns
+  once that adapters are up with no fabric given.
+- `node add --fabric` hands its fabrics to doctor (`doctor --fabric
+  name:mlx5_0,…`), so they are checked as declared even though agent.yaml is
+  written afterwards.
+
+### A restart with no room beside it
+
+`ctl restart` places each replacement beside the replica it replaces. With
+no room for it — one GPU, every card that fits held by the replica itself —
+the old replica was stopped only after the drain grace period, five minutes
+of "restart replacing" with nothing saying why, which read as a restart that
+never happened; scaling to zero and back was the way out. And a replacement
+that fit nowhere even with the old replica gone was stopped for all the same,
+leaving the model with nothing. Now:
+
+- **The replica is the only thing in the way**, found by placing the
+  replacement on a copy of the inventory without it: stopped at once, the
+  replacement taking its place.
+- **The replacement fits nowhere even then**: the restart is given up and the
+  replica keeps serving. `ctl replicas` shows `restart refused:` and why.
+- **Room being made** — a lower class preempted, an idle on-demand replica
+  stopping — is waited for, as before.
+- **`ctl restart --in-place`** stops each replica first, whatever room there is.
+
+### Every GPU node gets engine builds it can run, by itself
+
+Which build of an engine a node installs follows its hardware — the CPU
+architecture, the driver, and the oldest GPU on it — from what each project
+publishes and how it was compiled. Each engine has a list of builds, best
+first; the node takes the first its hardware meets, and the first of those
+that installs and is seen to compute on the GPU. A build the table is wrong
+about costs its download, and the next one is tried.
+
+| | llama.cpp | vLLM |
+|---|---|---|
+| Driver 580+, Turing and newer | CUDA 13.3 | the version asked for, CUDA 13 (PyPI) |
+| Driver 525–579 | CUDA 12.8 (x64): its machine code for 8.6, 8.9, 12.0; the GPUs it ships as PTX need 570 | the version asked for, CUDA 12.9 (vLLM's GitHub release, PyTorch's cu129 torch) |
+| Volta (V100) | CUDA 12.8 on driver 570+, else Vulkan | **0.19.1**, the newest vLLM built for 7.0: no build of 0.20 or later has its kernels |
+| Pascal, Maxwell | CUDA 12.8 on 570+, else Vulkan | none: no vLLM is built for them |
+| arm64 GPU (GH200, Grace) | CUDA 13.3, else Vulkan — never the CPU build | CUDA 13 or 12.9, 8.0 and newer |
+| No NVIDIA GPU | the CPU build | none |
+
+Read from the projects' release pipelines: vLLM's x86 builds start at 7.5
+from 0.20.0, its PyPI build is CUDA 13 from 0.20, 0.19.1 was the last
+compiled for 7.0 (PyPI's torch 2.10 is too); llama.cpp's CUDA 12.8 build
+carries 5.0–9.0 as PTX and 8.6/8.9/12.0 as machine code, its CUDA 13.3 build
+starts at 7.5.
+
+- **A node running an older vLLM says so**: the cluster page's *vLLM build*
+  column shows the version installed, and the installation's detail says why
+  ("vLLM 0.19.1, the newest built for compute capability 7.0, instead of
+  0.30.0"). Profiles choose GPUs by class, so a model meant for the newer
+  vLLM is not placed on a V100 by accident.
+- **The install computes on the GPU before it counts**: a small tensor
+  operation for vLLM, the devices llama.cpp lists for its GPU builds.
+- **The build can still be chosen per node** on the cluster page (auto, CUDA
+  13, CUDA 12.9, none), with `llmgw ctl engine-build NODE vllm BUILD`, or
+  `POST /api/v1/nodes/{id}/engines`; kept in the store and sent to the node
+  at once.
+- **Whether an engine computes on its GPUs is read from every process
+  nvidia-smi lists**, not compute processes alone: the Vulkan build is a
+  graphics one. A llama.cpp shard is given `GGML_VK_VISIBLE_DEVICES` beside
+  `CUDA_VISIBLE_DEVICES`.
+
+### Engines per node, and nodes that do not run them all
+
+A cluster with older GPUs could not add them cleanly: `node add` refused a
+V100 node for the vLLM it could not run, though it was meant for llama.cpp,
+and `--ignore-checks` passed over every other problem with it.
+
+- **What vLLM cannot do is a warning**, in `llmgw doctor` and `node add`: the
+  node joins without it and serves what it does run. `node add` checks the
+  vLLM the control plane installs, not the binary's own default.
+- **`node add --disable-engine vllm`** keeps an engine off a node for good:
+  written as `engines.disabled` in its `agent.yaml`, never installed or
+  advertised there, and not checked by doctor.
+- **`engine_overrides`** in `server.yaml` give nodes picked by GPU class, id
+  or labels a different version of an engine, or none — for exceptions; the
+  build each GPU generation needs is chosen without it (above).
+- **The requirements follow the version**, as the table above sets out; the
+  threshold had been CUDA 13's for every version.
+- **An engine binary named in `agent.yaml` is used.** The agent had installed
+  its own beside it, which was then used first.
+- The node page shows each GPU's compute capability.
+
+### Serving from the mirror says where a model will run
+
+The serve form put every GGUF model on the CPU: it offered no choice, and its
+"Profile name" field, pre-filled "cpu", read as though renaming it would move
+the model to a GPU. Placement is decided by `gpu_class`, which the form never
+set. Now:
+
+- **Runs on** lists the fleet's GPU classes with how many cards are free, and
+  for llama.cpp the CPU. On a fleet with GPUs the class with the most free
+  cards is chosen.
+- **GPUs per replica** sets how many cards a replica takes on one node
+  (`gpus_per_node` for llama.cpp, `tp` for vLLM), and **parallel requests**
+  how many a llama.cpp replica serves at once.
+- The profile is named for where it runs (`t4`, `t4-x4`, `cpu`) rather than
+  typed.
+- A line under the form says what the model will run as — engine, CPU or how
+  many of which GPU, memory, context pool and how it is divided — and turns
+  red when the CPU is chosen with GPUs free. The message after adding repeats
+  it.
+- Choosing the CPU while GPUs are free is refused unless **run on the CPU
+  anyway** is ticked.
+
+### llama.cpp on several GPUs of one node
+
+A llama.cpp profile takes several of a node's GPUs with `gpus_per_node: N`:
+placement reserves N cards of its `gpu_class` on one node, the engine sees
+those N (`CUDA_VISIBLE_DEVICES`) and spreads the layers over them
+(`--split-mode layer`, in proportion to each card's free memory, which for
+whole cards is even). `tp` or `pp` on a llama.cpp profile was a width nothing
+placed — the replica stayed unplaced or ran on one card. It is now read as
+`gpus_per_node`, so a profile written that way is placed on that many cards
+after the upgrade, rather than dropped; only both given and disagreeing, or
+`nodes` above one, is refused.
+
+## v0.10.6 — 2026-10-01
+
+Published first as release candidates v0.10.6-rc.1 to rc.8; rc.5 failed its
+format check and has no binaries.
+
+Tried on the PoC cluster before release: LoRA adapters asked for by name
+through the gateway, two on one model, each answering as trained and the model
+itself unchanged; a saturated pool spilling to the next (8 of 12 concurrent
+requests on the preferred replica, 4 on the other); and a bronze request
+displaced from a full queue by a gold one, with 503 and Retry-After 5. Not tried
+live: a model served from an external provider, which needs a provider key.
+
+### Capacity from the admin pages
+
+What v0.10.5 made settable from the command line is now on the pages too:
+
+- **Quotas** has a Capacity section: every scope's guaranteed and most GPUs,
+  cores and memory beside what it holds, with a form to set, change or remove
+  one. It goes through the same checks as `llmgw ctl quota`.
+- **A model's card** on the mirror page sets its class, its owner (when there
+  are quotas to count against), and the memory and cores a replica reserves,
+  beside its replica counts. Only what changed is applied, and each change is
+  said.
+- **Room counts what a class may take.** The card, `llmgw ctl model target` and
+  the headroom API count the replicas a model could get by stopping those its
+  class outranks, separately from free room, and name whose they would be. A
+  critical model's card no longer refuses a replica that preemption would make
+  room for.
+
+### LoRA adapters
+
+A llama.cpp model can carry `adapters`, each a name and a GGUF LoRA file. The
+nodes fetch them with the model's weights and llama.cpp loads them all at start,
+applied to nothing (`--lora-scaled FILE:0`); a request that names an
+adapter goes to the model's replicas with `"lora": [{"id": i, "scale": s}]`,
+so one replica serves the model and all its adapters. Adapter names share the
+namespace of models, are listed in `/v1/models`, and follow key allowlists on
+their own — a key allowed an adapter does not reach the model under it. Adapter
+files are kept from weight cleanup like the model's own. Not yet for vLLM or
+SGLang profiles, whose adapters are loaded differently.
+
+### Finding and attaching adapters from the mirror page
+
+- **Explore** has a *LoRA adapters* view: adapters for any model, or — from a
+  held model's card, *find adapters on Hugging Face* — those made for the model
+  its repository was made from, since a GGUF repository is a quantisation and
+  adapters are made for the original.
+- **Opening a repository** reads the first bytes of each GGUF file and marks it
+  a *LoRA adapter* or a *model*. Repositories tagged "lora" hold a model with
+  the adapter merged in about as often as an adapter, and only the file says.
+- **Held adapter files** are marked in the list, and a model's card lists its
+  adapters, offers the held adapter files of its architecture to attach under a
+  name and scale, and removes them. A file that is not an adapter, or is for
+  another architecture, is refused before it reaches the catalog.
+- **`llmgw ctl model adapters MODEL [--add NAME=URL[@SCALE]] [--remove NAME]`**
+  lists and changes them from the command line, through
+  `POST /api/v1/models/{name}/adapters`, with the same checks.
+
+### Models served outside the cluster
+
+A catalog model can be an OpenAI-compatible API elsewhere: `external: {url,
+model, api_key_file}` and no profiles. Nothing is placed for it; the router
+sends its requests to the provider, rewritten to the provider's name for the
+model and with its key, read from the file on every publish so a rotated key is
+picked up. Named in another model's `fallbacks`, it is used only when the
+cluster has no capacity left. It is never scraped for metrics, and never reads
+as stale. Prices, budgets, limits, usage and key allowlists apply as to any
+model — and since its prompts leave the cluster, the allowlist is how to decide
+who may send them (docs/install.md). `ctl models` and the cluster page mark it
+external.
+
+### Pool spillover
+
+A model served from pools of different priority — H100 before A100, say — was
+routed to the preferred pool however full it was. Now a pool whose replicas are
+saturated, by the same measure admission uses (queue past its healthy depth, or
+KV cache past its ceiling), passes requests to the next pool that is not. When
+every pool is saturated the preferred one is kept: a pool that is just as full
+gains nothing.
+
+### Displacement: a higher band is not refused while a lower one waits
+
+Each band's queue was bounded, and nothing bounded them together, so a full
+gateway turned an interactive request away while batch requests sat queued
+behind it. `admission.queue_capacity_total` now bounds everything queued —
+by default half of every band's capacity together. Reached, a request is not
+refused while a lower band has any queued: the newest request of the lowest
+band is turned away instead, with a 503 and `Retry-After: 5`, so it backs off
+and returns. The newest, because the oldest have waited longest and are
+nearest their TTL. Counted as `displaced` in the admission rejection metric.
+
+### Releases can be signed
+
+`llmgw release keygen` makes an ed25519 key pair. With the private key in the
+CI secret `RELEASE_SIGNING_KEY`, releases carry `checksums.txt.sig`; with the
+public key in `internal/release/trusted.pub`, every binary built after refuses
+a release not signed with it, wherever it installs one. Off until then, and
+docs/install.md says in which order to turn it on.
+
+### Restarts borrow a lower class's room
+
+A restart places the replacement beside the replica it replaces, and where
+there is no free room it now makes some the way any placement does: by
+stopping a replica its class outranks, which gets its room back once the old
+replica has gone. The in-place restart — the model unserved while its
+replacement loads — is left for when nothing of a lower class is running. On a
+cluster where every model is standard, that is every full one: mark the models
+that can wait `opportunistic` (`llmgw ctl model class`, or the model's card).
+
 ## v0.10.5 — 2026-09-30

 Published first as release candidates v0.10.5-rc.1 to rc.14; rc.5 failed its
@@ -1550,9 +1550,19 @@ type GPUSensor struct {
 	PowerLimitW float64                `protobuf:"fixed64,5,opt,name=power_limit_w,json=powerLimitW,proto3" json:"power_limit_w,omitempty"`
 	// throttle names what is holding the clocks down now: sw_power_cap,
 	// hw_slowdown, sw_thermal, hw_thermal, power_brake.
-	Throttle      []string `protobuf:"bytes,6,rep,name=throttle,proto3" json:"throttle,omitempty"`
-	unknownFields protoimpl.UnknownFields
-	sizeCache     protoimpl.SizeCache
+	Throttle []string `protobuf:"bytes,6,rep,name=throttle,proto3" json:"throttle,omitempty"`
+	// How busy the card is, sampled with the rest, rather than read once at
+	// registration. -1 is not reported by the driver. memory_total_mb is set
+	// by every agent that sends these, so zero there is an older agent.
+	UtilizationPercent       float64 `protobuf:"fixed64,7,opt,name=utilization_percent,json=utilizationPercent,proto3" json:"utilization_percent,omitempty"`
+	MemoryUtilizationPercent float64 `protobuf:"fixed64,8,opt,name=memory_utilization_percent,json=memoryUtilizationPercent,proto3" json:"memory_utilization_percent,omitempty"`
+	MemoryUsedMb             float64 `protobuf:"fixed64,9,opt,name=memory_used_mb,json=memoryUsedMb,proto3" json:"memory_used_mb,omitempty"`
+	MemoryTotalMb            float64 `protobuf:"fixed64,10,opt,name=memory_total_mb,json=memoryTotalMb,proto3" json:"memory_total_mb,omitempty"`
+	SmClockMhz               float64 `protobuf:"fixed64,11,opt,name=sm_clock_mhz,json=smClockMhz,proto3" json:"sm_clock_mhz,omitempty"`
+	// ecc_uncorrected counts uncorrected ECC errors since the driver loaded.
+	EccUncorrected float64 `protobuf:"fixed64,12,opt,name=ecc_uncorrected,json=eccUncorrected,proto3" json:"ecc_uncorrected,omitempty"`
+	unknownFields  protoimpl.UnknownFields
+	sizeCache      protoimpl.SizeCache
 }

 func (x *GPUSensor) Reset() {
@@ -1627,6 +1637,48 @@ func (x *GPUSensor) GetThrottle() []string {
 	return nil
 }

+func (x *GPUSensor) GetUtilizationPercent() float64 {
+	if x != nil {
+		return x.UtilizationPercent
+	}
+	return 0
+}
+
+func (x *GPUSensor) GetMemoryUtilizationPercent() float64 {
+	if x != nil {
+		return x.MemoryUtilizationPercent
+	}
+	return 0
+}
+
+func (x *GPUSensor) GetMemoryUsedMb() float64 {
+	if x != nil {
+		return x.MemoryUsedMb
+	}
+	return 0
+}
+
+func (x *GPUSensor) GetMemoryTotalMb() float64 {
+	if x != nil {
+		return x.MemoryTotalMb
+	}
+	return 0
+}
+
+func (x *GPUSensor) GetSmClockMhz() float64 {
+	if x != nil {
+		return x.SmClockMhz
+	}
+	return 0
+}
+
+func (x *GPUSensor) GetEccUncorrected() float64 {
+	if x != nil {
+		return x.EccUncorrected
+	}
+	return 0
+}
+
 // ShardLoad is one engine process group's use.
 type ShardLoad struct {
 	state   protoimpl.MessageState `protogen:"open.v1"`
@@ -2370,9 +2422,12 @@ func (x *Provision) GetEngines() []*EngineWant {

 // EngineWant names one engine at one version.
 type EngineWant struct {
-	state         protoimpl.MessageState `protogen:"open.v1"`
-	Engine        string                 `protobuf:"bytes,1,opt,name=engine,proto3" json:"engine,omitempty"`   // "llamacpp"
-	Version       string                 `protobuf:"bytes,2,opt,name=version,proto3" json:"version,omitempty"` // "b10991"
+	state   protoimpl.MessageState `protogen:"open.v1"`
+	Engine  string                 `protobuf:"bytes,1,opt,name=engine,proto3" json:"engine,omitempty"`   // "llamacpp"
+	Version string                 `protobuf:"bytes,2,opt,name=version,proto3" json:"version,omitempty"` // "b10991"
+	// variant is the build to install, when one is chosen for the node rather
+	// than left to its hardware: "cu13", "cu129". Empty is the agent's choice.
+	Variant       string `protobuf:"bytes,3,opt,name=variant,proto3" json:"variant,omitempty"`
 	unknownFields protoimpl.UnknownFields
 	sizeCache     protoimpl.SizeCache
 }
@@ -2421,6 +2476,13 @@ func (x *EngineWant) GetVersion() string {
 	return ""
 }

+func (x *EngineWant) GetVariant() string {
+	if x != nil {
+		return x.Variant
+	}
+	return ""
+}
+
 // ProvisionStatus is where one engine's installation stands on a node.
 type ProvisionStatus struct {
 	state   protoimpl.MessageState `protogen:"open.v1"`
@@ -2873,8 +2935,11 @@ type EngineProfile struct {
 	Env             map[string]string      `protobuf:"bytes,9,rep,name=env,proto3" json:"env,omitempty" protobuf_key:"bytes,1,opt,name=key" protobuf_val:"bytes,2,opt,name=value"`
 	ModelPath       string                 `protobuf:"bytes,10,opt,name=model_path,json=modelPath,proto3" json:"model_path,omitempty"`
 	ServedModelName string                 `protobuf:"bytes,11,opt,name=served_model_name,json=servedModelName,proto3" json:"served_model_name,omitempty"`
-	unknownFields   protoimpl.UnknownFields
-	sizeCache       protoimpl.SizeCache
+	// adapters are LoRA adapters loaded beside the model, in order: the index
+	// is the id a request applies one by.
+	Adapters      []*LoraAdapter `protobuf:"bytes,12,rep,name=adapters,proto3" json:"adapters,omitempty"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
 }

 func (x *EngineProfile) Reset() {
@@ -2984,6 +3049,13 @@ func (x *EngineProfile) GetServedModelName() string {
 	return ""
 }

+func (x *EngineProfile) GetAdapters() []*LoraAdapter {
+	if x != nil {
+		return x.Adapters
+	}
+	return nil
+}
+
 // RemoveWeights deletes cached model files a node no longer needs.
 //
 // The paths are computed by the control plane from the catalog, but the agent
@@ -3777,6 +3849,60 @@ func (x *UpgradeReply) GetError() string {
 	return ""
 }

+// LoraAdapter is one adapter: the name clients ask for it by, and its source,
+// which the agent fetches as it does the model's weights.
+type LoraAdapter struct {
+	state         protoimpl.MessageState `protogen:"open.v1"`
+	Name          string                 `protobuf:"bytes,1,opt,name=name,proto3" json:"name,omitempty"`
+	Source        string                 `protobuf:"bytes,2,opt,name=source,proto3" json:"source,omitempty"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
+}
+
+func (x *LoraAdapter) Reset() {
+	*x = LoraAdapter{}
+	mi := &file_llmgw_v1_agent_proto_msgTypes[36]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *LoraAdapter) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*LoraAdapter) ProtoMessage() {}
+
+func (x *LoraAdapter) ProtoReflect() protoreflect.Message {
+	mi := &file_llmgw_v1_agent_proto_msgTypes[36]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use LoraAdapter.ProtoReflect.Descriptor instead.
+func (*LoraAdapter) Descriptor() ([]byte, []int) {
+	return file_llmgw_v1_agent_proto_rawDescGZIP(), []int{36}
+}
+
+func (x *LoraAdapter) GetName() string {
+	if x != nil {
+		return x.Name
+	}
+	return ""
+}
+
+func (x *LoraAdapter) GetSource() string {
+	if x != nil {
+		return x.Source
+	}
+	return ""
+}
+
 var File_llmgw_v1_agent_proto protoreflect.FileDescriptor

 const file_llmgw_v1_agent_proto_rawDesc = "" +
@@ -3925,14 +4051,22 @@ const file_llmgw_v1_agent_proto_rawDesc = "" +
 	" \x01(\x01R\tcpuPowerW\x12 \n" +
 	"\finlet_temp_c\x18\v \x01(\x01R\n" +
 	"inletTempC\x12'\n" +
-	"\x04gpus\x18\f \x03(\v2\x13.llmgw.v1.GPUSensorR\x04gpus\"\xb5\x01\n" +
+	"\x04gpus\x18\f \x03(\v2\x13.llmgw.v1.GPUSensorR\x04gpus\"\xbd\x03\n" +
 	"\tGPUSensor\x12\x14\n" +
 	"\x05index\x18\x01 \x01(\x05R\x05index\x12\x15\n" +
 	"\x06temp_c\x18\x02 \x01(\x01R\x05tempC\x12\"\n" +
 	"\rmemory_temp_c\x18\x03 \x01(\x01R\vmemoryTempC\x12\x17\n" +
 	"\apower_w\x18\x04 \x01(\x01R\x06powerW\x12\"\n" +
 	"\rpower_limit_w\x18\x05 \x01(\x01R\vpowerLimitW\x12\x1a\n" +
-	"\bthrottle\x18\x06 \x03(\tR\bthrottle\"m\n" +
+	"\bthrottle\x18\x06 \x03(\tR\bthrottle\x12/\n" +
+	"\x13utilization_percent\x18\a \x01(\x01R\x12utilizationPercent\x12<\n" +
+	"\x1amemory_utilization_percent\x18\b \x01(\x01R\x18memoryUtilizationPercent\x12$\n" +
+	"\x0ememory_used_mb\x18\t \x01(\x01R\fmemoryUsedMb\x12&\n" +
+	"\x0fmemory_total_mb\x18\n" +
+	" \x01(\x01R\rmemoryTotalMb\x12 \n" +
+	"\fsm_clock_mhz\x18\v \x01(\x01R\n" +
+	"smClockMhz\x12'\n" +
+	"\x0fecc_uncorrected\x18\f \x01(\x01R\x0eeccUncorrected\"m\n" +
 	"\tShardLoad\x12\x19\n" +
 	"\bshard_id\x18\x01 \x01(\tR\ashardId\x12\x1b\n" +
 	"\tcpu_cores\x18\x02 \x01(\x01R\bcpuCores\x12(\n" +
@@ -4004,11 +4138,12 @@ const file_llmgw_v1_agent_proto_rawDesc = "" +
 	"\x03key\x18\x01 \x01(\tR\x03key\x12\x14\n" +
 	"\x05value\x18\x02 \x01(\tR\x05value:\x028\x01\";\n" +
 	"\tProvision\x12.\n" +
-	"\aengines\x18\x01 \x03(\v2\x14.llmgw.v1.EngineWantR\aengines\">\n" +
+	"\aengines\x18\x01 \x03(\v2\x14.llmgw.v1.EngineWantR\aengines\"X\n" +
 	"\n" +
 	"EngineWant\x12\x16\n" +
 	"\x06engine\x18\x01 \x01(\tR\x06engine\x12\x18\n" +
-	"\aversion\x18\x02 \x01(\tR\aversion\"\x8b\x01\n" +
+	"\aversion\x18\x02 \x01(\tR\aversion\x12\x18\n" +
+	"\avariant\x18\x03 \x01(\tR\avariant\"\x8b\x01\n" +
 	"\x0fProvisionStatus\x12\x16\n" +
 	"\x06engine\x18\x01 \x01(\tR\x06engine\x12\x18\n" +
 	"\aversion\x18\x02 \x01(\tR\aversion\x12\x18\n" +
@@ -4066,7 +4201,7 @@ const file_llmgw_v1_agent_proto_rawDesc = "" +
 	"profile_fp\x18\x0f \x01(\tR\tprofileFp\x12\x17\n" +
 	"\aapi_key\x18\x10 \x01(\tR\x06apiKey\x12\x1f\n" +
 	"\vcpu_threads\x18\x11 \x01(\x05R\n" +
-	"cpuThreads\"\xfd\x02\n" +
+	"cpuThreads\"\xb0\x03\n" +
 	"\rEngineProfile\x12\x12\n" +
 	"\x04name\x18\x01 \x01(\tR\x04name\x12\x16\n" +
 	"\x06engine\x18\x02 \x01(\tR\x06engine\x12\x1b\n" +
@@ -4080,7 +4215,8 @@ const file_llmgw_v1_agent_proto_rawDesc = "" +
 	"\n" +
 	"model_path\x18\n" +
 	" \x01(\tR\tmodelPath\x12*\n" +
-	"\x11served_model_name\x18\v \x01(\tR\x0fservedModelName\x1a6\n" +
+	"\x11served_model_name\x18\v \x01(\tR\x0fservedModelName\x121\n" +
+	"\badapters\x18\f \x03(\v2\x15.llmgw.v1.LoraAdapterR\badapters\x1a6\n" +
 	"\bEnvEntry\x12\x10\n" +
 	"\x03key\x18\x01 \x01(\tR\x03key\x12\x14\n" +
 	"\x05value\x18\x02 \x01(\tR\x05value:\x028\x01\"%\n" +
@@ -4143,7 +4279,10 @@ const file_llmgw_v1_agent_proto_rawDesc = "" +
 	"\n" +
 	"request_id\x18\x01 \x01(\tR\trequestId\x12\x18\n" +
 	"\aversion\x18\x02 \x01(\tR\aversion\x12\x14\n" +
-	"\x05error\x18\x03 \x01(\tR\x05error2O\n" +
+	"\x05error\x18\x03 \x01(\tR\x05error\"9\n" +
+	"\vLoraAdapter\x12\x12\n" +
+	"\x04name\x18\x01 \x01(\tR\x04name\x12\x16\n" +
+	"\x06source\x18\x02 \x01(\tR\x06source2O\n" +
 	"\fAgentService\x12?\n" +
 	"\aConnect\x12\x16.llmgw.v1.AgentMessage\x1a\x18.llmgw.v1.ControlMessage(\x010\x01B\x93\x01\n" +
 	"\fcom.llmgw.v1B\n" +
@@ -4162,7 +4301,7 @@ func file_llmgw_v1_agent_proto_rawDescGZIP() []byte {
 }

 var file_llmgw_v1_agent_proto_enumTypes = make([]protoimpl.EnumInfo, 1)
-var file_llmgw_v1_agent_proto_msgTypes = make([]protoimpl.MessageInfo, 42)
+var file_llmgw_v1_agent_proto_msgTypes = make([]protoimpl.MessageInfo, 43)
 var file_llmgw_v1_agent_proto_goTypes = []any{
 	(ShardStatus_State)(0),    // 0: llmgw.v1.ShardStatus.State
 	(*AgentMessage)(nil),      // 1: llmgw.v1.AgentMessage
@@ -4201,12 +4340,13 @@ var file_llmgw_v1_agent_proto_goTypes = []any{
 	(*FabricTestReply)(nil),   // 34: llmgw.v1.FabricTestReply
 	(*UpgradeAgent)(nil),      // 35: llmgw.v1.UpgradeAgent
 	(*UpgradeReply)(nil),      // 36: llmgw.v1.UpgradeReply
-	nil,                       // 37: llmgw.v1.Register.LabelsEntry
-	nil,                       // 38: llmgw.v1.Register.EngineVersionsEntry
-	nil,                       // 39: llmgw.v1.GPU.NicPathsEntry
-	nil,                       // 40: llmgw.v1.IBPort.CountersEntry
-	nil,                       // 41: llmgw.v1.Heartbeat.EngineVersionsEntry
-	nil,                       // 42: llmgw.v1.EngineProfile.EnvEntry
+	(*LoraAdapter)(nil),       // 37: llmgw.v1.LoraAdapter
+	nil,                       // 38: llmgw.v1.Register.LabelsEntry
+	nil,                       // 39: llmgw.v1.Register.EngineVersionsEntry
+	nil,                       // 40: llmgw.v1.GPU.NicPathsEntry
+	nil,                       // 41: llmgw.v1.IBPort.CountersEntry
+	nil,                       // 42: llmgw.v1.Heartbeat.EngineVersionsEntry
+	nil,                       // 43: llmgw.v1.EngineProfile.EnvEntry
 }
 var file_llmgw_v1_agent_proto_depIdxs = []int32{
 	3,  // 0: llmgw.v1.AgentMessage.register:type_name -> llmgw.v1.Register
@@ -4226,24 +4366,24 @@ var file_llmgw_v1_agent_proto_depIdxs = []int32{
 	31, // 14: llmgw.v1.ControlMessage.fabric_scan:type_name -> llmgw.v1.FabricScanRequest
 	33, // 15: llmgw.v1.ControlMessage.fabric_test:type_name -> llmgw.v1.FabricTestRequest
 	35, // 16: llmgw.v1.ControlMessage.upgrade:type_name -> llmgw.v1.UpgradeAgent
-	37, // 17: llmgw.v1.Register.labels:type_name -> llmgw.v1.Register.LabelsEntry
+	38, // 17: llmgw.v1.Register.labels:type_name -> llmgw.v1.Register.LabelsEntry
 	7,  // 18: llmgw.v1.Register.gpus:type_name -> llmgw.v1.GPU
 	8,  // 19: llmgw.v1.Register.fabrics:type_name -> llmgw.v1.FabricAttachment
-	38, // 20: llmgw.v1.Register.engine_versions:type_name -> llmgw.v1.Register.EngineVersionsEntry
+	39, // 20: llmgw.v1.Register.engine_versions:type_name -> llmgw.v1.Register.EngineVersionsEntry
 	6,  // 21: llmgw.v1.Register.running_shards:type_name -> llmgw.v1.RunningShard
 	9,  // 22: llmgw.v1.Register.weights:type_name -> llmgw.v1.LocalWeights
 	5,  // 23: llmgw.v1.Register.storage:type_name -> llmgw.v1.Storage
 	4,  // 24: llmgw.v1.Register.host:type_name -> llmgw.v1.Host
 	15, // 25: llmgw.v1.Register.imex:type_name -> llmgw.v1.IMEX
 	0,  // 26: llmgw.v1.RunningShard.state:type_name -> llmgw.v1.ShardStatus.State
-	39, // 27: llmgw.v1.GPU.nic_paths:type_name -> llmgw.v1.GPU.NicPathsEntry
+	40, // 27: llmgw.v1.GPU.nic_paths:type_name -> llmgw.v1.GPU.NicPathsEntry
 	12, // 28: llmgw.v1.NodeLoad.shards:type_name -> llmgw.v1.ShardLoad
 	11, // 29: llmgw.v1.NodeLoad.gpus:type_name -> llmgw.v1.GPUSensor
-	40, // 30: llmgw.v1.IBPort.counters:type_name -> llmgw.v1.IBPort.CountersEntry
+	41, // 30: llmgw.v1.IBPort.counters:type_name -> llmgw.v1.IBPort.CountersEntry
 	14, // 31: llmgw.v1.IBPort.subnet_managers:type_name -> llmgw.v1.SubnetManager
 	5,  // 32: llmgw.v1.Heartbeat.storage:type_name -> llmgw.v1.Storage
 	9,  // 33: llmgw.v1.Heartbeat.weights:type_name -> llmgw.v1.LocalWeights
-	41, // 34: llmgw.v1.Heartbeat.engine_versions:type_name -> llmgw.v1.Heartbeat.EngineVersionsEntry
+	42, // 34: llmgw.v1.Heartbeat.engine_versions:type_name -> llmgw.v1.Heartbeat.EngineVersionsEntry
 	20, // 35: llmgw.v1.Heartbeat.provision:type_name -> llmgw.v1.ProvisionStatus
 	10, // 36: llmgw.v1.Heartbeat.load:type_name -> llmgw.v1.NodeLoad
 	13, // 37: llmgw.v1.Heartbeat.ib_ports:type_name -> llmgw.v1.IBPort
@@ -4254,14 +4394,15 @@ var file_llmgw_v1_agent_proto_depIdxs = []int32{
 	0,  // 42: llmgw.v1.ShardStatus.state:type_name -> llmgw.v1.ShardStatus.State
 	22, // 43: llmgw.v1.ShardStatus.fetch:type_name -> llmgw.v1.FetchProgress
 	24, // 44: llmgw.v1.StartShard.profile:type_name -> llmgw.v1.EngineProfile
-	42, // 45: llmgw.v1.EngineProfile.env:type_name -> llmgw.v1.EngineProfile.EnvEntry
-	1,  // 46: llmgw.v1.AgentService.Connect:input_type -> llmgw.v1.AgentMessage
-	2,  // 47: llmgw.v1.AgentService.Connect:output_type -> llmgw.v1.ControlMessage
-	47, // [47:48] is the sub-list for method output_type
-	46, // [46:47] is the sub-list for method input_type
-	46, // [46:46] is the sub-list for extension type_name
-	46, // [46:46] is the sub-list for extension extendee
-	0,  // [0:46] is the sub-list for field type_name
+	43, // 45: llmgw.v1.EngineProfile.env:type_name -> llmgw.v1.EngineProfile.EnvEntry
+	37, // 46: llmgw.v1.EngineProfile.adapters:type_name -> llmgw.v1.LoraAdapter
+	1,  // 47: llmgw.v1.AgentService.Connect:input_type -> llmgw.v1.AgentMessage
+	2,  // 48: llmgw.v1.AgentService.Connect:output_type -> llmgw.v1.ControlMessage
+	48, // [48:49] is the sub-list for method output_type
+	47, // [47:48] is the sub-list for method input_type
+	47, // [47:47] is the sub-list for extension type_name
+	47, // [47:47] is the sub-list for extension extendee
+	0,  // [0:47] is the sub-list for field type_name
 }

 func init() { file_llmgw_v1_agent_proto_init() }
@@ -4296,7 +4437,7 @@ func file_llmgw_v1_agent_proto_init() {
 			GoPackagePath: reflect.TypeOf(x{}).PkgPath(),
 			RawDescriptor: unsafe.Slice(unsafe.StringData(file_llmgw_v1_agent_proto_rawDesc), len(file_llmgw_v1_agent_proto_rawDesc)),
 			NumEnums:      1,
-			NumMessages:   42,
+			NumMessages:   43,
 			NumExtensions: 0,
 			NumServices:   1,
 		},
@@ -304,6 +304,16 @@ message GPUSensor {
  // throttle names what is holding the clocks down now: sw_power_cap,
  // hw_slowdown, sw_thermal, hw_thermal, power_brake.
  repeated string throttle = 6;
+  // How busy the card is, sampled with the rest, rather than read once at
+  // registration. -1 is not reported by the driver. memory_total_mb is set
+  // by every agent that sends these, so zero there is an older agent.
+  double utilization_percent = 7;
+  double memory_utilization_percent = 8;
+  double memory_used_mb = 9;
+  double memory_total_mb = 10;
+  double sm_clock_mhz = 11;
+  // ecc_uncorrected counts uncorrected ECC errors since the driver loaded.
+  double ecc_uncorrected = 12;
 }

 // ShardLoad is one engine process group's use.
@@ -465,6 +475,9 @@ message Provision {
 message EngineWant {
  string engine = 1;  // "llamacpp"
  string version = 2; // "b10991"
+  // variant is the build to install, when one is chosen for the node rather
+  // than left to its hardware: "cu13", "cu129". Empty is the agent's choice.
+  string variant = 3;
 }

 // ProvisionStatus is where one engine's installation stands on a node.
@@ -572,6 +585,9 @@ message EngineProfile {
  map<string, string> env = 9;
  string model_path = 10;
  string served_model_name = 11;
+  // adapters are LoRA adapters loaded beside the model, in order: the index
+  // is the id a request applies one by.
+  repeated LoraAdapter adapters = 12;
 }

 // RemoveWeights deletes cached model files a node no longer needs.
@@ -704,3 +720,10 @@ message UpgradeReply {
  string version = 2;
  string error = 3;
 }
+
+// LoraAdapter is one adapter: the name clients ask for it by, and its source,
+// which the agent fetches as it does the model's weights.
+message LoraAdapter {
+  string name = 1;
+  string source = 2;
+}
@@ -487,6 +487,298 @@
        "defaults": {},
        "overrides": []
      }
+    },
+    {
+      "type": "row",
+      "title": "GPUs",
+      "id": 14,
+      "gridPos": {
+        "x": 0,
+        "y": 30,
+        "w": 24,
+        "h": 1
+      },
+      "collapsed": false,
+      "panels": []
+    },
+    {
+      "type": "timeseries",
+      "title": "GPU busy, by card",
+      "description": "Share of each sample period a kernel was running, sampled every heartbeat.",
+      "datasource": {
+        "type": "prometheus",
+        "uid": "llmgw-prom"
+      },
+      "gridPos": {
+        "x": 0,
+        "y": 31,
+        "w": 12,
+        "h": 7
+      },
+      "id": 15,
+      "targets": [
+        {
+          "expr": "llmgw_gpu_utilization_percent",
+          "legendFormat": "{{node}} gpu {{gpu}}",
+          "refId": "A"
+        }
+      ],
+      "fieldConfig": {
+        "defaults": {
+          "unit": "percent",
+          "min": 0,
+          "max": 100
+        },
+        "overrides": []
+      },
+      "options": {
+        "legend": {
+          "displayMode": "list",
+          "placement": "bottom"
+        },
+        "tooltip": {
+          "mode": "multi"
+        }
+      }
+    },
+    {
+      "type": "timeseries",
+      "title": "GPU memory in use, by card",
+      "description": "Framebuffer in use. Engines reserve most of it at start.",
+      "datasource": {
+        "type": "prometheus",
+        "uid": "llmgw-prom"
+      },
+      "gridPos": {
+        "x": 12,
+        "y": 31,
+        "w": 12,
+        "h": 7
+      },
+      "id": 16,
+      "targets": [
+        {
+          "expr": "llmgw_gpu_memory_used_bytes / llmgw_gpu_memory_total_bytes",
+          "legendFormat": "{{node}} gpu {{gpu}}",
+          "refId": "A"
+        }
+      ],
+      "fieldConfig": {
+        "defaults": {
+          "unit": "percentunit",
+          "min": 0,
+          "max": 1
+        },
+        "overrides": []
+      },
+      "options": {
+        "legend": {
+          "displayMode": "list",
+          "placement": "bottom"
+        },
+        "tooltip": {
+          "mode": "multi"
+        }
+      }
+    },
+    {
+      "type": "timeseries",
+      "title": "GPU memory bandwidth busy, by card",
+      "description": "Decode is bound by memory: high here with compute low is a card waiting on memory.",
+      "datasource": {
+        "type": "prometheus",
+        "uid": "llmgw-prom"
+      },
+      "gridPos": {
+        "x": 0,
+        "y": 38,
+        "w": 12,
+        "h": 7
+      },
+      "id": 17,
+      "targets": [
+        {
+          "expr": "llmgw_gpu_memory_utilization_percent",
+          "legendFormat": "{{node}} gpu {{gpu}}",
+          "refId": "A"
+        }
+      ],
+      "fieldConfig": {
+        "defaults": {
+          "unit": "percent",
+          "min": 0,
+          "max": 100
+        },
+        "overrides": []
+      },
+      "options": {
+        "legend": {
+          "displayMode": "list",
+          "placement": "bottom"
+        },
+        "tooltip": {
+          "mode": "multi"
+        }
+      }
+    },
+    {
+      "type": "timeseries",
+      "title": "GPU temperature",
+      "description": "Throttling usually starts in the high 80s.",
+      "datasource": {
+        "type": "prometheus",
+        "uid": "llmgw-prom"
+      },
+      "gridPos": {
+        "x": 12,
+        "y": 38,
+        "w": 12,
+        "h": 7
+      },
+      "id": 18,
+      "targets": [
+        {
+          "expr": "llmgw_gpu_temperature_celsius",
+          "legendFormat": "{{node}} gpu {{gpu}}",
+          "refId": "A"
+        }
+      ],
+      "fieldConfig": {
+        "defaults": {
+          "unit": "celsius"
+        },
+        "overrides": []
+      },
+      "options": {
+        "legend": {
+          "displayMode": "list",
+          "placement": "bottom"
+        },
+        "tooltip": {
+          "mode": "multi"
+        }
+      }
+    },
+    {
+      "type": "timeseries",
+      "title": "GPU power against its limit",
+      "description": "1 is the enforced limit; a card held there is power-capped.",
+      "datasource": {
+        "type": "prometheus",
+        "uid": "llmgw-prom"
+      },
+      "gridPos": {
+        "x": 0,
+        "y": 45,
+        "w": 12,
+        "h": 7
+      },
+      "id": 19,
+      "targets": [
+        {
+          "expr": "llmgw_gpu_power_watts / llmgw_gpu_power_limit_watts",
+          "legendFormat": "{{node}} gpu {{gpu}}",
+          "refId": "A"
+        }
+      ],
+      "fieldConfig": {
+        "defaults": {
+          "unit": "percentunit",
+          "min": 0,
+          "max": 1.1
+        },
+        "overrides": []
+      },
+      "options": {
+        "legend": {
+          "displayMode": "list",
+          "placement": "bottom"
+        },
+        "tooltip": {
+          "mode": "multi"
+        }
+      }
+    },
+    {
+      "type": "timeseries",
+      "title": "GPU clock",
+      "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"
+        }
+      }
    }
  ]
 }
@@ -200,6 +200,40 @@ groups:
            is skipped by placement rather than failing, so this is quiet until
            somebody wonders why a model will not land there.

+  - name: llmgw-gpus
+    rules:
+      - alert: GPUThermalThrottling
+        expr: llmgw_gpu_throttled{reason=~"sw_thermal|hw_thermal|hw_slowdown"} == 1
+        for: 10m
+        labels: { severity: warning }
+        annotations:
+          summary: "{{ $labels.node }} GPU {{ $labels.gpu }} is slowed by heat ({{ $labels.reason }})"
+          description: >-
+            Ten minutes held back by temperature: every replica on the card is
+            slower than its measurements say. Usually airflow, a fan, or the
+            room — the node page shows the card's temperature over time.
+
+      - alert: GPUUncorrectedECCErrors
+        expr: llmgw_gpu_ecc_errors > 0
+        for: 1m
+        labels: { severity: critical }
+        annotations:
+          summary: "{{ $labels.node }} GPU {{ $labels.gpu }} has {{ $value }} uncorrected ECC errors"
+          description: >-
+            Uncorrected memory errors since the driver loaded: what runs on the
+            card may be computing on corrupted memory. Drain the node
+            (llmgw ctl drain) and have the card looked at.
+
+      - alert: GPUHot
+        expr: llmgw_gpu_temperature_celsius > 87
+        for: 10m
+        labels: { severity: warning }
+        annotations:
+          summary: "{{ $labels.node }} GPU {{ $labels.gpu }} at {{ $value }} °C"
+          description: >-
+            Ten minutes above 87 °C, where most data-centre cards begin to
+            throttle.
+
  - name: monitoring-host
    rules:
      - alert: MonitoringDiskFilling
@@ -367,7 +367,7 @@ Nothing to do: the agent installs them once the node joins, under

 | Engine | On | How |
 |---|---|---|
-| llama.cpp | every node | the release build for its processor, CUDA where there is a GPU, checked against GitHub's checksum |
+| llama.cpp | every node | the release build for its processor and GPU — CUDA 13.3, CUDA 12.8 (x64), or Vulkan for GPUs and drivers the CUDA builds do not cover; the CPU build only where there is no NVIDIA GPU — checked against GitHub's checksum and seen to list the GPU |
 | vLLM | nodes with an NVIDIA GPU it can run on | a Python environment built by uv from PyPI, about 8 GB |
 | SGLang | only when you ask for it | the same way as vLLM |

@@ -376,8 +376,17 @@ node gets the same dependencies whenever it is built, and installed with
 `--require-hashes` against the sha256 PyPI publishes. The resolved list stays in
 the environment as `llmgw-requirements.txt`. It needs:

- **driver 580 or newer and a Turing GPU or newer** (compute capability 7.5):
-  vLLM is built for CUDA 13. A V100 cannot run it;
+- **the build its hardware runs**, chosen by the node. Driver 580 and newer:
+  the CUDA 13 build from PyPI. Driver 525 to 579: the same version built for
+  CUDA 12.9, from vLLM's GitHub release (checked against the digest GitHub
+  publishes) with PyTorch's CUDA 12.9 torch. Both have kernels for Turing
+  (7.5) and newer, 8.0 and newer on aarch64. **A V100** (Volta, 7.0) gets
+  **vLLM 0.19.1**, the newest built for it — no build from 0.20 on has its
+  kernels — and the cluster page shows it running that version; run `llmgw
+  ctl engine-check vllm` there once, since llmgw's flags are checked against
+  the current one. Pascal and older run no vLLM. Whatever is chosen, the
+  install ends by computing something on the GPU; a build that cannot is
+  passed over for the next;
 - **about 25 GB free** under the engines directory while it installs;
 - **gcc**, which Triton builds its kernel launcher with. The agent installs it.

@@ -394,8 +403,51 @@ engines:
 ```

 A changed version reaches every node when it next connects. To use an engine
-you installed yourself instead, set its version to `""` and name its path in
-the node's `agent.yaml` (`engines.vllm.bin`, `engines.sglang.python`).
+you installed yourself on a node, name its path in that node's `agent.yaml`
+(`engines.llamacpp.bin`, `engines.vllm.bin`, `engines.sglang.python`): the agent
+then installs none of its own beside it.
+
+#### Choosing a node's vLLM build
+
+Which build a GPU node installs is chosen on the cluster page, in the **vLLM
+build** column — *auto* (what its hardware runs, the default), *CUDA 13*,
+*CUDA 12.9*, or *none* — or from the command line:
+
+```
+llmgw ctl engine-build gpu07 vllm cu129
+```
+
+The choice is kept by the control plane and the node is told at once; it
+installs the build in the background, and the column shows how that stands.
+
+#### Different GPUs, different engines
+
+A cluster of two GPU generations may need two builds of an engine, or one on
+only some nodes. `engine_overrides` picks nodes — by a GPU class they have, by
+id, or by labels, every one given having to match — and lays engines over the
+cluster's for them, in order:
+
+```yaml
+engines:
+  vllm: "0.30.0"              # CUDA 13: Turing and newer
+engine_overrides:
+  - gpu_class: a10g           # pin a class to a version you have tested
+    engines: { vllm: "0.29.0" }
+  - labels: { role: gguf }    # nodes meant for llama.cpp alone
+    engines: { vllm: "" }     # none
+```
+
+A node told nothing it can run of an engine joins without it: `llmgw doctor`
+and `node add` warn, and the node serves what it does run. To keep an engine
+off a node for good, whatever the control plane says, add the node with
+`llmgw node add --disable-engine vllm`, which writes `engines.disabled: [vllm]`
+into its `agent.yaml` and leaves vLLM out of doctor's checks. That is narrower
+than `--ignore-checks`, which passes over every other problem as well.
+
+A llama.cpp CUDA build is checked for the GPU it is meant for: installing it
+asks it which devices it sees, and a replica placed on GPUs is not ready until
+its process holds memory on each of them. A build missing a CUDA library fails
+there, naming the library, rather than serving on the CPU.

 After changing a version, check it still takes the flags llmgw builds for it,
 which drift between releases:
@@ -671,6 +723,74 @@ llmgw ctl explain-placement qwen2.5-coder-3b-instruct
 which names every candidate and why each was rejected, rather than leaving you
 to infer it.

+### LoRA adapters
+
+A llama.cpp model can carry LoRA adapters, each asked for by its own name as
+though it were a model:
+
+```yaml
+- name: qwen2.5-coder-7b-instruct
+  source: http://mirror:8081/qwen2.5-coder-7b-instruct/5d567d26/model.gguf
+  adapters:
+    - name: coder-sql
+      source: http://mirror:8081/coder-sql/v1/adapter.gguf
+      scale: 1.0          # optional; 1 when left out
+  profiles: [...]
+```
+
+Every replica of the model fetches every adapter and loads it at start, applied
+to nothing. A request for `coder-sql` goes to the model's replicas with the
+adapter applied to it alone; a request for the model gets the model. Adapter
+names share the namespace of models and aliases, appear in `/v1/models`, and
+follow key allowlists on their own: a key allowed `coder-sql` reaches that
+adapter and not the model under it, while a key allowed the model reaches its
+adapters too. Adding or removing an adapter is a change to the model; running
+replicas load the new set when restarted (`llmgw ctl restart --model M`).
+Besides the mirror page's model card, adapters are attached from the command
+line, which reads the file first and refuses one that is not an adapter for
+this model's architecture:
+
+    llmgw ctl model adapters coder --add coder-sql=http://mirror:8081/sql-lora/v1/sql.gguf@0.8
+    llmgw ctl model adapters coder --remove coder-sql
+
+The adapter file must be a GGUF LoRA made for this model — llama.cpp's
+`convert_lora_to_gguf.py` makes one from a PEFT adapter — and not a model with
+the adapter already merged in, which is what most "LoRA GGUF" repositories hold.
+
+### Models served outside the cluster
+
+A model can be an OpenAI-compatible API elsewhere — a hosted provider, or
+another gateway — rather than replicas here. Nothing is placed for it; the
+router sends its requests on, with the provider's key and the provider's name
+for the model:
+
+```yaml
+- name: cloud-mini
+  source: external
+  external:
+    url: https://api.openai.com          # without /v1
+    model: gpt-4o-mini                   # what the provider calls it
+    api_key_file: /etc/llmgw/secrets/openai.key
+```
+
+Named in another model's `fallbacks`, it is used only when the cluster has no
+capacity left for that model:
+
+```yaml
+- name: qwen2.5-coder-7b-instruct
+  fallbacks: [cloud-mini]
+  ...
+```
+
+The key is read from its file on the control plane, never stored in the
+catalog, and a key rotated in the file is used within seconds. A model whose key
+file cannot be read is not offered. Everything else applies as to any model:
+prices and budgets, rate limits, usage records, and key allowlists.
+
+**Its prompts leave the cluster.** Decide who may send them: a key or project
+whose allowlist leaves the external model out cannot reach it, as a fallback or
+by name. The cluster page and `ctl models` mark it `external`.
+
 ### How many replicas, and how much each one carries

 Replica count is changed while the cluster runs, without editing the
@@ -964,7 +1084,39 @@ A replica may come back marked **`ready (stale profile)`**. That means it is
 running a profile that has since been edited — the reconciler leaves a serving
 replica alone on purpose, because churning one costs more than the difference.
 `llmgw ctl restart <replica>` replaces it with one on the new profile, starting
-the new one before stopping the old.
+the new one before stopping the old. Where the old one is all that stands in
+the new one's way — a single GPU — it is stopped first, at once, and the model
+is unserved while the new one loads; where the new profile fits nowhere even
+then, the restart is refused and the old one keeps serving, which `ctl
+replicas` shows as `restart refused:` with the reason. `--in-place` asks for
+stop-then-start whatever room there is.
+
+### Signed releases
+
+Every install above checks a release against its published checksums, which
+proves the download is what the server holds — not that the server holds what
+was built. Signing closes that: CI signs the checksums, and a binary that
+trusts a key installs only releases signed with it. Off until a key exists;
+turning it on, once:
+
+```
+llmgw release keygen
+```
+
+1. Put the **private** key in the repository's Actions secret
+   `RELEASE_SIGNING_KEY`. From the next tag, releases carry
+   `checksums.txt.sig`.
+2. Put the **public** key in `internal/release/trusted.pub` and release. Every
+   binary built from then on carries it and refuses releases that are not
+   signed with it — `ctl upgrade`, agents upgrading themselves, and
+   `llmgw upgrade` alike.
+
+In that order: a build that trusts a key refuses unsigned releases, so the
+secret has to be signing before the key is trusted. The first trusting release
+is itself installed by a binary that trusts nothing; every one after is checked
+by the one before. `LLMGW_RELEASE_KEY` overrides the built-in key, for rotating
+it. The onboarding script fetches with `curl` and checks checksums only; a node
+is checked from its first upgrade on.

 ## Backing up

@@ -1121,9 +1273,10 @@ requests. On a slow model, set `timeouts.generation` higher.
  (step 2) or every request costs zero, and a budget is something the dashboard
  displays rather than something that stops anything. The quotas page says so
  where it can, which is not the same as being told before you rely on it.
- **vLLM needs CUDA 13.** The version the agent installs needs driver 580
-  and a Turing GPU or newer. On older hardware, set `engines.vllm` to `""` and
-  install a vLLM built for your CUDA yourself.
+- **A V100 runs vLLM 0.19.1**, not the version the rest of the cluster runs:
+  no build from 0.20 on has Volta kernels, and the node picks the newest that
+  has. llmgw's flags are checked against the current version, so run `llmgw
+  ctl engine-check vllm` on a V100 node once.
 - **Two control planes have not been run for long.** The leader election and
  agent failover work and there is a pattern in `deploy/ha/`, but every cluster
  this has been run on in anger has had one. A single control plane is the
@@ -15,12 +15,16 @@ rather than repeated.

 | Key | Type | In |
 |---|---|---|
+| `name` | `string` | [Adapter](#adapter) |
+| `source` | `string` | [Adapter](#adapter) |
+| `scale` | `float64` | [Adapter](#adapter) |
 | `enabled` | `bool` | [Admission](#admission) |
 | `bands` | `[]Band` | [Admission](#admission) |
 | `default_band` | `string` | [Admission](#admission) |
 | `band_header` | `string` | [Admission](#admission) |
 | `queue_ttl` | `time.Duration` | [Admission](#admission) |
 | `queue_capacity_per_band` | `int` | [Admission](#admission) |
+| `queue_capacity_total` | `int` | [Admission](#admission) |
 | `interval` | `time.Duration` | [Admission](#admission) |
 | `enabled` | `bool` | [Auth](#auth) |
 | `store` | `string` | [Auth](#auth) |
@@ -77,6 +81,13 @@ rather than repeated.
 | `pool` | `string` | [Endpoint](#endpoint) |
 | `weight` | `float64` | [Endpoint](#endpoint) |
 | `priority` | `int` | [Endpoint](#endpoint) |
+| `gpu_class` | `string` | [EngineOverride](#engineoverride) |
+| `nodes` | `[]string` | [EngineOverride](#engineoverride) |
+| `labels` | `map[string]string` | [EngineOverride](#engineoverride) |
+| `engines` | `map[string]string` | [EngineOverride](#engineoverride) |
+| `url` | `string` | [External](#external) |
+| `model` | `string` | [External](#external) |
+| `api_key_file` | `string` | [External](#external) |
 | `contact` | `string` | [Landing](#landing) |
 | `notice` | `string` | [Landing](#landing) |
 | `bin` | `string` | [LlamaCppConfig](#llamacppconfig) |
@@ -94,6 +105,8 @@ rather than repeated.
 | `scaling` | `Scaling` | [Model](#model) |
 | `class` | `string` | [Model](#model) |
 | `owner` | `string` | [Model](#model) |
+| `external` | `*External` | [Model](#model) |
+| `adapters` | `[]Adapter` | [Model](#model) |
 | `fallbacks` | `[]string` | [Model](#model) |
 | `routing_profile` | `string` | [Model](#model) |
 | `slo` | `SLO` | [Model](#model) |
@@ -174,6 +187,7 @@ rather than repeated.
 | `tracing` | `Tracing` | [Server](#server) |
 | `mirror` | `Mirror` | [Server](#server) |
 | `engines` | `map[string]string` | [Server](#server) |
+| `engine_overrides` | `[]EngineOverride` | [Server](#server) |
 | `backup_dir` | `string` | [Server](#server) |
 | `releases` | `string` | [Server](#server) |
 | `router` | `Router` | [Server](#server) |
@@ -195,6 +209,27 @@ rather than repeated.
 | `launch_mode` | `string` | [VLLMConfig](#vllmconfig) |
 | `ray_bin` | `string` | [VLLMConfig](#vllmconfig) |

+## Adapter
+
+### `name`
+
+`string`
+
+Name is what clients ask for, like a model name.
+
+### `source`
+
+`string`
+
+Source is the adapter file (GGUF for llama.cpp), fetched by each node
+as the model's weights are.
+
+### `scale`
+
+`float64`
+
+Scale is how strongly it is applied; zero takes 1.
+
 ## Admission

 ### `enabled`
@@ -228,6 +263,15 @@ should not be able to promote themselves.

 `int`

+### `queue_capacity_total`
+
+`int`
+
+QueueCapacityTotal bounds everything queued across bands. Reached, a
+request is not refused while a lower band has any queued: the newest of
+the lowest band is turned away instead, with a 503 and Retry-After.
+Zero takes half of all the bands' capacities together.
+
 ### `interval`

 `time.Duration`
@@ -606,6 +650,54 @@ called. Empty means vLLM's names.

 `int`

+## EngineOverride
+
+### `gpu_class`
+
+`string`
+
+GPUClass picks nodes with a GPU of this class ("v100-32g").
+
+### `nodes`
+
+`[]string`
+
+Nodes picks nodes by id.
+
+### `labels`
+
+`map[string]string`
+
+Labels picks nodes carrying all of these.
+
+### `engines`
+
+`map[string]string`
+
+Engines are laid over the cluster's: a version, or "" for none.
+
+## External
+
+### `url`
+
+`string`
+
+URL is the provider's base address, without /v1: https://api.openai.com.
+
+### `model`
+
+`string`
+
+Model is what the provider calls it; requests are rewritten to it.
+
+### `api_key_file`
+
+`string`
+
+APIKeyFile holds the provider's key, on the control plane. A file
+rather than the key itself, so the key is in neither the catalog, the
+store nor the audit log; empty for a provider that asks for none.
+
 ## Landing

 ### `contact`
@@ -755,6 +847,28 @@ Owner is the part of the tenancy tree whose capacity quota this model
 counts against, written org:ID, team:ID or project:ID. Empty is shared:
 held by the cluster, counted against nobody.

+### `external`
+
+`*External`
+
+External is a model served by someone else — an OpenAI-compatible API
+outside the cluster — rather than by replicas placed here. Such a model
+has no profiles and nothing is placed for it; the router sends its
+requests to the provider. Most useful named in another model's
+fallbacks: used only when the cluster has no capacity left. Its
+prompts leave the cluster, so which keys may name it is worth an
+allowlist.
+
+Takes a [External](#external) block.
+
+### `adapters`
+
+`[]Adapter`
+
+Adapters are LoRA adapters served on this model's replicas, each asked
+for by its own name as though it were a model. Every replica loads
+every adapter, applying one only to a request that names it.
+
 ### `fallbacks`

 `[]string`
@@ -1329,6 +1443,16 @@ never trusted to already have what somebody once put there.
 Unset keeps DefaultEngines. A version replaces the default; an empty
 one leaves that engine to be installed by hand.

+### `engine_overrides`
+
+`[]EngineOverride`
+
+EngineOverrides give some nodes different engines from the rest: a
+vLLM built for CUDA 12 on the Volta cards CUDA 13 dropped, or none at
+all on nodes meant only for llama.cpp. Each picks nodes and lays its
+engines over the cluster's, in order, later ones last; an empty version
+removes that engine from the nodes it picks.
+
 ### `backup_dir`

 `string`
@@ -314,6 +314,17 @@ func (a *Agent) settleAdopted(sh *supervisor.Shard, o supervisor.Orphan) {
 			"shard", o.ShardID, "replica", o.ReplicaID, "err", err)
 		return
 	}
+	// An engine started by an agent older than the GPU check may be one
+	// serving a GPU placement on the CPU. Stopped, so its replacement starts
+	// through the check and fails there in words, if the build still sees
+	// no GPU — rather than going on unnoticed after the upgrade that could
+	// have caught it.
+	if err := a.onItsGPUs(ctx, eng, o.Spec, sh.Status().PID); err != nil {
+		a.log.Error("stopping an adopted engine that computes on the CPU though placed on GPUs",
+			"shard", o.ShardID, "replica", o.ReplicaID, "err", err)
+		_ = a.sup.Stop(ctx, o.ShardID)
+		return
+	}
 	a.log.Info("adopted engine is serving",
 		"shard", o.ShardID, "replica", o.ReplicaID)
 }
@@ -450,7 +461,7 @@ func (a *Agent) session(ctx context.Context, server string) error {
 	// vLLM accepted a vLLM shard and failed it minutes later.
 	reg.EngineVersions = a.engines(parent)
 	reg.EnginesProbed = true
-	reg.InstallableEngines = provision.Installable()
+	reg.InstallableEngines = slices.DeleteFunc(provision.Installable(), a.disabled)
 	reg.Features = []string{FeatureFabricTest}
 	reg.RunningShards = a.runningShards()
 	a.mu.Lock()
@@ -649,6 +660,9 @@ func (a *Agent) sampleLoad() *llmgwv1.NodeLoad {
 			out.Gpus = append(out.Gpus, &llmgwv1.GPUSensor{
 				Index: int32(g.Index), TempC: g.TempC, MemoryTempC: g.MemoryTempC,
 				PowerW: g.PowerW, PowerLimitW: g.PowerLimitW, Throttle: g.Throttle,
+				UtilizationPercent: g.UtilPercent, MemoryUtilizationPercent: g.MemUtilPercent,
+				MemoryUsedMb: g.MemoryUsedMB, MemoryTotalMb: g.MemoryTotalMB,
+				SmClockMhz: g.SMClockMHz, EccUncorrected: g.ECCUncorrected,
 			})
 		}
 	}
@@ -891,6 +905,17 @@ func (a *Agent) startShard(ctx context.Context, m *llmgwv1.StartShard, out chan<
 		stop()
 		spec.Profile.ModelPath = path
 	}
+	// Adapters are fetched like the model, and before the engine starts: it
+	// loads them at start, and one missing is a replica that answers for the
+	// base model under the adapter's name.
+	for i, ad := range spec.Profile.Adapters {
+		path, err := a.weights().Ensure(ctx, ad.Path)
+		if err != nil {
+			report(supervisor.Failed, fmt.Sprintf("adapter %s: %v", ad.Name, err), nil)
+			return
+		}
+		spec.Profile.Adapters[i].Path = path
+	}

 	cmds, err := eng.Commands(spec)
 	if err != nil {
@@ -920,6 +945,14 @@ func (a *Agent) startShard(ctx context.Context, m *llmgwv1.StartShard, out chan<
 	} else {
 		err = a.sup.WaitAlive(ctx, sh)
 	}
+	if err == nil {
+		err = a.onItsGPUs(ctx, eng, spec, sh.Status().PID)
+		if err != nil {
+			// Serving, but not where it was placed: stopped, so the GPU it
+			// holds in the inventory is not taken by a CPU process.
+			_ = a.sup.Stop(ctx, shardID)
+		}
+	}
 	if err != nil {
 		report(supervisor.Failed, err.Error(), sh.Logs(20))
 		return
@@ -0,0 +1,32 @@
+package agent
+
+import (
+	"testing"
+
+	llmgwv1 "gitea.mynodes.xyz/ilker/llmgw/api/gen/llmgw/v1"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/engines"
+)
+
+func TestAnEngineDisabledOnTheNodeIsNeitherInstalledNorAdvertised(t *testing.T) {
+	a := &Agent{}
+	a.cfg.Engines.Disabled = []string{"vllm"}
+	a.setProvision(&llmgwv1.Provision{Engines: []*llmgwv1.EngineWant{
+		{Engine: "llamacpp", Version: "b1"}, {Engine: "vllm", Version: "0.30.0"},
+	}})
+	if len(a.provWants) != 1 || a.provWants[0].Engine != "llamacpp" {
+		t.Errorf("wants %+v: vLLM is disabled here and must not be installed", a.provWants)
+	}
+	a.configuredVers = map[string]string{"vllm": "0.25.0"}
+	a.installedVers = map[string]string{"llamacpp": "b1"}
+	if got := a.engineReportLocked(); got["vllm"] != "" || got["llamacpp"] != "b1" {
+		t.Errorf("advertised %v: a disabled engine found on disk is still not offered", got)
+	}
+}
+
+func TestAnEngineNamedByHandIsNotInstalledOverIt(t *testing.T) {
+	a := &Agent{reg: engines.NewRegistry(engines.Config{LlamaCpp: engines.LlamaCppConfig{Bin: "/opt/own/llama-server"}})}
+	a.setProvision(&llmgwv1.Provision{Engines: []*llmgwv1.EngineWant{{Engine: "llamacpp", Version: "b1"}}})
+	if len(a.provWants) != 0 {
+		t.Errorf("wants %+v: an installed build would be used over the operator's own", a.provWants)
+	}
+}
@@ -0,0 +1,64 @@
+package discovery
+
+import (
+	"context"
+	"encoding/xml"
+	"fmt"
+	"os/exec"
+	"strconv"
+	"strings"
+)
+
+// GPUProcesses reports which processes hold memory on each GPU, by the GPU's
+// nvidia-smi index: what a process actually computes on, as opposed to what it
+// was told to use. A llama.cpp started with --n-gpu-layers on a GPU whose
+// backend did not load runs on the CPU without a word, and this is where that
+// shows — the process is on no GPU's list.
+//
+// From the full query rather than --query-compute-apps, which lists compute
+// processes only: llama.cpp's Vulkan build is listed as a graphics one, and
+// would have read as not on its GPU at all.
+//
+// A node with no nvidia-smi returns nil and no error, as discovery does.
+func GPUProcesses(ctx context.Context, binary string) (map[int][]int, error) {
+	if binary == "" {
+		if _, err := exec.LookPath("nvidia-smi"); err != nil {
+			return nil, nil
+		}
+		binary = "nvidia-smi"
+	}
+	out, err := exec.CommandContext(ctx, binary, "-q", "-x").Output()
+	if err != nil {
+		return nil, fmt.Errorf("running %s: %w", binary, err)
+	}
+	return parseGPUProcesses(out)
+}
+
+// smiLog is the part of nvidia-smi's XML report this reads: each GPU, in
+// index order, and the processes on it.
+type smiLog struct {
+	GPUs []struct {
+		Processes []struct {
+			PID  string `xml:"pid"`
+			Type string `xml:"type"`
+		} `xml:"processes>process_info"`
+	} `xml:"gpu"`
+}
+
+// parseGPUProcesses reads the report. GPUs are listed in index order, which
+// is nvidia-smi's own numbering — the one placement assigns by.
+func parseGPUProcesses(report []byte) (map[int][]int, error) {
+	var log smiLog
+	if err := xml.Unmarshal(report, &log); err != nil {
+		return nil, fmt.Errorf("reading nvidia-smi's report: %w", err)
+	}
+	out := map[int][]int{}
+	for i, g := range log.GPUs {
+		for _, p := range g.Processes {
+			if pid, err := strconv.Atoi(strings.TrimSpace(p.PID)); err == nil {
+				out[i] = append(out[i], pid)
+			}
+		}
+	}
+	return out, nil
+}
@@ -0,0 +1,38 @@
+package discovery
+
+import (
+	"reflect"
+	"testing"
+)
+
+func TestGPUProcessesOfEveryKindAreListedByIndex(t *testing.T) {
+	report := `<?xml version="1.0" ?>
+<nvidia_smi_log>
+  <driver_version>580.65.06</driver_version>
+  <gpu id="00000000:17:00.0"><product_name>Tesla V100-SXM2-32GB</product_name>
+    <processes></processes>
+  </gpu>
+  <gpu id="00000000:65:00.0"><product_name>Tesla V100-SXM2-32GB</product_name>
+    <processes>
+      <process_info><pid>4100</pid><type>C</type><process_name>llama-server</process_name></process_info>
+      <process_info><pid>4300</pid><type>G</type><process_name>llama-server</process_name></process_info>
+    </processes>
+  </gpu>
+  <gpu id="00000000:B3:00.0"><processes>
+      <process_info><pid>4200</pid><type>C+G</type></process_info>
+  </processes></gpu>
+</nvidia_smi_log>`
+	got, err := parseGPUProcesses([]byte(report))
+	if err != nil {
+		t.Fatal(err)
+	}
+	// A Vulkan engine is a graphics process ("G") and is on its GPU all the
+	// same; a compute-only query would have missed it.
+	want := map[int][]int{1: {4100, 4300}, 2: {4200}}
+	if !reflect.DeepEqual(got, want) {
+		t.Errorf("got %v, want %v", got, want)
+	}
+	if _, err := parseGPUProcesses([]byte("not xml")); err == nil {
+		t.Error("a report that does not parse was read as no processes")
+	}
+}
@@ -41,6 +41,14 @@ type GPUSensor struct {
 	// is running free. Idle and application clock settings are left out;
 	// they are not the card being held back.
 	Throttle []string
+	// How busy it is, read with the rest every sample. Discovery reads the
+	// same once, at start, which made a busy card look as idle as it was
+	// when the agent came up. -1 is not reported by the driver; zero
+	// MemoryTotalMB is a sample without these at all.
+	UtilPercent, MemUtilPercent float64
+	MemoryUsedMB, MemoryTotalMB float64
+	SMClockMHz                  float64
+	ECCUncorrected              float64
 }

 // sensorEvery bounds how often nvidia-smi is run. Heartbeats come more often
@@ -187,13 +195,33 @@ var throttleBits = []struct {
 // gpus asks nvidia-smi for each card's heat and power. The memory temperature
 // is not known to every driver, and a field nvidia-smi does not know fails the
 // whole query, so it is asked for first and dropped on failure.
+// gpuFields are what each sample asks nvidia-smi for, in order. The busy
+// figures follow the throttle reasons; the memory temperature is last, being
+// the one field a driver may reject outright.
+const (
+	gpuBase  = "index,temperature.gpu,power.draw,enforced.power.limit,clocks_throttle_reasons.active"
+	gpuBusy  = ",utilization.gpu,utilization.memory,memory.used,memory.total,clocks.sm,ecc.errors.uncorrected.volatile.total"
+	gpuHBMTC = ",temperature.memory"
+)
+
 func (s *SensorSampler) gpus(ctx context.Context) []GPUSensor {
-	withMem := "index,temperature.gpu,power.draw,enforced.power.limit,clocks_throttle_reasons.active,temperature.memory"
-	out, err := s.smi(ctx, withMem)
-	if err != nil {
-		if out, err = s.smi(ctx, strings.TrimSuffix(withMem, ",temperature.memory")); err != nil {
-			return nil
+	// Most fields first, fewer on a driver that refuses one: a query naming
+	// a field the driver does not know fails as a whole.
+	var out []byte
+	var busy bool
+	for _, q := range []struct {
+		fields string
+		busy   bool
+	}{{gpuBase + gpuBusy + gpuHBMTC, true}, {gpuBase + gpuBusy, true}, {gpuBase + gpuHBMTC, false}, {gpuBase, false}} {
+		var err error
+		if out, err = s.smi(ctx, q.fields); err == nil {
+			busy = q.busy
+			break
 		}
+		out = nil
+	}
+	if out == nil {
+		return nil
 	}
 	var gs []GPUSensor
 	for _, line := range strings.Split(strings.TrimSpace(string(out)), "\n") {
@@ -216,14 +244,31 @@ func (s *SensorSampler) gpus(ctx context.Context) []GPUSensor {
 				}
 			}
 		}
-		if len(f) > 5 {
-			g.MemoryTempC = num(f[5])
+		rest := f[5:]
+		if busy && len(rest) >= 6 {
+			g.UtilPercent, g.MemUtilPercent = numOrUnknown(rest[0]), numOrUnknown(rest[1])
+			g.MemoryUsedMB, g.MemoryTotalMB = num(rest[2]), num(rest[3])
+			g.SMClockMHz, g.ECCUncorrected = num(rest[4]), numOrUnknown(rest[5])
+			rest = rest[6:]
+		}
+		if len(rest) > 0 {
+			g.MemoryTempC = num(rest[0])
 		}
 		gs = append(gs, g)
 	}
 	return gs
 }

+// numOrUnknown is a figure, or -1 where nvidia-smi says it has none: a card
+// without ECC has no error count, which is not the same as none.
+func numOrUnknown(s string) float64 {
+	v, err := strconv.ParseFloat(s, 64)
+	if err != nil || v < 0 {
+		return -1
+	}
+	return v
+}
+
 // num reads a figure nvidia-smi may give as "[N/A]" or "N/A".
 func num(s string) float64 {
 	v, err := strconv.ParseFloat(s, 64)
@@ -167,3 +167,44 @@ func TestADriverThatDoesNotKnowMemoryTemperatureStillReportsTheRest(t *testing.T
 		t.Errorf("%+v", got.GPUs)
 	}
 }
+
+func TestEachSampleSaysHowBusyEachGPUIs(t *testing.T) {
+	// What the card is doing now, every sample: discovery read it once, at
+	// start, and a card busy since looked as idle as it was then.
+	smi := func(_ context.Context, fields string) ([]byte, error) {
+		if !strings.Contains(fields, "utilization.gpu") {
+			return nil, errors.New("asked without how busy the cards are")
+		}
+		return []byte(
+			"0, 71, 612.45, 700.00, 0x0000000000000000, 97, 64, 71234, 81559, 1980, 0, 84\n" +
+				"1, 30, 70.10, 700.00, 0x0000000000000001, 0, 0, 512, 81559, 345, [N/A], [N/A]\n"), nil
+	}
+	now := time.Now()
+	got := sampler(t.TempDir(), smi, &now).Sample(t.Context()).GPUs
+	if len(got) != 2 {
+		t.Fatalf("%+v", got)
+	}
+	g0, g1 := got[0], got[1]
+	if g0.UtilPercent != 97 || g0.MemUtilPercent != 64 || g0.MemoryUsedMB != 71234 || g0.MemoryTotalMB != 81559 ||
+		g0.SMClockMHz != 1980 || g0.ECCUncorrected != 0 || g0.MemoryTempC != 84 {
+		t.Errorf("card 0 %+v", g0)
+	}
+	// Idle is zero; a count the card does not keep is unknown, not zero.
+	if g1.UtilPercent != 0 || g1.ECCUncorrected != -1 || g1.MemoryTempC != 0 {
+		t.Errorf("card 1 %+v", g1)
+	}
+}
+
+func TestADriverThatDoesNotKnowTheBusyFieldsStillReportsHeatAndPower(t *testing.T) {
+	smi := func(_ context.Context, fields string) ([]byte, error) {
+		if strings.Contains(fields, "utilization.gpu") {
+			return nil, errors.New(`Field "ecc.errors.uncorrected.volatile.total" is not a valid field to query.`)
+		}
+		return []byte("0, 66, 250.0, 300.0, 0x0000000000000000, 70\n"), nil
+	}
+	now := time.Now()
+	got := sampler(t.TempDir(), smi, &now).Sample(t.Context()).GPUs
+	if len(got) != 1 || got[0].TempC != 66 || got[0].MemoryTempC != 70 || got[0].MemoryTotalMB != 0 {
+		t.Errorf("%+v", got)
+	}
+}
@@ -9,6 +9,7 @@
 package engines

 import (
+	"context"
 	"fmt"
 	"os"
 	"slices"
@@ -40,6 +41,9 @@ type Profile struct {
 	ModelPath string
 	// ServedModelName is the name clients use, which need not match the path.
 	ServedModelName string
+	// Adapters are LoRA adapters to load beside the model, in order: an
+	// adapter's index is the id requests apply it by.
+	Adapters []Adapter
 }

 // ShardSpec is one shard of one replica on this node.
@@ -143,6 +147,15 @@ type Engine interface {
 	MetricsPath() string
 }

+// CPUFallback is an engine that can come up on the CPU when it was placed on
+// GPUs, and keep serving there without an error. The agent checks such a shard
+// is on its GPUs before calling it ready.
+type CPUFallback interface {
+	FallsBackToCPU() bool
+	// Devices lists the GPUs the engine's build can see, for saying why.
+	Devices(ctx context.Context) ([]string, error)
+}
+
 // GPUInfo is the subset of GPU facts validation needs.
 type GPUInfo struct {
 	Index             int
@@ -200,6 +213,10 @@ func (r *Registry) Names() []string {

 // Config configures the engine adapters.
 type Config struct {
+	// Disabled are engines this node never installs or advertises, whatever
+	// the control plane asks: vLLM on a node of cards no build of it runs on,
+	// that is meant for llama.cpp alone.
+	Disabled []string       `koanf:"disabled"`
 	VLLM     VLLMConfig     `koanf:"vllm"`
 	SGLang   SGLangConfig   `koanf:"sglang"`
 	LlamaCpp LlamaCppConfig `koanf:"llamacpp"`
@@ -229,6 +246,11 @@ func NCCLEnv(s ShardSpec) map[string]string { return baseEnv(s) }
 func baseEnv(s ShardSpec) map[string]string {
 	env := map[string]string{
 		"CUDA_VISIBLE_DEVICES": joinInts(s.GPUIndices),
+		// The indices above are nvidia-smi's, which counts by PCI bus. CUDA
+		// counts fastest first unless told otherwise, and on a node with two
+		// kinds of GPU the shard would get a different card than placement
+		// gave it — one another shard may hold.
+		"CUDA_DEVICE_ORDER": "PCI_BUS_ID",
 	}

 	if len(s.HCAs) > 0 {
@@ -373,3 +395,10 @@ func quantizationOf(args []string) string {
 	}
 	return ""
 }
+
+// Adapter is a LoRA adapter an engine loads beside its model. Path is a
+// source until the agent has fetched it, a local file after.
+type Adapter struct {
+	Name string `json:"name"`
+	Path string `json:"path"`
+}
@@ -1,8 +1,10 @@
 package engines

 import (
+	"context"
 	"fmt"
 	"os"
+	"os/exec"
 	"path/filepath"
 	"runtime"
 	"sort"
@@ -65,25 +67,25 @@ func (*LlamaCpp) ServesHTTP(ShardSpec) bool { return true }

 // Validate checks the profile against the GPUs it would run on.
 //
-// The two refusals here are about not pretending: llama.cpp has no equivalent
-// of a tensor-parallel group spanning machines, and its multi-GPU support is a
-// different mechanism (--split-mode, --tensor-split) that this adapter does not
-// drive. Accepting either would produce a replica that starts and then serves
-// from one machine while the catalog believes it has several.
+// llama.cpp spreads a model over a node's GPUs by layer (--split-mode layer),
+// so gpus_per_node is how many it gets. It has no tensor parallelism and no
+// group spanning machines, and accepting tp or nodes would produce a replica
+// that serves from fewer cards, or one machine, while the catalog believes
+// otherwise.
 func (*LlamaCpp) Validate(p Profile, gpus []GPUInfo) error {
-	if err := validateCommon(p, gpus); err != nil {
-		return err
-	}
 	if p.Nodes > 1 {
 		return fmt.Errorf("profile %q: llamacpp does not span nodes; use vllm or sglang for a %d-node replica",
 			p.Name, p.Nodes)
 	}
-	if p.TP > 1 {
-		return fmt.Errorf("profile %q: llamacpp does not do tensor parallelism; "+
-			"it splits a model across local GPUs with --tensor-split, which this adapter does not drive",
-			p.Name)
+	if p.TP > 1 || p.PP > 1 {
+		return fmt.Errorf("profile %q: llamacpp splits a model across a node's GPUs by layer, not by tp or pp; "+
+			"write gpus_per_node: %d instead", p.Name, max(p.TP, 1)*max(p.PP, 1))
 	}
-	return nil
+	// The common checks count a replica's GPUs as tp; for llama.cpp the
+	// count is gpus_per_node.
+	q := p
+	q.TP = max(p.GPUsPerNode, 1)
+	return validateCommon(q, gpus)
 }

 // Commands returns the processes for one shard, which is always one process.
@@ -132,6 +134,13 @@ func (l *LlamaCpp) Commands(s ShardSpec) ([]Command, error) {
 	if len(s.GPUIndices) > 0 && !hasFlag(p.Args, "--n-gpu-layers", "-ngl") {
 		args = append(args, "--n-gpu-layers", "999")
 	}
+	// Several GPUs: the layers spread across all of them, which is the only
+	// split llama.cpp does without a matched interconnect. In proportion to
+	// each card's free memory, which for the whole cards placement gives is
+	// an even split.
+	if len(s.GPUIndices) > 1 && !hasFlag(p.Args, "--split-mode", "-sm") {
+		args = append(args, "--split-mode", "layer")
+	}
 	if len(s.GPUIndices) == 0 && !hasFlag(p.Args, "--threads", "-t") {
 		// What placement reserved first: it is what the node's other
 		// replicas were placed around.
@@ -144,12 +153,29 @@ func (l *LlamaCpp) Commands(s ShardSpec) ([]Command, error) {
 		}
 		args = append(args, "--threads", strconv.Itoa(n))
 	}
+	// Loaded at start, applied per request: a request that names an adapter
+	// carries "lora": [{"id": i, "scale": s}], and one that names none is the
+	// base model. Each is loaded at scale zero, not with --lora and
+	// --lora-init-without-apply: that leaves an adapter's scale at 1, and
+	// llama-server applies it at that scale to every request with no "lora"
+	// of its own — every request for the model itself.
+	for _, a := range p.Adapters {
+		args = append(args, "--lora-scaled", a.Path+":0")
+	}
 	args = append(args, p.Args...)

 	env := baseEnv(s)
 	if s.APIKey != "" {
 		env["LLAMA_API_KEY"] = s.APIKey
 	}
+	// The Vulkan build — what a GPU older than the CUDA builds gets — does
+	// not read CUDA_VISIBLE_DEVICES; this is its own. Vulkan numbers devices
+	// as the driver lists them, which on NVIDIA follows the same bus order
+	// as nvidia-smi; were it not to, the shard would not hold memory on its
+	// own GPUs, and starting it checks exactly that.
+	if len(s.GPUIndices) > 0 {
+		env["GGML_VK_VISIBLE_DEVICES"] = joinInts(s.GPUIndices)
+	}
 	return []Command{{
 		Name: "llama-server", Path: bin, Args: args, Env: materializeEnv(env),
 	}}, nil
@@ -225,3 +251,38 @@ func (l *LlamaCpp) bin() string {
 	}
 	return "llama-server"
 }
+
+// FallsBackToCPU says that llama.cpp keeps running when its GPU backend does
+// not load: --n-gpu-layers asks for layers on a GPU it then has none of, and
+// the server comes up and answers on the CPU, at a tenth of the speed, with no
+// error anywhere. Its readiness on a GPU placement needs checking separately.
+func (*LlamaCpp) FallsBackToCPU() bool { return true }
+
+// Devices lists the GPUs this build can compute on, as llama-server's
+// --list-devices names them ("CUDA0: NVIDIA L4 (22994 MiB, 22713 MiB free)").
+// Empty, on a machine with GPUs, is a build whose GPU backend did not load:
+// the CPU build, or a CUDA build missing a library its backend links.
+func (l *LlamaCpp) Devices(ctx context.Context) ([]string, error) {
+	out, err := exec.CommandContext(ctx, l.bin(), "--list-devices").CombinedOutput() //nolint:gosec // the engine's own binary
+	if err != nil {
+		return nil, fmt.Errorf("%s --list-devices: %v: %s", l.bin(), err, strings.TrimSpace(string(out)))
+	}
+	return parseDevices(string(out)), nil
+}
+
+// parseDevices reads --list-devices: the lines after "Available devices:",
+// which says "(none)" when there are none. Log lines before it are skipped.
+func parseDevices(out string) []string {
+	var devs []string
+	listing := false
+	for _, line := range strings.Split(out, "\n") {
+		line = strings.TrimSpace(line)
+		switch {
+		case strings.HasPrefix(line, "Available devices"):
+			listing = true
+		case listing && line != "" && line != "(none)":
+			devs = append(devs, line)
+		}
+	}
+	return devs
+}
@@ -306,3 +306,58 @@ func TestTheCoresPlacementReservedAreTheThreads(t *testing.T) {
 		t.Errorf("--threads %q; what placement reserved should win over the node's setting", v)
 	}
 }
+
+func TestAdaptersAreLoadedAtStartAndAppliedToNoneByDefault(t *testing.T) {
+	s := llamaSpec(nil)
+	s.Profile.Adapters = []Adapter{{Name: "sql", Path: "/w/sql.gguf"}, {Name: "docs", Path: "/w/docs.gguf"}}
+	cmds, err := (&LlamaCpp{}).Commands(s)
+	if err != nil {
+		t.Fatal(err)
+	}
+	args := strings.Join(cmds[0].Args, " ")
+	if !strings.Contains(args, "--lora-scaled /w/sql.gguf:0 --lora-scaled /w/docs.gguf:0") {
+		t.Errorf("args %s: both adapters, in order, loaded at scale zero", args)
+	}
+	plain, _ := (&LlamaCpp{}).Commands(llamaSpec(nil))
+	if strings.Contains(strings.Join(plain[0].Args, " "), "--lora") {
+		t.Error("a model with no adapters was given lora flags")
+	}
+}
+
+func TestDevicesAreReadFromListDevices(t *testing.T) {
+	cuda := "ggml_cuda_init: found 2 CUDA devices:\nAvailable devices:\n  CUDA0: NVIDIA L4 (22994 MiB, 22713 MiB free)\n  CUDA1: NVIDIA L4 (22994 MiB, 22713 MiB free)\n"
+	if got := parseDevices(cuda); len(got) != 2 || !strings.HasPrefix(got[0], "CUDA0: NVIDIA L4") {
+		t.Errorf("CUDA build: %q", got)
+	}
+	// A CUDA build whose backend did not load says what the CPU build says.
+	if got := parseDevices("Available devices:\n  (none)\n"); len(got) != 0 {
+		t.Errorf("no devices: %q", got)
+	}
+}
+
+func TestLlamaCppSpreadsAReplicaOverItsNodesGPUsByLayer(t *testing.T) {
+	four := Profile{Name: "t4x4", Engine: "llamacpp", GPUClass: "t4", TP: 1, GPUsPerNode: 4, ModelPath: "/m"}
+	if err := (&LlamaCpp{}).Validate(four, h100(4)); err != nil {
+		t.Fatalf("four GPUs on a four-GPU node: %v", err)
+	}
+	if err := (&LlamaCpp{}).Validate(four, h100(2)); err == nil {
+		t.Error("four GPUs were accepted on a node with two")
+	}
+
+	s := llamaSpec([]int{0, 1, 2, 3})
+	s.Profile.GPUsPerNode = 4
+	cmds, err := (&LlamaCpp{}).Commands(s)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if v, _ := flagValue(cmds[0].Args, "--split-mode"); v != "layer" {
+		t.Errorf("--split-mode %q over four GPUs", v)
+	}
+	if !slices.Contains(cmds[0].Env, "CUDA_VISIBLE_DEVICES=0,1,2,3") {
+		t.Errorf("env %v: the replica must see all four of its GPUs", cmds[0].Env)
+	}
+	one, _ := (&LlamaCpp{}).Commands(llamaSpec([]int{2}))
+	if hasFlag(one[0].Args, "--split-mode") {
+		t.Error("a one-GPU replica was given a split mode")
+	}
+}
@@ -0,0 +1,116 @@
+package agent
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"strconv"
+	"strings"
+	"time"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/discovery"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/engines"
+)
+
+// gpuProcesses and processGroup are variables so tests can stand in for
+// nvidia-smi and /proc.
+var (
+	gpuProcesses = discovery.GPUProcesses
+	processGroup = pgidOf
+	// gpuCheckWait is how long a shard has, once it answers, to show up on
+	// its GPUs: the weights are on them by then, but nvidia-smi can lag.
+	gpuCheckWait = 15 * time.Second
+)
+
+// onItsGPUs checks that a shard placed on GPUs computes on them.
+//
+// For an engine that falls back to the CPU when its GPU backend does not load,
+// "it answers" is not "it is ready": llama.cpp built for CUDA but missing
+// cuBLAS comes up on the CPU, answers at a tenth of the speed, and the GPU it
+// was given sits idle and reserved. What it computes on is what nvidia-smi
+// lists it on, so a shard none of whose processes hold memory on one of its
+// GPUs is refused, with what the engine itself can see.
+func (a *Agent) onItsGPUs(ctx context.Context, eng engines.Engine, spec engines.ShardSpec, pid int) error {
+	fb, ok := eng.(engines.CPUFallback)
+	if !ok || !fb.FallsBackToCPU() || len(spec.GPUIndices) == 0 || pid <= 0 {
+		return nil
+	}
+	group := processGroup(pid)
+	deadline := time.Now().Add(gpuCheckWait)
+	var missing []int
+	for {
+		procs, err := gpuProcesses(ctx, a.cfg.Discovery.NvidiaSMI)
+		if err != nil || procs == nil {
+			// No nvidia-smi to ask is no evidence either way, and a node that
+			// was given GPUs has one; say so and let the shard serve.
+			a.log.Warn("could not check that a GPU shard runs on its GPUs", "shard", spec.ShardID, "err", err)
+			return nil
+		}
+		missing = missing[:0]
+		for _, i := range spec.GPUIndices {
+			if !holds(procs[i], pid, group) {
+				missing = append(missing, i)
+			}
+		}
+		if len(missing) == 0 {
+			return nil
+		}
+		if time.Now().After(deadline) {
+			break
+		}
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-time.After(time.Second):
+		}
+	}
+	seen := "it lists no GPU at all"
+	if devs, err := fb.Devices(ctx); err != nil {
+		seen = "listing its devices failed: " + err.Error()
+	} else if len(devs) > 0 {
+		seen = "it lists " + strings.Join(devs, "; ")
+	}
+	return fmt.Errorf("placed on GPU %s but computing on the CPU: the engine holds no memory there. "+
+		"Its GPU backend did not load (%s). Check the build with `ldd` on its libggml-cuda.so, "+
+		"or run `llmgw doctor` on this node", joinIndices(missing), seen)
+}
+
+// holds reports whether any of pids is the shard's process or in its group:
+// an engine that computes in a child process is still the shard.
+func holds(pids []int, pid, group int) bool {
+	for _, p := range pids {
+		if p == pid || (group > 0 && processGroup(p) == group) {
+			return true
+		}
+	}
+	return false
+}
+
+// pgidOf reads a process's group from /proc; zero when it cannot.
+func pgidOf(pid int) int {
+	b, err := os.ReadFile("/proc/" + strconv.Itoa(pid) + "/stat")
+	if err != nil {
+		return 0
+	}
+	// The command name, in parentheses, may hold spaces; the fields after its
+	// closing parenthesis are state, ppid, pgrp.
+	s := string(b)
+	i := strings.LastIndexByte(s, ')')
+	if i < 0 {
+		return 0
+	}
+	f := strings.Fields(s[i+1:])
+	if len(f) < 3 {
+		return 0
+	}
+	n, _ := strconv.Atoi(f[2])
+	return n
+}
+
+func joinIndices(is []int) string {
+	s := make([]string, len(is))
+	for k, i := range is {
+		s[k] = strconv.Itoa(i)
+	}
+	return strings.Join(s, ", ")
+}
@@ -0,0 +1,91 @@
+package agent
+
+import (
+	"context"
+	"log/slog"
+	"strings"
+	"testing"
+	"time"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/engines"
+)
+
+// fallbackEngine is an engine that comes up on the CPU when its GPU backend
+// does not load, and lists the devices it can see.
+type fallbackEngine struct {
+	engines.Engine
+	devices []string
+}
+
+func (fallbackEngine) FallsBackToCPU() bool { return true }
+func (f fallbackEngine) Devices(context.Context) ([]string, error) {
+	return f.devices, nil
+}
+
+func TestAGPUShardComputingOnTheCPUIsNotReady(t *testing.T) {
+	procs := map[int][]int{}
+	oldProcs, oldGroup, oldWait := gpuProcesses, processGroup, gpuCheckWait
+	gpuProcesses = func(context.Context, string) (map[int][]int, error) { return procs, nil }
+	processGroup = func(pid int) int {
+		if pid >= 500 && pid < 600 {
+			return 500 // the shard's group: 500 and its children
+		}
+		return pid
+	}
+	gpuCheckWait = 0
+	t.Cleanup(func() { gpuProcesses, processGroup, gpuCheckWait = oldProcs, oldGroup, oldWait })
+
+	a := &Agent{log: slog.Default()}
+	spec := engines.ShardSpec{ShardID: "s", GPUIndices: []int{0, 1}}
+	eng := fallbackEngine{}
+
+	// On no GPU at all: refused, saying what the engine sees.
+	err := a.onItsGPUs(context.Background(), eng, spec, 500)
+	if err == nil || !strings.Contains(err.Error(), "computing on the CPU") || !strings.Contains(err.Error(), "lists no GPU") {
+		t.Fatalf("a shard on no GPU was let through: %v", err)
+	}
+	// On one of its two: refused, naming the one it is missing.
+	procs[0] = []int{500}
+	if err := a.onItsGPUs(context.Background(), eng, spec, 500); err == nil || !strings.Contains(err.Error(), "GPU 1 ") {
+		t.Fatalf("a shard on half its GPUs: %v", err)
+	}
+	// Another process on GPU 1 is not the shard.
+	procs[1] = []int{900}
+	if err := a.onItsGPUs(context.Background(), eng, spec, 500); err == nil {
+		t.Fatal("another process's memory was taken for the shard's")
+	}
+	// A child of the shard holding GPU 1 is the shard.
+	procs[1] = []int{900, 512}
+	if err := a.onItsGPUs(context.Background(), eng, spec, 500); err != nil {
+		t.Fatalf("a shard on both its GPUs: %v", err)
+	}
+
+	// An engine that fails rather than falling back is not checked, nor is a
+	// shard with no GPUs.
+	procs = map[int][]int{}
+	if err := a.onItsGPUs(context.Background(), struct{ engines.Engine }{}, spec, 500); err != nil {
+		t.Errorf("checked an engine that does not fall back: %v", err)
+	}
+	if err := a.onItsGPUs(context.Background(), eng, engines.ShardSpec{}, 500); err != nil {
+		t.Errorf("checked a CPU shard: %v", err)
+	}
+}
+
+func TestTheCheckWaitsForTheShardToShowUpOnItsGPU(t *testing.T) {
+	calls := 0
+	oldProcs, oldWait := gpuProcesses, gpuCheckWait
+	gpuProcesses = func(context.Context, string) (map[int][]int, error) {
+		calls++
+		if calls < 2 {
+			return map[int][]int{}, nil
+		}
+		return map[int][]int{0: {77}}, nil
+	}
+	gpuCheckWait = 5 * time.Second
+	t.Cleanup(func() { gpuProcesses, gpuCheckWait = oldProcs, oldWait })
+
+	a := &Agent{log: slog.Default()}
+	if err := a.onItsGPUs(context.Background(), fallbackEngine{}, engines.ShardSpec{GPUIndices: []int{0}}, 77); err != nil {
+		t.Fatalf("nvidia-smi lagging by a second failed the shard: %v", err)
+	}
+}
@@ -46,9 +46,30 @@ func (a *Agent) engineReportLocked() map[string]string {
 		out = map[string]string{}
 	}
 	maps.Copy(out, a.installedVers)
+	maps.DeleteFunc(out, func(e, _ string) bool { return a.disabled(e) })
 	return out
 }

+// byHand reports whether agent.yaml names this engine's binary, which opts
+// it out of being installed.
+func (a *Agent) byHand(engine string) bool {
+	if a.reg == nil {
+		return false
+	}
+	e, err := a.reg.Get(engine)
+	if err != nil {
+		return false
+	}
+	c, ok := e.(engines.Configured)
+	return ok && c.Configured()
+}
+
+// disabled reports whether agent.yaml says this node never runs an engine:
+// not installed when asked, and not advertised when found.
+func (a *Agent) disabled(engine string) bool {
+	return slices.Contains(a.cfg.Engines.Disabled, engine)
+}
+
 // provisionEvery is how often installed engines are checked again. A library
 // removed by an unrelated upgrade is found here rather than by the next
 // replica to start.
@@ -58,7 +79,16 @@ const provisionEvery = 10 * time.Minute
 func (a *Agent) setProvision(p *llmgwv1.Provision) {
 	wants := []provision.Want{}
 	for _, e := range p.GetEngines() {
-		wants = append(wants, provision.Want{Engine: e.GetEngine(), Version: e.GetVersion()})
+		if a.disabled(e.GetEngine()) {
+			continue // a control plane older than the node's say in it
+		}
+		if a.byHand(e.GetEngine()) {
+			// The operator named a binary for it: a local build, a CUDA
+			// runtime of their own. Installing one beside it would win —
+			// the installed build is used first — and quietly replace theirs.
+			continue
+		}
+		wants = append(wants, provision.Want{Engine: e.GetEngine(), Version: e.GetVersion(), Variant: e.GetVariant()})
 	}
 	a.mu.Lock()
 	a.provWants = wants
@@ -27,6 +27,9 @@ import (
 type Want struct {
 	Engine  string
 	Version string
+	// Variant is a build chosen for this node ("cu129"); empty leaves it to
+	// the node's hardware.
+	Variant string
 }

 // The states an engine's installation can be in.
@@ -135,7 +138,12 @@ func (in *Installer) one(ctx context.Context, w Want) Status {
 			"; install it by hand and name it in the agent's configuration"
 		return st
 	}
-	build, err := recipe.Choose(w.Version, in.Host)
+	if w.Variant != "" && !validVersion.MatchString(w.Variant) {
+		st.State = Unsupported
+		st.Detail = fmt.Sprintf("%q is not a build this agent will install", w.Variant)
+		return st
+	}
+	builds, err := recipe.Candidates(w.Version, w.Variant, in.Host)
 	if err != nil {
 		st.State, st.Detail = Unsupported, err.Error()
 		var nh NotHere
@@ -144,14 +152,18 @@ func (in *Installer) one(ctx context.Context, w Want) Status {
 		}
 		return st
 	}
-	st.Variant = build.Variant
-	dir := filepath.Join(in.root(), w.Engine, w.Version+"-"+build.Variant)

-	// Installed before: check it still runs, and trust nothing else about it.
-	if _, err := os.Stat(filepath.Join(dir, marker)); err == nil {
+	// Installed before: whichever candidate is, best first, checked that it
+	// still runs and trusted for nothing else.
+	for _, build := range builds {
+		dir := in.dirFor(w, build)
+		if _, err := os.Stat(filepath.Join(dir, marker)); err != nil {
+			continue
+		}
+		st := st.as(w, build)
 		// The packages are checked too. Removing one is exactly the kind of
-		// change an unrelated upgrade makes, and ldd below would catch it only
-		// as a missing library with no hint of which package held it.
+		// change an unrelated upgrade makes, and ldd below would catch it
+		// only as a missing library with no hint of which package held it.
 		if err := in.packages(ctx, build.Packages, &st); err != nil {
 			st.State, st.Detail = Failed, err.Error()
 			return st
@@ -161,24 +173,64 @@ func (in *Installer) one(ctx context.Context, w Want) Status {
 			return in.ready(st, build, bin, version)
 		}
 		in.log().Warn("an installed engine no longer runs; installing it again",
-			"engine", w.Engine, "version", w.Version, "err", err)
+			"engine", w.Engine, "version", st.Version, "variant", build.Variant, "err", err)
+		break
 	}

-	if err := in.install(ctx, recipe, w, build, dir, &st); err != nil {
-		st.State, st.Detail, st.Binary = Failed, err.Error(), ""
-		return st
+	// Each candidate in turn until one installs and is seen to work: a build
+	// the table was wrong about — a driver too old for it after all, a GPU
+	// it has no kernels for — costs its download, and the next is tried.
+	var tried []string
+	for _, build := range builds {
+		st := st.as(w, build)
+		if err := in.install(ctx, recipe, w, build, in.dirFor(w, build), &st); err != nil {
+			tried = append(tried, build.Variant+": "+err.Error())
+			in.log().Warn("an engine build did not work here; trying the next",
+				"engine", w.Engine, "variant", build.Variant, "err", err)
+			continue
+		}
+		bin, version, err := in.verify(ctx, in.dirFor(w, build), build)
+		if err != nil {
+			tried = append(tried, build.Variant+": "+err.Error())
+			continue
+		}
+		return in.ready(st, build, bin, version)
 	}
-	bin, version, err := in.verify(ctx, dir, build)
-	if err != nil {
-		st.State, st.Detail = Failed, err.Error()
-		return st
+	st.State, st.Binary = Failed, ""
+	st.Detail = strings.Join(tried, "; then ")
+	if len(builds) == 1 {
+		_, st.Detail, _ = strings.Cut(st.Detail, ": ")
 	}
-	return in.ready(st, build, bin, version)
+	return st
+}
+
+// dirFor is where a build of an engine is installed: by the version it is
+// and its variant, so two builds of one version sit side by side.
+func (in *Installer) dirFor(w Want, b Build) string {
+	return filepath.Join(in.root(), w.Engine, b.versionOr(w.Version)+"-"+b.Variant)
+}
+
+// versionOr is the version a build is: its own, when it is not the one asked
+// for.
+func (b Build) versionOr(asked string) string {
+	if b.Version != "" {
+		return b.Version
+	}
+	return asked
+}
+
+// as is a status for one build of what was asked for.
+func (st Status) as(w Want, b Build) Status {
+	st.Variant, st.Version = b.Variant, b.versionOr(w.Version)
+	return st
 }

 func (in *Installer) ready(st Status, b Build, bin, version string) Status {
 	st.State, st.Binary = Ready, bin
 	st.Detail = strings.TrimSpace(filepath.Base(bin) + " " + version)
+	if b.Note != "" {
+		st.Detail += "; " + b.Note
+	}
 	var missing []string
 	for _, f := range b.CPUFlags {
 		if !in.Host.HasCPUFlag(f) {
@@ -299,9 +351,55 @@ func (in *Installer) verify(ctx context.Context, dir string, b Build) (bin, vers
 	if err != nil {
 		return "", "", fmt.Errorf("%s --version failed: %v: %s", b.Binary, err, lastLine(out))
 	}
+	if b.GPUBackend != "" {
+		if err := in.verifyGPUBackend(tctx, dir, bin, b); err != nil {
+			return "", "", err
+		}
+	}
 	return bin, firstVersionLine(out), nil
 }

+// verifyGPUBackend checks a GPU build can compute on this node's GPUs.
+//
+// The binary's own ldd says nothing about it: the GPU backend is a library
+// the server opens at run time, and when it cannot — cuBLAS missing, a driver
+// too old — the server carries on with its CPU backend. So the build is asked
+// which devices it sees, and when it sees none, ldd on the backend says why.
+func (in *Installer) verifyGPUBackend(ctx context.Context, dir, bin string, b Build) error {
+	out, err := in.run(ctx, nil, bin, "--list-devices")
+	if err != nil {
+		return fmt.Errorf("%s --list-devices failed: %v: %s", b.Binary, err, lastLine(out))
+	}
+	if listsDevice(string(out)) {
+		return nil
+	}
+	why := "its GPU backend did not load"
+	if lib, err := findFile(dir, b.GPUBackend); err != nil {
+		why = "the build has no " + b.GPUBackend
+	} else if out, err := in.run(ctx, nil, "ldd", lib); err == nil {
+		if missing := lddMissing(string(out)); len(missing) > 0 {
+			why = b.GPUBackend + " cannot load: " + strings.Join(missing, "; ")
+		}
+	}
+	return fmt.Errorf("the %s build of %s sees no GPU, and would run on the CPU: %s", b.Variant, b.Binary, why)
+}
+
+// listsDevice reads llama-server --list-devices: a device line after
+// "Available devices:", which holds "(none)" when there is none.
+func listsDevice(out string) bool {
+	listing := false
+	for _, line := range strings.Split(out, "\n") {
+		line = strings.TrimSpace(line)
+		switch {
+		case strings.HasPrefix(line, "Available devices"):
+			listing = true
+		case listing && line != "" && line != "(none)":
+			return true
+		}
+	}
+	return false
+}
+
 // packages installs the operating system packages a build needs.
 //
 // Only those missing, and from the node's own package sources: a package the
@@ -89,7 +89,11 @@ type fakeSystem struct {
 	installed map[string]bool
 	tools     map[string]bool
 	lddOut    string
-	calls     []string
+	// devices is what llama-server --list-devices prints, and backendLdd
+	// what ldd says of its GPU backend.
+	devices    string
+	backendLdd string
+	calls      []string
 	// locked is how many more package commands find the lock held.
 	locked int
 }
@@ -127,8 +131,12 @@ func (f *fakeSystem) run(_ context.Context, _ []string, name string, args ...str
 			}
 		}
 		return nil, nil
+	case name == "ldd" && strings.HasSuffix(args[0], "libggml-cuda.so"):
+		return []byte(f.backendLdd), nil
 	case name == "ldd":
 		return []byte(f.lddOut), nil
+	case strings.HasSuffix(name, "llama-server") && args[0] == "--list-devices":
+		return []byte(f.devices), nil
 	case strings.HasSuffix(name, "llama-server"):
 		return []byte("version: 10991 (abc)\n"), nil
 	}
@@ -247,13 +255,38 @@ func TestAnArchiveCannotWriteOutsideItsDirectory(t *testing.T) {
 func TestTheBuildFollowsTheHardware(t *testing.T) {
 	gpu := ubuntu
 	gpu.NVIDIADriver = "570.86.15"
-	b, err := chooseLlamaCpp("b1", gpu)
-	if err != nil || !strings.HasPrefix(b.Variant, "cuda") || len(b.Assets) != 2 {
-		t.Errorf("an NVIDIA node got %+v, %v; expected the CUDA build and its runtime", b, err)
+	got, err := llamaCppCandidates("b1", "", gpu)
+	if err != nil || got[0].Variant != "cuda-12.8" || len(got[0].Assets) != 2 || got[len(got)-1].Variant != "vulkan" {
+		t.Errorf("an NVIDIA node on driver 570 got %+v, %v; expected CUDA 12.8 with its runtime, then Vulkan", got, err)
 	}
-	if _, err := chooseLlamaCpp("b1", discovery.Host{OSID: "ubuntu", Arch: "riscv64"}); err == nil {
+	if _, err := llamaCppCandidates("b1", "", discovery.Host{OSID: "ubuntu", Arch: "riscv64"}); err == nil {
 		t.Error("a build was chosen for an architecture llama.cpp does not publish for")
 	}
+	for _, c := range []struct {
+		name string
+		host discovery.Host
+		want []string
+	}{
+		// Driver 580 and Turing: CUDA 13.3 first.
+		{"current", discovery.Host{Arch: "amd64", NVIDIADriver: "580.65.06", MinComputeCapability: 8.0}, []string{"cuda-13.3", "cuda-12.8", "vulkan"}},
+		// A V100 on driver 550: CUDA 12.8 ships 7.0 as PTX, which needs 570.
+		{"V100, 550", discovery.Host{Arch: "amd64", NVIDIADriver: "550.54.15", MinComputeCapability: 7.0}, []string{"vulkan"}},
+		// An RTX 3090 on 550: 8.6 is machine code in the CUDA 12.8 build.
+		{"3090, 550", discovery.Host{Arch: "amd64", NVIDIADriver: "550.54.15", MinComputeCapability: 8.6}, []string{"cuda-12.8", "vulkan"}},
+		// A Pascal card: CUDA 12.8, not 13.3.
+		{"P40", discovery.Host{Arch: "amd64", NVIDIADriver: "575.51.03", MinComputeCapability: 6.1}, []string{"cuda-12.8", "vulkan"}},
+		// arm64 with a GPU: never the CPU build.
+		{"GH200", discovery.Host{Arch: "arm64", NVIDIADriver: "580.65.06", MinComputeCapability: 9.0}, []string{"cuda-13.3", "vulkan"}},
+	} {
+		got, err := llamaCppCandidates("b1", "", c.host)
+		var names []string
+		for _, b := range got {
+			names = append(names, b.Variant)
+		}
+		if err != nil || strings.Join(names, " ") != strings.Join(c.want, " ") {
+			t.Errorf("%s: %v, %v; want %v", c.name, names, err, c.want)
+		}
+	}

 	slow := ubuntu
 	slow.CPUFlags = nil
@@ -388,3 +421,32 @@ func TestALibrarysVersionLinkIsStillUnpacked(t *testing.T) {
 		t.Errorf("link %q %v", target, err)
 	}
 }
+
+func TestACUDABuildThatSeesNoGPUIsNotInstalled(t *testing.T) {
+	// Found in the field: libggml-cuda.so in the build, cuBLAS nowhere the
+	// loader looks, and llama-server serving a GPU placement on the CPU.
+	cuda := "bin-ubuntu-cuda-12.8-x64.tar.gz"
+	rel := newRelease(map[string][]byte{
+		"llama-b1-" + cuda: tarGz(t, map[string]string{
+			"llama-b1/llama-server": "#!/bin/sh\n", "llama-b1/libggml-cuda.so": "x"}, nil),
+		"cudart-llama-b1-" + cuda: tarGz(t, map[string]string{"libcudart.so.12": "x"}, nil),
+	})
+	gpu := ubuntu
+	gpu.NVIDIADriver = "570.86.15"
+	sys := &fakeSystem{installed: map[string]bool{"libgomp1": true},
+		devices:    "Available devices:\n  (none)\n",
+		backendLdd: "\tlibcublas.so.12 => not found\n\tlibcudart.so.12 => /x/libcudart.so.12\n"}
+	in, _ := installer(t, rel.serve(t), sys, gpu)
+
+	got := in.Converge(t.Context(), []Want{{Engine: "llamacpp", Version: "b1"}})
+	if got[0].State != Failed || !strings.Contains(got[0].Detail, "sees no GPU") ||
+		!strings.Contains(got[0].Detail, "libcublas.so.12 is not installed") {
+		t.Fatalf("expected a failure naming the missing library: %+v", got[0])
+	}
+
+	// The same build seeing its GPU is installed.
+	sys.devices = "Available devices:\n  CUDA0: NVIDIA L4 (22994 MiB, 22713 MiB free)\n"
+	if got := in.Converge(t.Context(), []Want{{Engine: "llamacpp", Version: "b1"}}); got[0].State != Ready {
+		t.Fatalf("a CUDA build that sees its GPU: %+v", got[0])
+	}
+}
@@ -36,6 +36,10 @@ const lockName = "llmgw-requirements.txt"
 func (in *Installer) installPython(ctx context.Context, r Recipe, w Want, b Build, dir string, st *Status) error {
 	ctx, cancel := context.WithTimeout(ctx, pythonInstallTimeout)
 	defer cancel()
+	// The version this build is, which for a GPU older than every build of
+	// the one asked for is an older one: its publication date is the one
+	// the dependencies are resolved as of.
+	w.Version = b.versionOr(w.Version)

 	if err := in.packages(ctx, b.Packages, st); err != nil {
 		return err
@@ -86,28 +90,67 @@ func (in *Installer) installPython(ctx context.Context, r Recipe, w Want, b Buil
 	}
 	python := filepath.Join(tmp, "bin", "python")

-	in.progress(st, "resolving %s against PyPI as of %s", strings.Join(b.Requirements, " "), asOf[:10])
+	reqs := b.Requirements
+	if b.Wheel != nil {
+		// The engine from its GitHub release, pinned to the digest GitHub
+		// publishes for the file: the same trust as an archive, and uv
+		// refuses a download that does not match it.
+		in.progress(st, "asking GitHub for %s %s's %s…%s", b.Wheel.Repo, b.Wheel.Tag, b.Wheel.Prefix, b.Wheel.Suffix)
+		sums, err := in.checksums(ctx, b.Wheel.Repo, b.Wheel.Tag)
+		if err != nil {
+			return err
+		}
+		var asset, sum string
+		for name, s := range sums {
+			if strings.HasPrefix(name, b.Wheel.Prefix) && strings.HasSuffix(name, b.Wheel.Suffix) {
+				asset, sum = name, s
+			}
+		}
+		if asset == "" {
+			return fmt.Errorf("%s %s publishes no %s…%s", b.Wheel.Repo, b.Wheel.Tag, b.Wheel.Prefix, b.Wheel.Suffix)
+		}
+		url := strings.TrimSuffix(in.downloadBase(), "/") + "/" + b.Wheel.Repo + "/releases/download/" +
+			b.Wheel.Tag + "/" + asset
+		reqs = append([]string{b.Wheel.Project + " @ " + url + "#sha256=" + sum}, reqs...)
+	}
+	// A second index, for the torch the engine was built on. Every package
+	// is taken from whichever of the two has the version resolved — both are
+	// the publishers' own — and each file is still pinned by its hash.
+	var index []string
+	if b.ExtraIndex != "" {
+		index = []string{"--extra-index-url", b.ExtraIndex, "--index-strategy", "unsafe-best-match"}
+	}
+
+	in.progress(st, "resolving %s against PyPI as of %s", strings.Join(reqs, " "), asOf[:10])
 	lock := filepath.Join(tmp, lockName)
-	args := []string{"pip", "compile", "--python", python, "--exclude-newer", asOf,
-		"--generate-hashes", "--quiet", "--output-file", lock, "-"}
-	if out, err := in.runStdin(ctx, env, strings.Join(b.Requirements, "\n")+"\n", uv.Binary, args...); err != nil {
-		return fmt.Errorf("resolving %s: %v: %s", strings.Join(b.Requirements, " "), err, tail(out))
+	args := append([]string{"pip", "compile", "--python", python, "--exclude-newer", asOf,
+		"--generate-hashes", "--quiet", "--output-file", lock}, index...)
+	args = append(args, "-")
+	if out, err := in.runStdin(ctx, env, strings.Join(reqs, "\n")+"\n", uv.Binary, args...); err != nil {
+		return fmt.Errorf("resolving %s: %v: %s", strings.Join(reqs, " "), err, tail(out))
 	}

 	in.progress(st, "downloading and installing %s and its dependencies: several gigabytes, which takes minutes",
-		strings.Join(b.Requirements, " "))
-	if out, err := in.run(ctx, env, uv.Binary, "pip", "install", "--python", python,
-		"--require-hashes", "--no-deps", "-r", lock); err != nil {
-		return fmt.Errorf("installing %s: %v: %s", strings.Join(b.Requirements, " "), err, tail(out))
+		strings.Join(reqs, " "))
+	install := append([]string{"pip", "install", "--python", python,
+		"--require-hashes", "--no-deps", "-r", lock}, index...)
+	if out, err := in.run(ctx, env, uv.Binary, install...); err != nil {
+		return fmt.Errorf("installing %s: %v: %s", strings.Join(reqs, " "), err, tail(out))
 	}

 	in.progress(st, "checking that %s runs", w.Engine)
 	if _, _, err := in.verify(ctx, tmp, b); err != nil {
 		return err
 	}
+	if b.GPUCheck && in.Host.NVIDIADriver != "" {
+		in.progress(st, "checking that %s (%s) computes on this node's GPU", w.Engine, b.Variant)
+		if err := in.gpuCheck(ctx, python, w, b); err != nil {
+			return err
+		}
+	}
 	m, _ := json.MarshalIndent(markerFile{
 		Engine: w.Engine, Version: w.Version, Variant: b.Variant,
-		Binary: filepath.Join("bin", b.Binary), Requirements: b.Requirements,
+		Binary: filepath.Join("bin", b.Binary), Requirements: reqs,
 		ExcludeNewer: asOf, UV: UVVersion, Installed: time.Now().UTC(),
 	}, "", "  ")
 	if err := os.WriteFile(filepath.Join(tmp, marker), m, 0o644); err != nil { //nolint:gosec // not secret
@@ -203,3 +246,26 @@ func tail(b []byte) string {
 	}
 	return strings.Join(lines, " / ")
 }
+
+// gpuProbe is what the environment is asked to run on the GPU: one small
+// operation, which needs the torch to have kernels for the card.
+const gpuProbe = `import torch
+assert torch.cuda.is_available(), "torch sees no CUDA device"
+x = torch.ones(64, device="cuda")
+print(float((x * 2).sum()), torch.cuda.get_device_name(0), torch.cuda.get_device_capability(0))`
+
+// gpuCheck runs gpuProbe in a freshly installed environment.
+func (in *Installer) gpuCheck(ctx context.Context, python string, w Want, b Build) error {
+	ctx, cancel := context.WithTimeout(ctx, 3*time.Minute)
+	defer cancel()
+	out, err := in.run(ctx, nil, python, "-c", gpuProbe)
+	if err != nil {
+		hint := ""
+		if strings.Contains(string(out), "no kernel image") {
+			hint = "; this build has no kernels for this GPU, so try the other build for this node"
+		}
+		return fmt.Errorf("%s %s (%s) installed but cannot compute on this node's GPU: %s%s",
+			w.Engine, w.Version, b.Variant, tail(out), hint)
+	}
+	return nil
+}
@@ -2,6 +2,7 @@ package provision

 import (
 	"context"
+	"errors"
 	"net/http"
 	"net/http/httptest"
 	"os"
@@ -23,6 +24,10 @@ type fakePython struct {
 	pmu      sync.Mutex
 	commands []string
 	failPip  bool
+	// noKernel makes the GPU check fail as a torch without kernels for the
+	// card does: in every build, or in noKernelIn only ("-cu13.").
+	noKernel   bool
+	noKernelIn string
 }

 func (f *fakePython) run(ctx context.Context, env []string, name string, args ...string) ([]byte, error) {
@@ -43,6 +48,11 @@ func (f *fakePython) run(ctx context.Context, env []string, name string, args ..
 		return nil, os.WriteFile(filepath.Join(venv, "bin", "vllm"), []byte("#!/bin/sh\n"), 0o755)
 	case strings.HasSuffix(name, "/bin/vllm") && args[0] == "--version":
 		return []byte("0.30.0\n"), nil
+	case strings.HasSuffix(name, "/bin/python") && args[0] == "-c":
+		if f.noKernel || (f.noKernelIn != "" && strings.Contains(name, f.noKernelIn)) {
+			return []byte("RuntimeError: CUDA error: no kernel image is available for execution on the device\n"), errExit
+		}
+		return []byte("128.0 Tesla V100-SXM2-32GB (7, 0)\n"), nil
 	}
 	return f.fakeSystem.run(ctx, env, name, args...)
 }
@@ -59,6 +69,8 @@ func pythonInstaller(t *testing.T, host discovery.Host, sys *fakePython) (*Insta
 	uv := newRelease(map[string][]byte{
 		"uv-x86_64-unknown-linux-gnu.tar.gz": tarGz(t,
 			map[string]string{"uv-x86_64-unknown-linux-gnu/uv": "#!/bin/sh\n"}, nil),
+		// The fake GitHub serves every repository's release from this one.
+		"vllm-0.30.0+cu129-cp38-abi3-manylinux_2_28_x86_64.whl": []byte("the cu129 wheel"),
 	})
 	gh := uv.serve(t)
 	mux := http.NewServeMux()
@@ -66,6 +78,9 @@ func pythonInstaller(t *testing.T, host discovery.Host, sys *fakePython) (*Insta
 		_, _ = w.Write([]byte(`{"urls":[{"upload_time_iso_8601":"2026-09-22T05:22:21.647189Z"},
 			{"upload_time_iso_8601":"2026-09-22T05:23:10.039243Z"}]}`))
 	})
+	mux.HandleFunc("/pypi/vllm/0.19.1/json", func(w http.ResponseWriter, _ *http.Request) {
+		_, _ = w.Write([]byte(`{"urls":[{"upload_time_iso_8601":"2026-04-02T10:00:00Z"}]}`))
+	})
 	pypi := httptest.NewServer(mux)
 	t.Cleanup(pypi.Close)
 	return &Installer{
@@ -130,8 +145,10 @@ func TestVLLMIsNotForEveryNode(t *testing.T) {
 		says  string
 	}{
 		{"no GPU", ubuntu, Skipped, "has none"},
-		{"old driver", discovery.Host{Arch: "amd64", NVIDIADriver: "550.54.15"}, Unsupported, "driver 580"},
-		{"V100", discovery.Host{Arch: "amd64", NVIDIADriver: "580.65.06", MinComputeCapability: 7.0}, Unsupported, "7.5"},
+		// Too old for CUDA 13 and for its CUDA 12.9 build both.
+		{"a driver too old for either", discovery.Host{Arch: "amd64", NVIDIADriver: "470.256.02"}, Unsupported, "driver 525"},
+		// No vLLM was ever built for Pascal.
+		{"P100", discovery.Host{Arch: "amd64", NVIDIADriver: "580.65.06", MinComputeCapability: 6.0}, Unsupported, "no vLLM is built for GPUs of compute capability 6.0"},
 	}
 	for _, c := range cases {
 		t.Run(c.name, func(t *testing.T) {
@@ -182,3 +199,114 @@ func TestAVersionPyPIDoesNotHaveIsSaidSo(t *testing.T) {
 		t.Errorf("got %s %q", got[0].State, got[0].Detail)
 	}
 }
+
+func TestAnOlderDriverGetsTheSameVLLMBuiltForCUDA12(t *testing.T) {
+	// A Turing card on a driver too old for CUDA 13: the same vLLM's CUDA
+	// 12.9 build, from the project's GitHub release, with the torch PyTorch
+	// publishes for 12.9.
+	volta := discovery.Host{OSID: "ubuntu", Arch: "amd64", NVIDIADriver: "550.54.15", MinComputeCapability: 7.5}
+	sys := &fakePython{fakeSystem: fakeSystem{installed: map[string]bool{}}}
+	in, gh := pythonInstaller(t, volta, sys)
+	got := in.Converge(t.Context(), []Want{{Engine: "vllm", Version: "0.30.0"}})
+	if got[0].State != Ready || got[0].Variant != "cu129" {
+		t.Fatalf("a Turing node on driver 550: %+v", got[0])
+	}
+	if _, err := os.Stat(filepath.Join(in.Root, "vllm", "0.30.0-cu129", marker)); err != nil {
+		t.Errorf("not installed under its own build's name: %v", err)
+	}
+	all := strings.Join(sys.commands, "\n")
+	wheel := "vllm-0.30.0+cu129-cp38-abi3-manylinux_2_28_x86_64.whl"
+	for _, want := range []string{
+		// Pinned to the digest GitHub publishes for the wheel.
+		"vllm @ " + in.DownloadBase + "/vllm-project/vllm/releases/download/v0.30.0/" + wheel + "#sha256=" +
+			strings.TrimPrefix(gh.digests[wheel], "sha256:"),
+		"--extra-index-url https://download.pytorch.org/whl/cu129 --index-strategy unsafe-best-match -",
+		"--require-hashes --no-deps -r",
+		// Computing on the card, not only importing.
+		"/bin/python -c import torch",
+	} {
+		if !strings.Contains(all, want) {
+			t.Errorf("no %q in:\n%s", want, all)
+		}
+	}
+	for _, c := range sys.commands {
+		if strings.Contains(c, "pip install") && !strings.Contains(c, "download.pytorch.org/whl/cu129") {
+			t.Errorf("installed without the index its torch comes from: %s", c)
+		}
+	}
+}
+
+func TestABuildChosenForANodeIsTheOneInstalled(t *testing.T) {
+	sys := &fakePython{fakeSystem: fakeSystem{installed: map[string]bool{}}}
+	in, _ := pythonInstaller(t, gpuNode, sys)
+	if got := in.Converge(t.Context(), []Want{{Engine: "vllm", Version: "0.30.0", Variant: "cu129"}}); got[0].Variant != "cu129" || got[0].State != Ready {
+		t.Errorf("cu129 chosen for a node that runs cu13 too: %+v", got[0])
+	}
+	if _, err := vllmCandidates("0.30.0", "cu200", gpuNode); err == nil {
+		t.Error("a build that does not exist was accepted")
+	}
+}
+
+func TestABuildThatCannotComputeOnTheCardIsNotInstalled(t *testing.T) {
+	sys := &fakePython{fakeSystem: fakeSystem{installed: map[string]bool{}}, noKernel: true}
+	in, _ := pythonInstaller(t, gpuNode, sys)
+	got := in.Converge(t.Context(), []Want{{Engine: "vllm", Version: "0.30.0"}})
+	if got[0].State != Failed || !strings.Contains(got[0].Detail, "no kernels for this GPU") {
+		t.Fatalf("a build without kernels for the card: %+v", got[0])
+	}
+	if _, err := os.Stat(filepath.Join(in.Root, "vllm", "0.30.0-cu13", marker)); err == nil {
+		t.Error("marked installed")
+	}
+}
+
+func TestOnlyACUDARequirementFallsBackToTheCUDA12Build(t *testing.T) {
+	// A machine vLLM is not published for is refused, not given the other
+	// build to fail with.
+	if _, err := vllmCandidates("0.30.0", "", discovery.Host{NVIDIADriver: "580.65.06"}); err == nil ||
+		!strings.Contains(err.Error(), "no Linux build") {
+		t.Errorf("an unknown architecture: %v", err)
+	}
+	if _, err := vllmCandidates("0.30.0", "cu129", ubuntu); !errors.As(err, new(NotHere)) {
+		t.Errorf("a node with no GPU chosen cu129: %v", err)
+	}
+	// Driver 580 and Turing: CUDA 13 first, the CUDA 12.9 build behind it.
+	got, err := vllmCandidates("0.30.0", "", gpuNode)
+	if err != nil || len(got) != 2 || got[0].Variant != "cu13" || got[1].Variant != "cu129" {
+		t.Errorf("candidates on a current node: %+v, %v", got, err)
+	}
+}
+
+func TestAV100GetsTheNewestVLLMBuiltForItByItself(t *testing.T) {
+	// 0.30.0 asked for; no build of it has Volta kernels. The node installs
+	// 0.19.1, the newest that has, and says so.
+	v100 := discovery.Host{OSID: "ubuntu", Arch: "amd64", NVIDIADriver: "550.54.15", MinComputeCapability: 7.0}
+	sys := &fakePython{fakeSystem: fakeSystem{installed: map[string]bool{}}}
+	in, _ := pythonInstaller(t, v100, sys)
+	got := in.Converge(t.Context(), []Want{{Engine: "vllm", Version: "0.30.0"}})[0]
+	if got.State != Ready || got.Version != "0.19.1" || got.Variant != "cu12" ||
+		!strings.Contains(got.Detail, "the newest built for compute capability 7.0, instead of 0.30.0") {
+		t.Fatalf("a V100: %+v", got)
+	}
+	if _, err := os.Stat(filepath.Join(in.Root, "vllm", "0.19.1-cu12", marker)); err != nil {
+		t.Errorf("not installed as what it is: %v", err)
+	}
+	all := strings.Join(sys.commands, "\n")
+	if !strings.Contains(all, "<<< vllm==0.19.1") || strings.Contains(all, "download.pytorch.org") {
+		t.Errorf("0.19.1 is PyPI's CUDA 12 build, with PyPI's torch:\n%s", all)
+	}
+	// Chosen cu13, which 0.19.1 has none of: refused, saying why.
+	if _, err := vllmCandidates("0.30.0", "cu13", v100); err == nil {
+		t.Error("a CUDA 13 build was offered to a V100")
+	}
+}
+
+func TestABuildThatFailsOnTheCardGivesWayToTheNext(t *testing.T) {
+	// The table says CUDA 13 runs here; the card says otherwise. The next
+	// build is installed rather than the node going without.
+	sys := &fakePython{fakeSystem: fakeSystem{installed: map[string]bool{}}, noKernelIn: "-cu13."}
+	in, _ := pythonInstaller(t, gpuNode, sys)
+	got := in.Converge(t.Context(), []Want{{Engine: "vllm", Version: "0.30.0"}})[0]
+	if got.State != Ready || got.Variant != "cu129" {
+		t.Fatalf("after the CUDA 13 build failed on the card: %+v", got)
+	}
+}
@@ -43,12 +43,41 @@ type Build struct {
 	// VersionArgs is what Binary is run with to prove the build works, when
 	// --version is not it.
 	VersionArgs []string
+	// GPUBackend, when set, is the shared library the build computes on GPUs
+	// with, loaded at run time rather than linked: "libggml-cuda.so". The
+	// build must list a device with it, or it is not installed — a llama.cpp
+	// whose CUDA backend cannot load runs on the CPU without complaint.
+	GPUBackend string
+	// Wheel, when set, is the engine itself as a wheel from a GitHub
+	// release rather than from PyPI, checked against the digest GitHub
+	// publishes for it. ExtraIndex is a package index besides PyPI that the
+	// rest resolves from too: the CUDA 12.9 torch the wheel was built on.
+	Wheel      *Wheel
+	ExtraIndex string
+	// GPUCheck, on a node with an NVIDIA GPU, has the installed environment
+	// compute something on one before the build counts as installed. A build
+	// with no kernels for the card — a GPU generation its torch dropped —
+	// imports and prints its version, and fails on its first model.
+	GPUCheck bool
+	// Version is the engine version this build is, when it is not the one
+	// asked for: the newest built for this machine's GPUs, where the one
+	// asked for has no kernels for them. Empty is the version asked for.
+	Version string
+	// Note says why this build, when it is not the plain one: shown beside
+	// it wherever the installation is.
+	Note string
 	// FreeGB is the disk space the installation needs before it starts: a
 	// Python engine is several gigabytes, and one that fills the disk halfway
 	// leaves the node worse off than one that never began.
 	FreeGB int
 }

+// Wheel is a package published as a GitHub release asset: the asset whose
+// name starts with Prefix and ends with Suffix.
+type Wheel struct {
+	Project, Repo, Tag, Prefix, Suffix string
+}
+
 // Recipe knows how to install one engine.
 type Recipe struct {
 	// Repo is the GitHub repository whose releases carry the builds; empty
@@ -56,8 +85,12 @@ type Recipe struct {
 	Repo string
 	// PyPI is the project an engine installed from PyPI is published as.
 	PyPI string
-	// Choose picks the build for a machine, or says why none fits.
-	Choose func(version string, h discovery.Host) (Build, error)
+	// Candidates are the builds that may run on a machine, best first, or
+	// why none can. The installer takes the first that installs and is seen
+	// to work — computing on the GPU, where there is one — so a build this
+	// table is wrong about costs a download, not a node. variant is a build
+	// chosen for the node; empty leaves the choice to the machine.
+	Candidates func(version, variant string, h discovery.Host) ([]Build, error)
 }

 // Recipes are the engines this agent can install.
@@ -66,10 +99,10 @@ type Recipe struct {
 // from PyPI by uv, which is itself installed like llama.cpp: a GitHub release
 // checked against its published checksum.
 var Recipes = map[string]Recipe{
-	"llamacpp": {Repo: "ggml-org/llama.cpp", Choose: chooseLlamaCpp},
-	"vllm":     {PyPI: "vllm", Choose: chooseVLLM},
-	"sglang":   {PyPI: "sglang", Choose: chooseSGLang},
-	"uv":       {Repo: "astral-sh/uv", Choose: chooseUV},
+	"llamacpp": {Repo: "ggml-org/llama.cpp", Candidates: llamaCppCandidates},
+	"vllm":     {PyPI: "vllm", Candidates: vllmCandidates},
+	"sglang":   {PyPI: "sglang", Candidates: sglangCandidates},
+	"uv":       {Repo: "astral-sh/uv", Candidates: uvCandidates},
 }

 // Installable are the engines this agent can install, as it tells the
@@ -99,7 +132,15 @@ type NotHere struct{ Reason string }

 func (e NotHere) Error() string { return e.Reason }

-func chooseUV(version string, h discovery.Host) (Build, error) {
+func uvCandidates(version, _ string, h discovery.Host) ([]Build, error) {
+	b, err := chooseUV(version, h)
+	if err != nil {
+		return nil, err
+	}
+	return []Build{b}, nil
+}
+
+func chooseUV(_ string, h discovery.Host) (Build, error) {
 	var arch string
 	switch h.Arch {
 	case "amd64":
@@ -116,33 +157,227 @@ func chooseUV(version string, h discovery.Host) (Build, error) {
 	}, nil
 }

-// minDriverForCUDA13 is the oldest NVIDIA driver CUDA 13 runs on. Every vLLM
-// since 0.26 and SGLang 0.5 is built on torch for CUDA 13.
-const minDriverForCUDA13 = 580
+// What builds need of the NVIDIA side of a machine, read from what each
+// project publishes and how it was built:
+//
+//   - CUDA 13 needs driver 580; CUDA 12 builds run on driver 525 and newer,
+//     by CUDA's minor version compatibility — except code shipped only as
+//     PTX, which the driver compiles and so needs a driver as new as the
+//     CUDA it was compiled with.
+//   - Which GPUs a build has kernels for is the build's own list, not what
+//     its CUDA supports.
+const (
+	minDriverForCUDA13 = 580
+	minDriverForCUDA12 = 525
+)

-// minComputeForCUDA13 is the oldest GPU CUDA 13 supports: Turing. It dropped
-// Volta, so a V100 cannot run these builds at all.
-const minComputeForCUDA13 = 7.5
+// gpuNeeds is what one build needs of a machine's GPUs and driver.
+type gpuNeeds struct {
+	driver int
+	cc     float64
+}

-// cudaGate says whether a node can run an engine built on CUDA 13.
-func cudaGate(engine string, h discovery.Host) error {
+// fits says why a machine cannot run a build, or nil.
+func (n gpuNeeds) fits(what string, h discovery.Host) error {
+	if major := leadingInt(h.NVIDIADriver); n.driver > 0 && major > 0 && major < n.driver {
+		return fmt.Errorf("%s needs NVIDIA driver %d or newer; this node has %s", what, n.driver, h.NVIDIADriver)
+	}
+	if cc := h.MinComputeCapability; n.cc > 0 && cc > 0 && cc < n.cc {
+		return fmt.Errorf("%s has kernels for compute capability %.1f and newer; this node has %.1f", what, n.cc, cc)
+	}
+	return nil
+}
+
+// candidate is a build and what it needs.
+type candidate struct {
+	b     Build
+	needs gpuNeeds
+	what  string
+}
+
+// pick keeps the candidates a machine can run, in order, and those a
+// variant chosen for the node names; with none left, it says why each went.
+func pick(engine, variant string, cands []candidate, h discovery.Host) ([]Build, error) {
+	var out []Build
+	var why []string
+	for _, c := range cands {
+		if variant != "" && c.b.Variant != variant {
+			continue
+		}
+		if err := c.needs.fits(c.what, h); err != nil {
+			why = append(why, err.Error())
+			continue
+		}
+		out = append(out, c.b)
+	}
+	switch {
+	case len(out) > 0:
+		return out, nil
+	case len(why) == 0 && variant != "":
+		var names []string
+		for _, c := range cands {
+			names = append(names, c.b.Variant)
+		}
+		return nil, fmt.Errorf("%s has no build %q here; it has %s", engine, variant, strings.Join(names, ", "))
+	}
+	return nil, fmt.Errorf("no build of %s runs on this node: %s", engine, strings.Join(why, "; "))
+}
+
+// --- vLLM ---
+
+// VLLMVariants are the builds of vLLM a node can be given by choice: CUDA
+// 13 from PyPI, and CUDA 12.9 from the project's GitHub release — the same
+// version, built for drivers CUDA 13 does not run on.
+var VLLMVariants = []string{"cu13", "cu129"}
+
+// pytorchCU129 is where the CUDA 12.9 build's torch is published. PyPI's
+// torch is built for CUDA 13; the cu129 vLLM wheel needs this one.
+const pytorchCU129 = "https://download.pytorch.org/whl/cu129"
+
+// vllmCUDA13Since is the first vLLM whose PyPI build is CUDA 13, read from
+// its releases: 0.19.1 publishes "+cu130" beside a CUDA 12 default, 0.20.0
+// "+cu129" beside a CUDA 13 one.
+const vllmCUDA13Since = "0.20"
+
+// vllmMinCompute is the oldest GPU a vLLM version's builds have kernels for,
+// from its release pipeline: on x86, 7.0 (Volta) up to 0.19.1 — the
+// Dockerfile's default "7.0 7.5 8.0 8.9 9.0 10.0 12.0" — and 7.5 from 0.20.0,
+// CUDA 12.9 build included; on arm64, built for Grace machines, 8.7 before
+// 0.20 and 8.0 since.
+func vllmMinCompute(version, arch string) float64 {
+	old := versionBefore(version, vllmCUDA13Since)
+	switch {
+	case arch == "arm64" && old:
+		return 8.7
+	case arch == "arm64":
+		return 8.0
+	case old:
+		return 7.0
+	}
+	return 7.5
+}
+
+// vllmNewestFor is the newest vLLM built for GPUs of compute capability cc
+// on an architecture, for GPUs older than every build of the version asked
+// for: a V100 gets 0.19.1, where 0.20 and later have nothing for it. Nothing
+// for a GPU no vLLM was built for.
+var vllmNewestFor = []struct {
+	arch    string
+	from    float64 // the oldest GPU it covers
+	version string
+}{
+	{"amd64", 7.0, "0.19.1"},
+}
+
+// vllmCandidates are the vLLM builds for a machine.
+//
+// The version asked for, in the build the machine runs: CUDA 13 where the
+// driver allows, the same version's CUDA 12.9 build where it does not. One
+// version across the cluster wherever the GPUs allow it, because the flags
+// llmgw builds are checked against it. Where they do not — GPUs older than
+// every build of that version — the newest version built for them, said so
+// beside the installation: vLLM on a V100 at all, rather than not.
+func vllmCandidates(version, variant string, h discovery.Host) ([]Build, error) {
+	arch, err := gpuArch("vLLM", h)
+	if err != nil {
+		return nil, err
+	}
+	v, note := version, ""
+	if cc := h.MinComputeCapability; cc > 0 && cc < vllmMinCompute(version, h.Arch) {
+		older := ""
+		for _, n := range vllmNewestFor {
+			if n.arch == h.Arch && cc >= n.from {
+				older = n.version
+			}
+		}
+		if older == "" {
+			return nil, fmt.Errorf("no vLLM is built for GPUs of compute capability %.1f on %s; "+
+				"the node serves llama.cpp models", cc, arch)
+		}
+		v, note = older, fmt.Sprintf("vLLM %s, the newest built for compute capability %.1f, instead of %s", older, cc, version)
+	}
+	pypi := func(variant string) Build {
+		b := Build{
+			Variant:      variant,
+			Python:       pythonForEngines,
+			Requirements: []string{"vllm==" + v},
+			Binary:       "vllm",
+			Packages:     compilerPackages,
+			FreeGB:       25,
+			GPUCheck:     true,
+			Note:         note,
+		}
+		if v != version {
+			b.Version = v
+		}
+		return b
+	}
+	minCC := vllmMinCompute(v, h.Arch)
+	if versionBefore(v, vllmCUDA13Since) {
+		// PyPI's own build is CUDA 12, and is the one there is; asked for
+		// cu129, this is it.
+		if variant == "cu129" {
+			variant = "cu12"
+		}
+		return pick("vLLM "+v, variant, []candidate{
+			{pypi("cu12"), gpuNeeds{minDriverForCUDA12, minCC}, "vLLM " + v},
+		}, h)
+	}
+	cu129 := pypi("cu129")
+	cu129.Requirements = nil
+	// Found in the release's assets rather than named: the manylinux tag
+	// in the file's name changes between versions (2_28, 2_31, 2_34).
+	cu129.Wheel = &Wheel{
+		Project: "vllm", Repo: "vllm-project/vllm", Tag: "v" + v,
+		Prefix: "vllm-" + v + "+cu129-", Suffix: "_" + map[string]string{"amd64": "x86_64", "arm64": "aarch64"}[h.Arch] + ".whl",
+	}
+	cu129.ExtraIndex = pytorchCU129
+	return pick("vLLM "+v, variant, []candidate{
+		{pypi("cu13"), gpuNeeds{minDriverForCUDA13, minCC}, "vLLM " + v + " built for CUDA 13"},
+		{cu129, gpuNeeds{minDriverForCUDA12, minCC}, "vLLM " + v + " built for CUDA 12.9"},
+	}, h)
+}
+
+// gpuArch says whether a Python GPU engine is published for this machine
+// and whether it has an NVIDIA GPU at all.
+func gpuArch(engine string, h discovery.Host) (string, error) {
 	switch h.Arch {
 	case "amd64", "arm64":
 	default:
-		return fmt.Errorf("%s publishes no Linux build for %s", engine, orUnknown(h.Arch))
+		return "", fmt.Errorf("%s publishes no Linux build for %s", engine, orUnknown(h.Arch))
 	}
 	if h.NVIDIADriver == "" {
-		return NotHere{engine + " runs on NVIDIA GPUs and this node has none; it serves llama.cpp models"}
+		return "", NotHere{engine + " runs on NVIDIA GPUs and this node has none; it serves llama.cpp models"}
 	}
-	if major := leadingInt(h.NVIDIADriver); major > 0 && major < minDriverForCUDA13 {
-		return fmt.Errorf("%s is built for CUDA 13, which needs NVIDIA driver %d or newer; this node has %s",
-			engine, minDriverForCUDA13, h.NVIDIADriver)
+	return h.Arch, nil
+}
+
+// --- SGLang ---
+
+// sglangCUDA13Since is the first SGLang built for CUDA 13.
+const sglangCUDA13Since = "0.5"
+
+func sglangCandidates(version, variant string, h discovery.Host) ([]Build, error) {
+	if _, err := gpuArch("SGLang", h); err != nil {
+		return nil, err
 	}
-	if cc := h.MinComputeCapability; cc > 0 && cc < minComputeForCUDA13 {
-		return fmt.Errorf("%s is built for CUDA 13, which supports GPUs of compute capability %.1f and newer; this node has %.1f",
-			engine, minComputeForCUDA13, cc)
+	b := Build{
+		Variant:      "cu13",
+		Python:       pythonForEngines,
+		Requirements: []string{"sglang==" + version},
+		// SGLang is launched as a module, so what the agent is given is the
+		// environment's interpreter, and the proof is importing it.
+		Binary:      "python",
+		VersionArgs: []string{"-c", "import sglang; print('sglang', sglang.__version__)"},
+		Packages:    compilerPackages,
+		FreeGB:      25,
+		GPUCheck:    true,
 	}
-	return nil
+	needs := gpuNeeds{minDriverForCUDA13, 7.5}
+	if versionBefore(version, sglangCUDA13Since) {
+		b.Variant, needs = "cu12", gpuNeeds{minDriverForCUDA12, 7.0}
+	}
+	return pick("SGLang "+version, variant, []candidate{{b, needs, "SGLang " + version}}, h)
 }

 // Python engines need a C compiler at run time: Triton builds a small
@@ -150,35 +385,27 @@ func cudaGate(engine string, h discovery.Host) error {
 // that load fails long after the installation looked finished.
 var compilerPackages = map[string][]string{"deb": {"gcc"}, "rpm": {"gcc"}}

-func chooseVLLM(version string, h discovery.Host) (Build, error) {
-	if err := cudaGate("vLLM", h); err != nil {
-		return Build{}, err
+// versionBefore reports whether version a is older than b, comparing the
+// dotted numbers in turn ("0.25.1" before "0.26"). A part that is not a
+// number counts as zero.
+func versionBefore(a, b string) bool {
+	if b == "" {
+		return false
 	}
-	return Build{
-		Variant:      "cu13",
-		Python:       pythonForEngines,
-		Requirements: []string{"vllm==" + version},
-		Binary:       "vllm",
-		Packages:     compilerPackages,
-		FreeGB:       25,
-	}, nil
-}
-
-func chooseSGLang(version string, h discovery.Host) (Build, error) {
-	if err := cudaGate("SGLang", h); err != nil {
-		return Build{}, err
+	pa, pb := strings.Split(strings.TrimPrefix(a, "v"), "."), strings.Split(b, ".")
+	for i := 0; i < max(len(pa), len(pb)); i++ {
+		x, y := 0, 0
+		if i < len(pa) {
+			x = leadingInt(pa[i])
+		}
+		if i < len(pb) {
+			y = leadingInt(pb[i])
+		}
+		if x != y {
+			return x < y
+		}
 	}
-	// SGLang is launched as a module, so what the agent is given is the
-	// environment's interpreter, and the proof is importing it.
-	return Build{
-		Variant:      "cu13",
-		Python:       pythonForEngines,
-		Requirements: []string{"sglang==" + version},
-		Binary:       "python",
-		VersionArgs:  []string{"-c", "import sglang; print('sglang', sglang.__version__)"},
-		Packages:     compilerPackages,
-		FreeGB:       25,
-	}, nil
+	return false
 }

 // leadingInt is the number a version string starts with: 580 of "580.65.06".
@@ -193,17 +420,30 @@ func leadingInt(s string) int {
 	return n
 }

-// cudaForLlamaCpp is the CUDA build taken on an NVIDIA node. The release also
-// carries a matching cudart archive, so the node needs a driver and nothing
-// else — no toolkit.
-const cudaForLlamaCpp = "12.8"
+// --- llama.cpp ---

-func chooseLlamaCpp(version string, h discovery.Host) (Build, error) {
-	// The published Linux builds are built on Ubuntu, but nothing about them
-	// is Ubuntu's: they need glibc 2.34 and the libstdc++ of GCC 12, which
-	// Debian 12 and RHEL 10 have and RHEL 9 does not. So the distribution's
-	// name decides nothing here. The build is tried everywhere, and ldd says
-	// whether this system can load it — in the words of the missing symbol.
+// llamaCUDA12Real are the GPUs llama.cpp's CUDA 12.8 build carries machine
+// code for. The rest of what it covers — 5.0, 6.1, 7.0, 7.5, 8.0, 9.0 — it
+// carries as PTX, which a driver older than the CUDA it came from cannot
+// compile.
+var llamaCUDA12Real = []float64{8.6, 8.9, 12.0}
+
+// llamaCppCandidates are the llama.cpp builds for a machine.
+//
+// On an NVIDIA node, the GPU builds best first: CUDA 13.3 (driver 580,
+// Turing and newer), CUDA 12.8 on x64 (Maxwell to Blackwell; driver 570 for
+// the GPUs it ships as PTX, 525 for those it ships as machine code), and
+// Vulkan, which runs on any card whose driver has Vulkan — the generations
+// CUDA builds no longer cover. Never the CPU build on an NVIDIA node: that is
+// a GPU placement served on the CPU, which is what installing checks for.
+// The CPU build where there is no NVIDIA GPU.
+//
+// The published Linux builds are built on Ubuntu, but nothing about them is
+// Ubuntu's: they need glibc 2.34 and the libstdc++ of GCC 12, which Debian 12
+// and RHEL 10 have and RHEL 9 does not. So the distribution's name decides
+// nothing here. The build is tried everywhere, and ldd says whether this
+// system can load it — in the words of the missing symbol.
+func llamaCppCandidates(version, variant string, h discovery.Host) ([]Build, error) {
 	var arch string
 	switch h.Arch {
 	case "amd64":
@@ -211,35 +451,53 @@ func chooseLlamaCpp(version string, h discovery.Host) (Build, error) {
 	case "arm64":
 		arch = "arm64"
 	default:
-		return Build{}, fmt.Errorf("llama.cpp publishes no Linux build for %s", orUnknown(h.Arch))
+		return nil, fmt.Errorf("llama.cpp publishes no Linux build for %s", orUnknown(h.Arch))
 	}
-
-	b := Build{
-		Variant: "cpu",
-		Assets:  []string{fmt.Sprintf("llama-%s-bin-ubuntu-%s.tar.gz", version, arch)},
-		// Everything else the build links is in the archive or in any current
-		// glibc system; OpenMP's runtime is the one a cloud image lacks, and
-		// the server does not start without it.
-		Packages: map[string][]string{"deb": {"libgomp1"}, "rpm": {"libgomp"}},
-		Binary:   "llama-server",
-	}
-	if arch == "x64" {
-		// The archive carries a backend per CPU generation and loads the best
-		// one, down to SSE4.2. Below AVX2 it runs, several times slower.
-		b.CPUFlags = []string{"avx2"}
-	}
-	// The CUDA build is published for x64 at this version; on arm64 it is a
-	// newer CUDA that the driver may not support, so an arm64 GPU node runs
-	// the CPU build until someone decides otherwise.
-	if h.NVIDIADriver != "" && arch == "x64" {
-		name := fmt.Sprintf("bin-ubuntu-cuda-%s-x64.tar.gz", cudaForLlamaCpp)
-		b.Variant = "cuda-" + cudaForLlamaCpp
-		b.Assets = []string{
-			fmt.Sprintf("llama-%s-%s", version, name),
-			fmt.Sprintf("cudart-llama-%s-%s", version, name),
+	base := func(variant string, assets ...string) Build {
+		b := Build{
+			Variant: variant,
+			Assets:  assets,
+			// Everything else the build links is in the archive or in any
+			// current glibc system; OpenMP's runtime is the one a cloud image
+			// lacks, and the server does not start without it.
+			Packages: map[string][]string{"deb": {"libgomp1"}, "rpm": {"libgomp"}},
+			Binary:   "llama-server",
 		}
+		if arch == "x64" {
+			// The archive carries a backend per CPU generation and loads the
+			// best one, down to SSE4.2. Below AVX2 it runs, several times
+			// slower.
+			b.CPUFlags = []string{"avx2"}
+		}
+		return b
 	}
-	return b, nil
+	if h.NVIDIADriver == "" {
+		return pick("llama.cpp", variant, []candidate{
+			{base("cpu", fmt.Sprintf("llama-%s-bin-ubuntu-%s.tar.gz", version, arch)), gpuNeeds{}, "llama.cpp"},
+		}, h)
+	}
+	cuda := func(cv string) Build {
+		name := fmt.Sprintf("bin-ubuntu-cuda-%s-%s.tar.gz", cv, arch)
+		// The release carries the CUDA runtime beside the build, so the node
+		// needs a driver and nothing else — no toolkit.
+		b := base("cuda-"+cv, fmt.Sprintf("llama-%s-%s", version, name), fmt.Sprintf("cudart-llama-%s-%s", version, name))
+		b.GPUBackend = "libggml-cuda.so"
+		return b
+	}
+	vulkan := base("vulkan", fmt.Sprintf("llama-%s-bin-ubuntu-vulkan-%s.tar.gz", version, arch))
+	vulkan.GPUBackend = "libggml-vulkan.so"
+	vulkan.Packages = map[string][]string{"deb": {"libgomp1", "libvulkan1"}, "rpm": {"libgomp", "vulkan-loader"}}
+
+	cuda12Driver := 570
+	if slices.Contains(llamaCUDA12Real, h.MinComputeCapability) {
+		cuda12Driver = minDriverForCUDA12
+	}
+	cands := []candidate{{cuda("13.3"), gpuNeeds{minDriverForCUDA13, 7.5}, "llama.cpp built for CUDA 13.3"}}
+	if arch == "x64" {
+		cands = append(cands, candidate{cuda("12.8"), gpuNeeds{cuda12Driver, 5.0}, "llama.cpp built for CUDA 12.8"})
+	}
+	cands = append(cands, candidate{vulkan, gpuNeeds{}, "llama.cpp built for Vulkan"})
+	return pick("llama.cpp", variant, cands, h)
 }

 func orUnknown(s string) string {
@@ -2,6 +2,8 @@ package weights

 import (
 	"context"
+	"crypto/sha256"
+	"encoding/hex"
 	"encoding/json"
 	"fmt"
 	"io"
@@ -26,8 +28,9 @@ type manifest struct {
 	Model   string `json:"model"`
 	Version string `json:"version"`
 	Files   []struct {
-		Name  string `json:"name"`
-		Bytes int64  `json:"bytes"`
+		Name   string `json:"name"`
+		Bytes  int64  `json:"bytes"`
+		SHA256 string `json:"sha256"`
 	} `json:"files"`
 	Bytes int64 `json:"bytes"`
 }
@@ -87,6 +90,25 @@ func (f *Fetcher) ensureSet(ctx context.Context, source string) (string, error)
 		}
 	}

+	// Each file against the checksum the mirror took when it fetched it,
+	// before the set is called complete: a shard damaged on the way loads as
+	// a model that answers nonsense, not as an error.
+	for _, file := range m.Files {
+		if file.SHA256 == "" {
+			continue
+		}
+		target := filepath.Join(dir, file.Name)
+		got, err := fileSHA256(target)
+		if err != nil {
+			return "", err
+		}
+		if !strings.EqualFold(got, file.SHA256) {
+			_ = os.Remove(target)
+			return "", fmt.Errorf("%s%s: sha256 %s, the mirror has %s; removed, to be fetched again",
+				source, file.Name, got, file.SHA256)
+		}
+	}
+
 	if err := os.WriteFile(done, []byte(source+"\n"), 0o644); err != nil {
 		return "", err
 	}
@@ -108,6 +130,12 @@ func (f *Fetcher) manifestFor(ctx context.Context, source string) (manifest, err
 		return manifest{}, fmt.Errorf("asking %s what it holds: %w", source, err)
 	}
 	defer resp.Body.Close()
+	if resp.StatusCode == http.StatusConflict {
+		// Still downloading on the mirror, or a shard missing: said in the
+		// mirror's words, which name what is not there yet.
+		why, _ := io.ReadAll(io.LimitReader(resp.Body, 4<<10))
+		return manifest{}, fmt.Errorf("%s is not complete on the mirror yet: %s", source, strings.TrimSpace(string(why)))
+	}
 	if resp.StatusCode != http.StatusOK {
 		return manifest{}, fmt.Errorf("%s answered %d", source, resp.StatusCode)
 	}
@@ -127,3 +155,16 @@ func (f *Fetcher) manifestFor(ctx context.Context, source string) (manifest, err
 	}
 	return m, nil
 }
+
+func fileSHA256(path string) (string, error) {
+	f, err := os.Open(path) //nolint:gosec // a file this fetcher wrote
+	if err != nil {
+		return "", err
+	}
+	defer f.Close()
+	h := sha256.New()
+	if _, err := io.Copy(h, f); err != nil {
+		return "", err
+	}
+	return hex.EncodeToString(h.Sum(nil)), nil
+}
@@ -1,6 +1,8 @@
 package weights

 import (
+	"crypto/sha256"
+	"encoding/hex"
 	"fmt"
 	"net/http"
 	"net/http/httptest"
@@ -165,3 +167,40 @@ func TestOnlyATrailingSlashMeansASet(t *testing.T) {
 		}
 	}
 }
+
+func TestASetFileThatDoesNotMatchTheMirrorsChecksumIsRefused(t *testing.T) {
+	good := "the shard as mirrored"
+	sum := sha256.Sum256([]byte(good))
+	served := "the shard damaged on the way"
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if strings.HasSuffix(r.URL.Path, "/") {
+			fmt.Fprintf(w, `{"model":"m","version":"v","files":[{"name":"s.safetensors","bytes":%d,"sha256":%q}]}`,
+				len(served), hex.EncodeToString(sum[:]))
+			return
+		}
+		_, _ = w.Write([]byte(served))
+	}))
+	t.Cleanup(srv.Close)
+	f := &Fetcher{CacheDir: t.TempDir()}
+	if _, err := f.Ensure(t.Context(), srv.URL+"/m/v/"); err == nil || !strings.Contains(err.Error(), "sha256") {
+		t.Fatalf("a damaged shard was accepted: %v", err)
+	}
+	served = good
+	// The size differs, so it is fetched again; this time it matches.
+	if _, err := f.Ensure(t.Context(), srv.URL+"/m/v/"); err != nil {
+		t.Fatalf("the shard as mirrored: %v", err)
+	}
+}
+
+func TestASetStillDownloadingOnTheMirrorSaysSo(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		http.Error(w, "m/v: this version is not complete: still downloading model-00002-of-00002.safetensors", http.StatusConflict)
+	}))
+	t.Cleanup(srv.Close)
+	f := &Fetcher{CacheDir: t.TempDir()}
+	_, err := f.Ensure(t.Context(), srv.URL+"/m/v/")
+	if err == nil || !strings.Contains(err.Error(), "not complete on the mirror yet") ||
+		!strings.Contains(err.Error(), "model-00002-of-00002") {
+		t.Fatalf("got %v", err)
+	}
+}
@@ -143,6 +143,7 @@ func newCtlCmd() *cobra.Command {
 		newRetryCmd(client),
 		newUpgradeCmd(client),
 		newEngineCheckCmd(),
+		newEngineBuildCmd(client),
 		newToolCheckCmd(),
 		newWeightsCmd(client, &asJSON),
 		newCapacityCmd(client, &asJSON),
@@ -254,6 +255,8 @@ func modelStatus(m adminapi.ModelView) string {
 	switch {
 	case m.Parked:
 		return "parked"
+	case m.External != "":
+		return "external: " + m.External
 	case m.HeldFor != "":
 		return "gave room to " + m.HeldFor
 	case m.Unplaceable:
@@ -12,8 +12,9 @@ import (

 func newRestartCmd(client func() *ctlClient) *cobra.Command {
 	var (
-		models []string
-		all    bool
+		models  []string
+		all     bool
+		inPlace bool
 	)
 	c := &cobra.Command{
 		Use:   "restart [REPLICA...]",
@@ -22,10 +23,14 @@ func newRestartCmd(client func() *ctlClient) *cobra.Command {
 			"an edited profile, a new engine version.\n\n" +
 			"Each replacement is placed beside the replica it replaces, and the old one\n" +
 			"keeps serving until the new one is ready. Replicas of one model are\n" +
-			"replaced one at a time. Where nothing else has room for a second copy —\n" +
-			"the usual case on a small CPU fleet — the old one is stopped after the\n" +
-			"drain grace period (5 minutes by default) and the new one starts in its\n" +
-			"place; a model with only that replica is unserved while it loads.\n\n" +
+			"replaced one at a time.\n\n" +
+			"Where there is no room for a second copy and the old replica is all that\n" +
+			"is in the way — one GPU, every card that fits held — it is stopped at once\n" +
+			"and the new one starts in its place; a model with only that replica is\n" +
+			"unserved while it loads. Where the new one would not fit even then — a\n" +
+			"profile edited to need more than any node has — the restart is given up\n" +
+			"and the old replica keeps serving: llmgw ctl replicas says so, and why.\n\n" +
+			"--in-place stops each replica first, whatever room there is.\n\n" +
 			"  llmgw ctl restart qwen2.5-7b-3f2a91c0\n" +
 			"  llmgw ctl restart --model qwen2.5-7b\n" +
 			"  llmgw ctl restart --all",
@@ -33,7 +38,7 @@ func newRestartCmd(client func() *ctlClient) *cobra.Command {
 			if len(args) == 0 && len(models) == 0 && !all {
 				return errors.New("name replicas, --model, or --all")
 			}
-			body, _ := json.Marshal(adminapi.RestartRequest{Replicas: args, Models: models, All: all})
+			body, _ := json.Marshal(adminapi.RestartRequest{Replicas: args, Models: models, All: all, InPlace: inPlace})
 			var res adminapi.RestartResult
 			if err := client().post(cmd.Context(), "/api/v1/restart", body, &res); err != nil {
 				return err
@@ -44,6 +49,7 @@ func newRestartCmd(client func() *ctlClient) *cobra.Command {
 	}
 	c.Flags().StringArrayVar(&models, "model", nil, "every replica of this model; repeatable")
 	c.Flags().BoolVar(&all, "all", false, "every ready replica in the cluster")
+	c.Flags().BoolVar(&inPlace, "in-place", false, "stop each replica before starting its replacement")
 	return c
 }

@@ -23,6 +23,7 @@ func newDoctorCmd() *cobra.Command {
 		opts    doctor.Options
 		asJSON  bool
 		verbose bool
+		fabrics []string
 	)
 	cmd := &cobra.Command{
 		Use:   "doctor",
@@ -38,8 +39,21 @@ that file when the report is not enough to go on.`,
 		Args: cobra.NoArgs,
 		RunE: func(cmd *cobra.Command, _ []string) error {
 			doctor.SetConfigLoader(config.LoadAgent)
-			if opts.VLLMVersion == "" {
+			for _, spec := range fabrics {
+				name, hcas, _ := strings.Cut(spec, ":")
+				if name == "" || hcas == "" {
+					return fmt.Errorf("--fabric %q: give it as name:mlx5_0,mlx5_1", spec)
+				}
+				if opts.Fabrics == nil {
+					opts.Fabrics = map[string][]string{}
+				}
+				opts.Fabrics[name] = strings.Split(hcas, ",")
+			}
+			switch opts.VLLMVersion {
+			case "":
 				opts.VLLMVersion = config.DefaultEngines["vllm"]
+			case "none":
+				opts.VLLMVersion = ""
 			}
 			ctx, cancel := context.WithTimeout(cmd.Context(), 60*time.Second)
 			defer cancel()
@@ -66,7 +80,8 @@ that file when the report is not enough to go on.`,
 	f.StringVar(&opts.ConfigPath, "config", "", "agent configuration (default /etc/llmgw/agent.yaml when present)")
 	f.StringVar(&opts.IBRoot, "ib-root", discovery.SysfsRoot, "sysfs InfiniBand directory")
 	f.StringVar(&opts.NvidiaSMI, "nvidia-smi", "", "path to nvidia-smi")
-	f.StringVar(&opts.VLLMVersion, "vllm", "", "the vLLM version to check for (default: the one this llmgw installs)")
+	f.StringVar(&opts.VLLMVersion, "vllm", "", "the vLLM version to check for (default: the one this llmgw installs; none to skip)")
+	f.StringArrayVar(&fabrics, "fabric", nil, "a fabric about to be declared, name:mlx5_0,mlx5_1, checked as if it were; repeatable")
 	f.BoolVar(&asJSON, "json", false, "write every fact and finding as JSON, to send back")
 	f.BoolVar(&verbose, "verbose", false, "also print the raw nvidia-smi topology")
 	return cmd
@@ -0,0 +1,39 @@
+package cli
+
+import (
+	"encoding/json"
+	"fmt"
+
+	"github.com/spf13/cobra"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/adminapi"
+)
+
+func newEngineBuildCmd(client func() *ctlClient) *cobra.Command {
+	return &cobra.Command{
+		Use:   "engine-build <node> <engine> <build>",
+		Short: "Choose which build of an engine a GPU node installs",
+		Long: "A node installs the build of vLLM its hardware runs: CUDA 13 where the\n" +
+			"driver and GPUs allow, else the same version built for CUDA 12.9. This\n" +
+			"chooses instead, as the cluster page's vLLM build column does:\n\n" +
+			"  auto    what the hardware runs (the default)\n" +
+			"  cu13    the CUDA 13 build, from PyPI\n" +
+			"  cu129   the CUDA 12.9 build, from vLLM's GitHub release\n" +
+			"  none    no vLLM on this node\n\n" +
+			"The node is told at once and installs it in the background; a build it\n" +
+			"cannot compute with on its GPU fails to install, saying why, and the\n" +
+			"cluster page shows it. A V100 runs no build of vLLM 0.20 or later, and\n" +
+			"gets 0.19.1, the newest built for it, by itself.\n\n" +
+			"  llmgw ctl engine-build gpu07 vllm cu129",
+		Args: cobra.ExactArgs(3),
+		RunE: func(cmd *cobra.Command, args []string) error {
+			body, _ := json.Marshal(adminapi.NodeEngineRequest{Engine: args[1], Build: args[2]})
+			if err := client().post(cmd.Context(), "/api/v1/nodes/"+args[0]+"/engines", body, nil); err != nil {
+				return err
+			}
+			fmt.Fprintf(cmd.OutOrStdout(), "%s: %s build %s; the node installs it in the background (llmgw ctl nodes shows it)\n",
+				args[0], args[1], args[2])
+			return nil
+		},
+	}
+}
@@ -3,6 +3,7 @@ package cli
 import (
 	"context"
 	"fmt"
+	"io"
 	"os/exec"
 	"path/filepath"
 	"sort"
@@ -126,6 +127,14 @@ func newEngineCheckCmd() *cobra.Command {
 			}
 			sort.Strings(unknown)

+			// Flags accepted is not the same as running where it is placed: a
+			// GPU build whose backend does not load accepts every flag and
+			// computes on the CPU. Asked of the engine here, not of a help.
+			var deviceErr error
+			if fb, ok := eng.(engines.CPUFallback); ok && help == "" {
+				deviceErr = checkDevices(cmd.Context(), name, fb, out)
+			}
+
 			fmt.Fprintf(out, "%s: %d flags checked", name, checked)
 			if skipped > 0 {
 				fmt.Fprintf(out, ", %d commands skipped", skipped)
@@ -137,7 +146,7 @@ func newEngineCheckCmd() *cobra.Command {
 					return fmt.Errorf("nothing was checked; install the engine here or pass --help-output")
 				}
 				fmt.Fprintln(out, "every flag llmgw builds is accepted")
-				return nil
+				return deviceErr
 			}
 			fmt.Fprintf(out, "\n%d flags not mentioned by the program they are passed to:\n", len(unknown))
 			for _, f := range unknown {
@@ -249,3 +258,22 @@ func isFlagChar(b byte) bool {
 	return b == '-' || b == '_' || (b >= 'a' && b <= 'z') ||
 		(b >= 'A' && b <= 'Z') || (b >= '0' && b <= '9')
 }
+
+// checkDevices says which GPUs an engine that can fall back to the CPU sees on
+// this machine, and fails when the machine has NVIDIA GPUs and it sees none.
+func checkDevices(ctx context.Context, name string, fb engines.CPUFallback, out io.Writer) error {
+	if _, err := exec.LookPath("nvidia-smi"); err != nil {
+		fmt.Fprintln(out, "no NVIDIA GPU here: which devices it computes on was not checked")
+		return nil
+	}
+	devs, err := fb.Devices(ctx)
+	if err != nil {
+		return err
+	}
+	if len(devs) == 0 {
+		return fmt.Errorf("this machine has NVIDIA GPUs and %s sees none: it would serve a GPU placement on the CPU. "+
+			"Its GPU backend did not load; ldd on its libggml-cuda.so names the missing library", name)
+	}
+	fmt.Fprintf(out, "computes on: %s\n", strings.Join(devs, "; "))
+	return nil
+}
@@ -2,6 +2,7 @@ package cli

 import (
 	"encoding/json"
+	"errors"
 	"fmt"
 	"log/slog"
 	"net"
@@ -56,52 +57,69 @@ func newMirrorAddCmd(dir *string) *cobra.Command {
 	var model, version, as string

 	c := &cobra.Command{
-		Use:   "add <url>",
+		Use:   "add <url>...",
 		Short: "Download a model into the mirror",
 		Long: "Downloads to a temporary name and renames when complete, so an\n" +
 			"interrupted download never leaves something a catalog could be pointed\n" +
 			"at. Reports the checksum: a mirror is the one place in this system where\n" +
 			"a corrupted file is copied to every node rather than caught once.\n\n" +
+			"Several URLs go into one version, as the files of one model: the shards\n" +
+			"of a safetensors set with its config.json and tokenizer, for vLLM. The\n" +
+			"version is offered to nodes only once it is whole — nothing still\n" +
+			"downloading, every shard its index names present — and its source is\n" +
+			"the version, ending in a slash.\n\n" +
 			"Refuses to replace a version that is already here. Add a new one beside\n" +
 			"it instead — that is what makes the old one still available when the new\n" +
 			"one turns out to be worse.",
-		Args: cobra.ExactArgs(1),
+		Args: cobra.MinimumNArgs(1),
 		RunE: func(cmd *cobra.Command, args []string) error {
-			source := args[0]
-			file := as
-			if file == "" {
-				file = mirror.LastPathElement(source)
-			}
 			if model == "" || version == "" {
 				return fmt.Errorf("--model and --version are required: they are the layout")
 			}
-
-			e := mirror.Entry{Model: model, Version: version, File: file}
-			out := cmd.OutOrStdout()
-			fmt.Fprintf(out, "fetching %s\n", source)
-
-			added, err := mirror.Add(cmd.Context(), *dir, source, e, func(done, total int64) {
-				if total > 0 {
-					fmt.Fprintf(out, "  %s of %s (%d%%)\n",
-						fileSize(done), fileSize(total), done*100/total)
-					return
-				}
-				fmt.Fprintf(out, "  %s so far\n", fileSize(done))
-			})
-			if err != nil {
-				return err
+			if as != "" && len(args) > 1 {
+				return fmt.Errorf("--as names one file; with %d URLs each keeps its own name", len(args))
 			}
-			fmt.Fprintf(out, "\n%s/%s/%s  %s\n  sha256 %s\n",
-				added.Entry.Model, added.Entry.Version, added.Entry.File,
-				fileSize(added.Entry.Bytes), added.SHA256)
-			fmt.Fprintf(out, "\nPoint a model at it with:\n  source: %s\n",
-				added.Entry.URL("http://<this-host>:8081"))
+			out := cmd.OutOrStdout()
+			var last mirror.Added
+			for _, source := range args {
+				file := as
+				if file == "" {
+					file = mirror.LastPathElement(source)
+				}
+				e := mirror.Entry{Model: model, Version: version, File: file}
+				fmt.Fprintf(out, "fetching %s\n", source)
+				added, err := mirror.Add(cmd.Context(), *dir, source, e, func(done, total int64) {
+					if total > 0 {
+						fmt.Fprintf(out, "  %s of %s (%d%%)\n",
+							fileSize(done), fileSize(total), done*100/total)
+						return
+					}
+					fmt.Fprintf(out, "  %s so far\n", fileSize(done))
+				})
+				if err != nil {
+					return err
+				}
+				fmt.Fprintf(out, "%s/%s/%s  %s\n  sha256 %s\n",
+					added.Entry.Model, added.Entry.Version, added.Entry.File,
+					fileSize(added.Entry.Bytes), added.SHA256)
+				last = added
+			}
+			src := last.Entry.URL("http://<this-host>:8081")
+			if len(args) > 1 {
+				src = strings.TrimSuffix(mirror.Entry{Model: model, Version: version}.URL("http://<this-host>:8081"), "/") + "/"
+				if m, err := mirror.ManifestFor(*dir, model, version); err != nil {
+					fmt.Fprintf(out, "\nNot yet offered to nodes: %v\n", err)
+				} else {
+					fmt.Fprintf(out, "\n%d files, %s, whole.\n", len(m.Files), fileSize(m.Bytes))
+				}
+			}
+			fmt.Fprintf(out, "\nPoint a model at it with:\n  source: %s\n", src)
 			return nil
 		},
 	}
 	c.Flags().StringVar(&model, "model", "", "the model's name, one directory level")
 	c.Flags().StringVar(&version, "version", "", "the version, one directory level; a date works well")
-	c.Flags().StringVar(&as, "as", "", "the filename to save as; taken from the URL when omitted")
+	c.Flags().StringVar(&as, "as", "", "the filename to save as, for one URL; taken from the URL when omitted")
 	return c
 }

@@ -405,7 +423,14 @@ func withManifest(root string, next http.Handler) http.Handler {
 			return
 		}
 		m, err := mirror.ManifestFor(root, model, version)
-		if err != nil {
+		switch {
+		case errors.Is(err, mirror.ErrIncomplete):
+			// Said, not hidden behind a 404: a node asking now is told why
+			// it cannot have the version yet, and tries again later.
+			w.Header().Set("Retry-After", "60")
+			http.Error(w, err.Error(), http.StatusConflict)
+			return
+		case err != nil:
 			http.NotFound(w, r)
 			return
 		}
@@ -5,6 +5,9 @@ import (
 	"fmt"
 	"os"
 	"os/user"
+	"slices"
+	"strconv"
+	"strings"
 	"time"

 	"github.com/knadh/koanf/parsers/yaml"
@@ -31,7 +34,7 @@ func newModelCmd(client func() *ctlClient, asJSON *bool) *cobra.Command {
 	}
 	c.AddCommand(newModelAddCmd(client), newModelRemoveCmd(client), newModelPlanCmd(client),
 		newResourcesCmd(client), newClassCmd(client), newOwnerCmd(client),
-		newTargetCmd(client, asJSON),
+		newTargetCmd(client, asJSON), newAdaptersCmd(client),
 		newModelServeCmd(client))
 	return c
 }
@@ -264,3 +267,69 @@ func newClassCmd(client func() *ctlClient) *cobra.Command {
 		},
 	}
 }
+
+func newAdaptersCmd(client func() *ctlClient) *cobra.Command {
+	var add, remove []string
+	c := &cobra.Command{
+		Use:   "adapters <model> [--add NAME=URL[@SCALE]] [--remove NAME]",
+		Short: "List, add or remove a model's LoRA adapters",
+		Long: "An adapter is asked for by its NAME as though it were a model, and is served\n" +
+			"by this model's replicas with it applied. URL is the adapter file, usually in\n" +
+			"the mirror; it must be a GGUF LoRA adapter for this model's architecture,\n" +
+			"which is checked before it is attached. Replicas load a changed set when\n" +
+			"restarted: llmgw ctl restart --model MODEL.",
+		Args: cobra.ExactArgs(1),
+		RunE: func(cmd *cobra.Command, args []string) error {
+			name := args[0]
+			var models []adminapi.ModelView
+			if err := client().get(cmd.Context(), "/api/v1/models", &models); err != nil {
+				return err
+			}
+			var current []catalog.Adapter
+			found := false
+			for _, m := range models {
+				if m.Name == name {
+					current, found = m.Adapters, true
+				}
+			}
+			if !found {
+				return fmt.Errorf("no model %q", name)
+			}
+			out := cmd.OutOrStdout()
+			if len(add) == 0 && len(remove) == 0 {
+				if len(current) == 0 {
+					fmt.Fprintf(out, "%s has no adapters\n", name)
+				}
+				for _, a := range current {
+					fmt.Fprintf(out, "%s\tscale %g\t%s\n", a.Name, a.EffectiveScale(), a.Source)
+				}
+				return nil
+			}
+			next := slices.DeleteFunc(slices.Clone(current), func(a catalog.Adapter) bool { return slices.Contains(remove, a.Name) })
+			for _, spec := range add {
+				n, rest, ok := strings.Cut(spec, "=")
+				if !ok || n == "" || rest == "" {
+					return fmt.Errorf("--add %q is not NAME=URL[@SCALE]", spec)
+				}
+				ad := catalog.Adapter{Name: n, Source: rest}
+				if i := strings.LastIndex(rest, "@"); i > 0 {
+					if s, err := strconv.ParseFloat(rest[i+1:], 64); err == nil {
+						ad.Source, ad.Scale = rest[:i], s
+					}
+				}
+				next = append(next, ad)
+			}
+			body, _ := json.Marshal(adminapi.AdaptersRequest{Adapters: next})
+			var res adminapi.AdaptersRequest
+			if err := client().post(cmd.Context(), "/api/v1/models/"+name+"/adapters", body, &res); err != nil {
+				return err
+			}
+			fmt.Fprintf(out, "%s serves %d adapter(s); restart its replicas to load them: llmgw ctl restart --model %s\n",
+				name, len(res.Adapters), name)
+			return nil
+		},
+	}
+	c.Flags().StringArrayVar(&add, "add", nil, "NAME=URL[@SCALE]; repeatable")
+	c.Flags().StringArrayVar(&remove, "remove", nil, "an adapter's name; repeatable")
+	return c
+}
@@ -12,6 +12,7 @@ import (

 	"github.com/spf13/cobra"

+	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/provision"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/config"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/control/adminapi"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/onboard"
@@ -113,13 +114,14 @@ engines are kept.`,
 	f.StringVar(&na.opts.MgmtAddr, "mgmt-addr", "", "address engines are reached at (default: found by the agent)")
 	f.StringVar(&na.opts.MgmtIface, "mgmt-if", "", "interface NCCL bootstraps over (default: the one holding that address)")
 	f.StringVar(&na.opts.LlamaCppBin, "llamacpp-bin", "", "a llama-server you installed, instead of the one the agent installs")
-	f.StringVar(&na.opts.VLLMBin, "vllm-bin", "", "a vllm you installed; also set engines.vllm: \"\" in server.yaml")
+	f.StringVar(&na.opts.VLLMBin, "vllm-bin", "", "a vllm you installed, instead of the one the agent installs")
 	f.StringVar(&na.opts.SGLangPython, "sglang-python", "", "python of an SGLang virtualenv you installed")
 	f.StringVar(&na.opts.NvidiaSMI, "nvidia-smi", "", "where nvidia-smi is, when it is not on PATH")
 	f.StringVar(&na.opts.IBRoot, "ib-root", "", "the sysfs InfiniBand directory, when not /sys/class/infiniband")
 	f.StringVar(&na.opts.PKI, "pki", defaultPKIDir, "directory holding the certificate authority")
 	f.StringVar(&na.opts.Releases, "releases", "https://gitea.mynodes.xyz/ilker/llmgw/releases/download", "where release files are downloaded from")
 	f.BoolVar(&na.opts.IgnoreChecks, "ignore-checks", false, "add the node even when doctor finds a problem")
+	f.StringArrayVar(&na.opts.DisableEngines, "disable-engine", nil, "an engine this node never installs, e.g. vllm on GPUs it does not run on; repeatable")
 	f.StringVar(&na.admin, "admin", "", "the control plane's admin API, to watch the node join (default: from server.yaml, else 127.0.0.1:9090)")
 	f.StringVar(&na.serverCfg, "config", "/etc/llmgw/server.yaml", "the control plane's configuration, for the admin API's address")
 	f.BoolVar(&na.waitAll, "wait", false, "wait for every engine, vLLM included, not only llama.cpp")
@@ -148,6 +150,23 @@ func (na *nodeAdd) prepare() error {
 			}
 		}
 	}
+	for _, e := range o.DisableEngines {
+		if !slices.Contains(provision.Installable(), e) {
+			return fmt.Errorf("--disable-engine %s: the engines are %s", e, strings.Join(provision.Installable(), ", "))
+		}
+	}
+	// What the control plane will ask the node to install, so doctor checks
+	// that vLLM rather than its own default. Its overrides for a GPU class
+	// are not known until the node says what it has, and are not applied
+	// here; doctor's verdict on vLLM is a warning either way.
+	if o.VLLMVersion == "" {
+		if cfg, err := config.Load(na.serverCfg); err == nil {
+			o.VLLMVersion = cfg.WantedEngines()["vllm"]
+			if o.VLLMVersion == "" {
+				o.VLLMVersion = "none"
+			}
+		}
+	}
 	o.Version = version.Version
 	if o.Binary == "" {
 		exe, err := os.Executable()
@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"net/url"
 	"strconv"
+	"strings"

 	"github.com/spf13/cobra"

@@ -192,6 +193,9 @@ func newTargetCmd(client func() *ctlClient, asJSON *bool) *cobra.Command {
 			switch {
 			case v.More == 0:
 				fmt.Fprintln(out, "It already has them.")
+			case v.Fits && len(v.Preempts) > 0:
+				fmt.Fprintf(out, "%d more; there is room once replicas of %s are stopped, which its class may do.\n",
+					v.More, strings.Join(v.Preempts, ", "))
 			case v.Fits:
 				fmt.Fprintf(out, "%d more, and there is room for them.\n", v.More)
 			default:
@@ -0,0 +1,67 @@
+package cli
+
+import (
+	"errors"
+	"fmt"
+	"os"
+
+	"github.com/spf13/cobra"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/release"
+)
+
+func newReleaseCmd() *cobra.Command {
+	c := &cobra.Command{
+		Use:   "release",
+		Short: "Sign releases, and make the key they are signed with",
+	}
+	c.AddCommand(&cobra.Command{
+		Use:   "keygen",
+		Short: "Make a release signing key pair",
+		Long: "Prints a public key and a private key. The private key goes into the CI\n" +
+			"secret RELEASE_SIGNING_KEY and nowhere else; the public key into\n" +
+			"internal/release/trusted.pub, which every build from then on carries\n" +
+			"and checks releases against. Run once, on a machine you trust.",
+		Args: cobra.NoArgs,
+		RunE: func(cmd *cobra.Command, _ []string) error {
+			pub, priv, err := release.GenerateKey()
+			if err != nil {
+				return err
+			}
+			fmt.Fprintf(cmd.OutOrStdout(), "public:  %s\nprivate: %s\n", pub, priv)
+			fmt.Fprintln(cmd.ErrOrStderr(), "\nThe private key signs every release: put it in the CI secret "+
+				"RELEASE_SIGNING_KEY, then clear this terminal's scrollback.")
+			return nil
+		},
+	})
+	c.AddCommand(&cobra.Command{
+		Use:   "sign FILE",
+		Short: "Sign a release's checksums, writing FILE.sig",
+		Long: "Reads the private key from RELEASE_SIGNING_KEY, never from an argument,\n" +
+			"so it does not land in a shell history or a process list.",
+		Args: cobra.ExactArgs(1),
+		RunE: func(cmd *cobra.Command, args []string) error {
+			key := os.Getenv("RELEASE_SIGNING_KEY")
+			if key == "" {
+				return errors.New("RELEASE_SIGNING_KEY is not set")
+			}
+			data, err := os.ReadFile(args[0])
+			if err != nil {
+				return err
+			}
+			sig, err := release.Sign(key, data)
+			if err != nil {
+				return err
+			}
+			if pub := release.TrustedKey(); pub != "" {
+				// Signed with a key this build would not accept is a release
+				// every upgraded machine refuses: stopped here instead.
+				if err := release.Verify(pub, data, sig); err != nil {
+					return fmt.Errorf("the signing key does not match this build's trusted key: %w", err)
+				}
+			}
+			return os.WriteFile(args[0]+".sig", []byte(sig+"\n"), 0o644) //nolint:gosec // public
+		},
+	})
+	return c
+}
@@ -48,6 +48,7 @@ func Root() *cobra.Command {
 		newDoctorCmd(),
 		newNodeCmd(),
 		newSelfUpgradeCmd(),
+		newReleaseCmd(),
 	)
 	return root
 }
@@ -5,19 +5,23 @@ import (
 	"crypto/tls"
 	"errors"
 	"fmt"
-	"gitea.mynodes.xyz/ilker/llmgw/internal/control/modelserve"
-	"gitea.mynodes.xyz/ilker/llmgw/internal/gguf"
-	"gitea.mynodes.xyz/ilker/llmgw/internal/sourcecheck"
-	"github.com/prometheus/client_golang/prometheus"
 	"log/slog"
+	"maps"
 	"net"
 	"net/http"
 	"os/signal"
+	"slices"
 	"strings"
 	"sync"
 	"syscall"
 	"time"

+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/inventory"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/modelserve"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/gguf"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/sourcecheck"
+	"github.com/prometheus/client_golang/prometheus"
+
 	"github.com/spf13/cobra"
 	"golang.org/x/net/http2"
 	"golang.org/x/net/http2/h2c"
@@ -491,13 +495,21 @@ func startControlPlane(ctx context.Context, cfg config.Server, cat *catalog.Live
 			// And what each model was measured to serve, for scaling and for
 			// choosing between profiles.
 			ctl.Reconciler().SetMeasurements(control.Measurements(st, log))
+			// The engine builds chosen per node on the cluster page.
+			ctl.SetEngineChoices(control.NewEngineChoices(ctx, st, log))
 			log.Info("competing to be the control plane that places work",
 				"owner", lead.Owner())
 		}
 	}
 	svc := agentsvc.New(ctl.Inventory(), ctl.Reconciler(), requireIdentity, log)
-	svc.SetEngines(cfg.WantedEngines())
-	log.Info("nodes install these engines", "engines", cfg.WantedEngines())
+	svc.SetEnginesFor(func(n inventory.Node) map[string]string {
+		return cfg.EnginesFor(n.ID, slices.Sorted(maps.Keys(n.GPUsByClass())), n.Labels)
+	})
+	log.Info("nodes install these engines", "engines", cfg.WantedEngines(), "overrides", len(cfg.EngineOverrides))
+	if ch := ctl.EngineChoices(); ch != nil {
+		svc.SetEngineChoices(ch.For)
+		ch.OnChange(svc.Reprovision)
+	}
 	// A standby must refuse agents as well as refusing to place. Commands
 	// travel down the stream the agent opened, so a standby holding one is a
 	// node the leader cannot command — and the leader's placements would fail
@@ -8,6 +8,7 @@ package config
 import (
 	"fmt"
 	"os"
+	"slices"
 	"strings"
 	"time"

@@ -85,6 +86,13 @@ type Server struct {
 	// one leaves that engine to be installed by hand.
 	Engines map[string]string `koanf:"engines"`

+	// EngineOverrides give some nodes different engines from the rest: a
+	// vLLM built for CUDA 12 on the Volta cards CUDA 13 dropped, or none at
+	// all on nodes meant only for llama.cpp. Each picks nodes and lays its
+	// engines over the cluster's, in order, later ones last; an empty version
+	// removes that engine from the nodes it picks.
+	EngineOverrides []EngineOverride `koanf:"engine_overrides"`
+
 	// BackupDir is where `llmgw server backup` writes, and where the newest
 	// archive's age is read for llmgw_backup_last_success_timestamp_seconds.
 	// Empty means /var/backups/llmgw-server.
@@ -378,7 +386,12 @@ type Admission struct {

 	QueueTTL             time.Duration `koanf:"queue_ttl"`
 	QueueCapacityPerBand int           `koanf:"queue_capacity_per_band"`
-	Interval             time.Duration `koanf:"interval"`
+	// QueueCapacityTotal bounds everything queued across bands. Reached, a
+	// request is not refused while a lower band has any queued: the newest of
+	// the lowest band is turned away instead, with a 503 and Retry-After.
+	// Zero takes half of all the bands' capacities together.
+	QueueCapacityTotal int           `koanf:"queue_capacity_total"`
+	Interval           time.Duration `koanf:"interval"`
 }

 // Prefix configures approximate prefix routing.
@@ -543,6 +556,14 @@ func Load(path string) (Server, error) {
 // Validate rejects configurations that would otherwise fail confusingly at
 // request time.
 func (s Server) Validate() error {
+	for i, o := range s.EngineOverrides {
+		if o.GPUClass == "" && len(o.Nodes) == 0 && len(o.Labels) == 0 {
+			return fmt.Errorf("engine_overrides[%d] picks no nodes: give it a gpu_class, nodes or labels", i)
+		}
+		if len(o.Engines) == 0 {
+			return fmt.Errorf("engine_overrides[%d] changes no engine: give it engines, with \"\" for none", i)
+		}
+	}
 	if s.Listen == "" {
 		return fmt.Errorf("listen is empty")
 	}
@@ -621,6 +642,58 @@ var DefaultEngines = map[string]string{
 	"vllm":     "0.30.0",
 }

+// EngineOverride gives the nodes it picks different engines. Every selector
+// set must match; at least one must be set.
+type EngineOverride struct {
+	// GPUClass picks nodes with a GPU of this class ("v100-32g").
+	GPUClass string `koanf:"gpu_class"`
+	// Nodes picks nodes by id.
+	Nodes []string `koanf:"nodes"`
+	// Labels picks nodes carrying all of these.
+	Labels map[string]string `koanf:"labels"`
+	// Engines are laid over the cluster's: a version, or "" for none.
+	Engines map[string]string `koanf:"engines"`
+}
+
+// Picks reports whether the override applies to a node with this id, these
+// GPU classes and these labels.
+func (o EngineOverride) Picks(node string, classes []string, labels map[string]string) bool {
+	if o.GPUClass == "" && len(o.Nodes) == 0 && len(o.Labels) == 0 {
+		return false
+	}
+	if o.GPUClass != "" && !slices.Contains(classes, o.GPUClass) {
+		return false
+	}
+	if len(o.Nodes) > 0 && !slices.Contains(o.Nodes, node) {
+		return false
+	}
+	for k, v := range o.Labels {
+		if labels[k] != v {
+			return false
+		}
+	}
+	return true
+}
+
+// EnginesFor is what one node installs: WantedEngines with every override
+// that picks it laid over, in order.
+func (s Server) EnginesFor(node string, classes []string, labels map[string]string) map[string]string {
+	out := s.WantedEngines()
+	for _, o := range s.EngineOverrides {
+		if !o.Picks(node, classes, labels) {
+			continue
+		}
+		for k, v := range o.Engines {
+			if v == "" {
+				delete(out, k)
+				continue
+			}
+			out[k] = v
+		}
+	}
+	return out
+}
+
 // WantedEngines is DefaultEngines with the configuration laid over it.
 func (s Server) WantedEngines() map[string]string {
 	out := map[string]string{}
@@ -0,0 +1,38 @@
+package config
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestEachNodeGetsTheEnginesItsOverridesSay(t *testing.T) {
+	s := Server{
+		Engines: map[string]string{"vllm": "0.30.0"},
+		EngineOverrides: []EngineOverride{
+			// Volta: CUDA 13 dropped it, CUDA 12 vLLM still runs.
+			{GPUClass: "v100-32g", Engines: map[string]string{"vllm": "0.25.1"}},
+			// Nodes meant for llama.cpp alone.
+			{Labels: map[string]string{"role": "gguf"}, Engines: map[string]string{"vllm": ""}},
+		},
+	}
+	if got := s.EnginesFor("h1", []string{"h100-80g"}, nil); got["vllm"] != "0.30.0" || got["llamacpp"] == "" {
+		t.Errorf("an H100 node: %v", got)
+	}
+	if got := s.EnginesFor("v1", []string{"v100-32g"}, nil); got["vllm"] != "0.25.1" {
+		t.Errorf("a V100 node: %v", got)
+	}
+	got := s.EnginesFor("v2", []string{"v100-32g"}, map[string]string{"role": "gguf"})
+	if _, has := got["vllm"]; has || got["llamacpp"] == "" {
+		t.Errorf("a V100 node for llama.cpp alone, the later override last: %v", got)
+	}
+
+	for _, bad := range []EngineOverride{
+		{Engines: map[string]string{"vllm": ""}},
+		{GPUClass: "v100-32g"},
+	} {
+		s := Server{Listen: ":8080", EngineOverrides: []EngineOverride{bad}}
+		if err := s.Validate(); err == nil || !strings.Contains(err.Error(), "engine_overrides[0]") {
+			t.Errorf("%+v: %v", bad, err)
+		}
+	}
+}
@@ -0,0 +1,71 @@
+package adminapi
+
+import (
+	"bytes"
+	"encoding/binary"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/catalog"
+)
+
+// ggufStart is the start of a GGUF file with these string keys.
+func ggufStart(kv ...string) []byte {
+	var b bytes.Buffer
+	w := func(v any) { _ = binary.Write(&b, binary.LittleEndian, v) }
+	str := func(s string) { w(uint64(len(s))); b.WriteString(s) }
+	b.WriteString("GGUF")
+	w(uint32(3))
+	w(uint64(0))
+	w(uint64(len(kv) / 2))
+	for i := 0; i < len(kv); i += 2 {
+		str(kv[i])
+		w(uint32(8)) // string
+		str(kv[i+1])
+	}
+	b.Write(make([]byte, 4096))
+	return b.Bytes()
+}
+
+func TestOnlyAnAdapterForTheModelsArchitectureIsAttached(t *testing.T) {
+	files := map[string][]byte{
+		"/base.gguf":  ggufStart("general.architecture", "qwen2"),
+		"/model.gguf": ggufStart("general.architecture", "qwen2", "general.type", "model"),
+		"/llama.gguf": ggufStart("general.architecture", "llama", "general.type", "adapter"),
+		"/lora.gguf":  ggufStart("general.architecture", "qwen2", "general.type", "adapter"),
+	}
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusPartialContent)
+		_, _ = w.Write(files[r.URL.Path])
+	}))
+	defer srv.Close()
+
+	models := []catalog.Model{{
+		Name: "coder", Source: srv.URL + "/base.gguf",
+		Scaling:  catalog.Scaling{Min: 1, Max: 1},
+		Profiles: []catalog.Profile{{Name: "cpu", Engine: "llamacpp", MinMemoryGB: 8}},
+	}}
+	for _, tc := range []struct {
+		file, want string
+		code       int
+	}{
+		{"/model.gguf", "not a LoRA adapter", http.StatusBadRequest},
+		{"/llama.gguf", "for llama models", http.StatusBadRequest},
+		{"/lora.gguf", "", http.StatusOK},
+	} {
+		a, _ := setup(t, models)
+		fm := &fakeModels{}
+		a.SetModels(fm)
+		mux := http.NewServeMux()
+		a.Register(mux)
+		req := httptest.NewRequest(http.MethodPost, "/api/v1/models/coder/adapters",
+			strings.NewReader(`{"adapters":[{"name":"fix","source":"`+srv.URL+tc.file+`"}]}`))
+		rec := httptest.NewRecorder()
+		mux.ServeHTTP(rec, req)
+		if rec.Code != tc.code || !strings.Contains(rec.Body.String(), tc.want) {
+			t.Errorf("%s: %d %s; want %d mentioning %q", tc.file, rec.Code, rec.Body.String(), tc.code, tc.want)
+		}
+	}
+}
@@ -9,17 +9,21 @@ package adminapi
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
-	"gitea.mynodes.xyz/ilker/llmgw/internal/control/modelserve"
-	"gitea.mynodes.xyz/ilker/llmgw/internal/release"
 	"io"
 	"net/http"
+	"net/url"
 	"slices"
 	"strconv"
 	"strings"
 	"time"
 	"unicode"

+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/modelserve"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/gguf"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/release"
+
 	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/discovery"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/bench"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/control"
@@ -60,6 +64,7 @@ func (a *API) Register(mux *http.ServeMux) {
 	mux.HandleFunc("POST /api/v1/models/{name}/scale", a.scaleModel)
 	mux.HandleFunc("POST /api/v1/models/{name}/resources", a.setResources)
 	a.registerQuotas(mux)
+	mux.HandleFunc("POST /api/v1/models/{name}/adapters", a.setAdapters)
 	mux.HandleFunc("GET /api/v1/models/{name}/target", a.target)
 	mux.HandleFunc("POST /api/v1/models/{name}/class", func(w http.ResponseWriter, r *http.Request) {
 		if a.models == nil {
@@ -107,6 +112,7 @@ func (a *API) Register(mux *http.ServeMux) {
 	mux.HandleFunc("GET /api/v1/explain-placement", a.explainPlacement)
 	mux.HandleFunc("POST /api/v1/nodes/{id}/state", a.setNodeState)
 	mux.HandleFunc("POST /api/v1/nodes/{id}/upgrade", a.upgradeNode)
+	mux.HandleFunc("POST /api/v1/nodes/{id}/engines", a.setNodeEngine)
 	mux.HandleFunc("DELETE /api/v1/nodes/{id}", a.removeNode)
 	mux.HandleFunc("POST /api/v1/fabric-test", a.fabricTest)
 	mux.HandleFunc("GET /api/v1/nodes/{id}/logs", a.nodeLogs)
@@ -152,6 +158,10 @@ type NodeView struct {
 	// Provision is where installing each wanted engine stands. An engine is in
 	// Engines only once its installation here is ready.
 	Provision []inventory.EngineInstall `json:"provision,omitempty"`
+	// EngineBuilds are the builds chosen for this node, by engine: "auto"
+	// (its hardware decides), a build ("cu129"), or "none". Only on a node
+	// with GPUs, the only kind with a choice to make.
+	EngineBuilds map[string]string `json:"engine_builds,omitempty"`
 	// OS is the operating system and architecture, as the agent read them.
 	OS string `json:"os,omitempty"`
 	// Warnings are facts about the machine that make it slow or unusable
@@ -204,10 +214,13 @@ type NodeView struct {

 // GPUView is one card as a node's page shows it.
 type GPUView struct {
-	Index       int    `json:"index"`
-	Name        string `json:"name"`
-	NUMA        int    `json:"numa"`
-	NVLinkPeers []int  `json:"nvlink_peers,omitempty"`
+	Index int    `json:"index"`
+	Name  string `json:"name"`
+	// ComputeCapability is what decides which engine builds run on the card:
+	// CUDA 13 builds need 7.5, so a V100 (7.0) runs only older ones.
+	ComputeCapability string `json:"compute_capability,omitempty"`
+	NUMA              int    `json:"numa"`
+	NVLinkPeers       []int  `json:"nvlink_peers,omitempty"`
 	// NearestNIC is the adapter closest to the card, and how close.
 	NearestNIC  string `json:"nearest_nic,omitempty"`
 	NICPath     string `json:"nic_path,omitempty"`
@@ -312,10 +325,15 @@ type ModelView struct {
 	Warnings []string `json:"warnings,omitempty"`
 	// Class is its capacity class; Preempted the replicas it lost to a higher
 	// class; HeldFor the model it is giving room to now, if any.
-	Class     string `json:"class"`
-	Owner     string `json:"owner,omitempty"`
-	Preempted int    `json:"preempted,omitempty"`
-	HeldFor   string `json:"held_for,omitempty"`
+	Class string `json:"class"`
+	Owner string `json:"owner,omitempty"`
+	// External is where a model served outside the cluster goes, provider
+	// host and its name for the model; empty for one served here.
+	External string `json:"external,omitempty"`
+	// Adapters are the LoRA adapters it serves.
+	Adapters  []catalog.Adapter `json:"adapters,omitempty"`
+	Preempted int               `json:"preempted,omitempty"`
+	HeldFor   string            `json:"held_for,omitempty"`
 }

 // ReplicaView is a replica as `ctl replicas` shows it.
@@ -408,6 +426,10 @@ type HeadroomView struct {
 	Blocked []string `json:"blocked,omitempty"`
 	// TokPerSec is what one more replica was measured to serve; 0 unmeasured.
 	TokPerSec float64 `json:"tok_per_sec,omitempty"`
+	// ByPreempting is how many more would fit by stopping replicas this
+	// model's class may take; Preempts the models those belong to.
+	ByPreempting int      `json:"by_preempting,omitempty"`
+	Preempts     []string `json:"preempts,omitempty"`
 }

 // RejectionView is one rejected candidate.
@@ -457,6 +479,13 @@ func (a *API) Nodes() []NodeView {
 		}
 		slices.Sort(v.Engines)
 		v.Provision = n.Provision
+		if ch := a.ctl.EngineChoices(); ch != nil && len(n.GPUs) > 0 {
+			chosen := ch.For(n.ID)
+			v.EngineBuilds = map[string]string{"vllm": "auto"}
+			for e, b := range chosen {
+				v.EngineBuilds[e] = b
+			}
+		}
 		if h := n.Host; h.OSID != "" {
 			v.OS = strings.TrimSpace(h.OSID + " " + h.OSVersion + " " + h.Arch)
 			if h.Arch == "amd64" && !h.HasCPUFlag("avx2") {
@@ -484,7 +513,8 @@ func (a *API) Nodes() []NodeView {
 		v.Load = loadView(n.Load, time.Now())
 		v.NVLinkDomain, v.IMEX, v.IMEXChannels = n.NVLinkDomain(), n.IMEX.State, n.IMEX.Channels
 		for _, g := range n.GPUs {
-			gv := GPUView{Index: g.Index, Name: g.Name, NUMA: g.NUMANode, NVLinkPeers: g.NVLinkPeers}
+			gv := GPUView{Index: g.Index, Name: g.Name, ComputeCapability: g.ComputeCapability,
+				NUMA: g.NUMANode, NVLinkPeers: g.NVLinkPeers}
 			for hca, rel := range g.NICPaths {
 				if gv.NearestNIC == "" || discovery.PathRank(rel) < discovery.PathRank(gv.NICPath) ||
 					(discovery.PathRank(rel) == discovery.PathRank(gv.NICPath) && hca < gv.NearestNIC) {
@@ -565,6 +595,7 @@ func (a *API) Models() []ModelView {
 			Desired:   h.Desired, Failures: h.Failures,
 			Parked: h.Parked, Reason: h.Reason, Unplaceable: h.Unplaceable,
 			Class: m.EffectiveClass(), Owner: m.Owner, Preempted: h.Preempted, HeldFor: h.HeldFor,
+			External: externalLabel(m), Adapters: m.Adapters,
 		}
 		for _, p := range m.Profiles {
 			v.Profiles = append(v.Profiles, p.Name)
@@ -647,6 +678,7 @@ type Models interface {
 	Resources(ctx context.Context, name, profile string, memoryGB, cpuCores int, by string) (catalog.Model, error)
 	SetClass(ctx context.Context, name, class, by string) (catalog.Model, error)
 	SetOwner(ctx context.Context, name, owner, by string) (catalog.Model, error)
+	SetAdapters(ctx context.Context, name string, adapters []catalog.Adapter, by string) (catalog.Model, error)
 }

 // ResourcesRequest sets what a model's profiles reserve of a node.
@@ -967,7 +999,8 @@ func (a *API) Headroom(model string, limit int) (HeadroomView, bool) {
 		return HeadroomView{}, false
 	}
 	return HeadroomView{Model: model, Current: h.Current, More: h.More, Nodes: h.Nodes,
-		Limited: h.Limited, Blocked: h.Blocked, TokPerSec: h.TokPerSec}, true
+		Limited: h.Limited, Blocked: h.Blocked, TokPerSec: h.TokPerSec,
+		ByPreempting: h.ByPreempting, Preempts: h.Preempts}, true
 }

 func (a *API) Explain(model string) ExplainView {
@@ -1222,6 +1255,9 @@ type RestartRequest struct {
 	Replicas []string `json:"replicas,omitempty"`
 	Models   []string `json:"models,omitempty"`
 	All      bool     `json:"all,omitempty"`
+	// InPlace stops each replica before starting its replacement, rather
+	// than starting the replacement beside it first.
+	InPlace bool `json:"in_place,omitempty"`
 }

 // RestartResult lists the replicas queued.
@@ -1245,7 +1281,11 @@ func (a *API) restart(w http.ResponseWriter, r *http.Request) {
 		writeError(w, http.StatusBadRequest, "nothing ready matches; name replicas, models, or all")
 		return
 	}
-	if err := a.ctl.Reconciler().Restart(ids...); err != nil {
+	restart := a.ctl.Reconciler().Restart
+	if req.InPlace {
+		restart = a.ctl.Reconciler().RestartInPlace
+	}
+	if err := restart(ids...); err != nil {
 		writeError(w, http.StatusNotFound, err.Error())
 		return
 	}
@@ -1438,3 +1478,114 @@ func reservationWarnings(m catalog.Model, measuredGB float64) []string {
 	}
 	return out
 }
+
+// externalLabel is "api.openai.com · gpt-4o-mini" for a model served elsewhere.
+func externalLabel(m catalog.Model) string {
+	if !m.IsExternal() {
+		return ""
+	}
+	host := m.External.URL
+	if u, err := url.Parse(host); err == nil && u.Host != "" {
+		host = u.Host
+	}
+	return host + " · " + m.External.Model
+}
+
+// AdaptersRequest replaces a model's LoRA adapters with these.
+type AdaptersRequest struct {
+	Adapters []catalog.Adapter `json:"adapters"`
+}
+
+// setAdapters replaces a model's adapters, after reading each new file to
+// check it is an adapter for the model's architecture: one that is not is a
+// replica that fails to start, found minutes later on a node.
+func (a *API) setAdapters(w http.ResponseWriter, r *http.Request) {
+	if a.models == nil {
+		writeError(w, http.StatusNotImplemented, "this control plane has no store")
+		return
+	}
+	var req AdaptersRequest
+	if err := json.NewDecoder(io.LimitReader(r.Body, 1<<16)).Decode(&req); err != nil {
+		writeError(w, http.StatusBadRequest, "the request is not valid JSON: "+err.Error())
+		return
+	}
+	name := r.PathValue("name")
+	m, ok := a.ctl.Catalog().Resolve(name)
+	if !ok {
+		writeError(w, http.StatusBadRequest, "no model "+name)
+		return
+	}
+	base, baseErr := gguf.KindFromURL(r.Context(), nil, m.Source)
+	have := map[string]bool{}
+	for _, ad := range m.Adapters {
+		have[ad.Source] = true
+	}
+	for _, ad := range req.Adapters {
+		if have[ad.Source] {
+			continue // already checked when it was attached
+		}
+		h, err := gguf.KindFromURL(r.Context(), nil, ad.Source)
+		switch {
+		case err != nil:
+			writeError(w, http.StatusBadRequest, fmt.Sprintf("adapter %s: %v", ad.Name, err))
+			return
+		case !h.IsAdapter():
+			writeError(w, http.StatusBadRequest, fmt.Sprintf("adapter %s: %s is a model, not a LoRA adapter", ad.Name, ad.Source))
+			return
+		case baseErr == nil && h.Architecture != base.Architecture:
+			writeError(w, http.StatusBadRequest, fmt.Sprintf("adapter %s is for %s models and %s is %s",
+				ad.Name, h.Architecture, m.Name, base.Architecture))
+			return
+		}
+	}
+	out, err := a.models.SetAdapters(r.Context(), m.Name, req.Adapters, actorOf(r))
+	if err != nil {
+		writeError(w, http.StatusBadRequest, err.Error())
+		return
+	}
+	writeJSON(w, http.StatusOK, AdaptersRequest{Adapters: out.Adapters})
+}
+
+// NodeEngineRequest chooses one engine's build for a node.
+type NodeEngineRequest struct {
+	Engine string `json:"engine"`
+	// Build is "auto", a build such as "cu129", or "none".
+	Build string `json:"build"`
+}
+
+// setNodeEngine chooses an engine's build for a node, and has the node
+// install it.
+func (a *API) setNodeEngine(w http.ResponseWriter, r *http.Request) {
+	ch := a.ctl.EngineChoices()
+	if ch == nil {
+		writeError(w, http.StatusNotImplemented, "this control plane has no store to keep the choice in")
+		return
+	}
+	id := r.PathValue("id")
+	if _, ok := a.ctl.Inventory().Get(id); !ok {
+		writeError(w, http.StatusNotFound, "no node "+id)
+		return
+	}
+	var req NodeEngineRequest
+	if err := json.NewDecoder(io.LimitReader(r.Body, 1<<12)).Decode(&req); err != nil {
+		writeError(w, http.StatusBadRequest, "the request is not valid JSON: "+err.Error())
+		return
+	}
+	if err := ch.Set(r.Context(), id, req.Engine, req.Build, actorOf(r)); err != nil {
+		writeError(w, http.StatusBadRequest, err.Error())
+		return
+	}
+	writeJSON(w, http.StatusOK, req)
+}
+
+// SetNodeEngine chooses an engine's build for a node, for the cluster page.
+func (a *API) SetNodeEngine(ctx context.Context, node, engine, build, by string) error {
+	ch := a.ctl.EngineChoices()
+	if ch == nil {
+		return errors.New("this control plane has no store to keep the choice in")
+	}
+	if _, ok := a.ctl.Inventory().Get(node); !ok {
+		return errors.New("no node " + node)
+	}
+	return ch.Set(ctx, node, engine, build, by)
+}
@@ -258,6 +258,10 @@ type fakeModels struct {
 	err     error
 }

+func (f *fakeModels) SetAdapters(_ context.Context, name string, as []catalog.Adapter, _ string) (catalog.Model, error) {
+	return catalog.Model{Name: name, Adapters: as}, f.err
+}
+
 func (f *fakeModels) SetOwner(_ context.Context, name, owner, _ string) (catalog.Model, error) {
 	return catalog.Model{Name: name, Owner: owner}, f.err
 }
@@ -0,0 +1,63 @@
+package adminapi
+
+import (
+	"context"
+	"log/slog"
+	"net/http"
+	"testing"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/inventory"
+)
+
+// choiceStore keeps engine choices in memory.
+type choiceStore struct{ m map[string]map[string]string }
+
+func (c *choiceStore) NodeEngines(context.Context) (map[string]map[string]string, error) {
+	return c.m, nil
+}
+func (c *choiceStore) SetNodeEngine(_ context.Context, node, engine, variant, _ string) error {
+	if c.m[node] == nil {
+		c.m[node] = map[string]string{}
+	}
+	c.m[node][engine] = variant
+	return nil
+}
+
+func TestAnEngineBuildIsChosenForANodeAndTheNodeIsTold(t *testing.T) {
+	a, ctl := setup(t, testModels(), gpuNode("gpu07", 8), inventory.Node{ID: "cpu01", State: inventory.Ready})
+	ch := control.NewEngineChoices(t.Context(), &choiceStore{m: map[string]map[string]string{}}, slog.New(slog.DiscardHandler))
+	var told []string
+	ch.OnChange(func(node string) error { told = append(told, node); return nil })
+	ctl.SetEngineChoices(ch)
+
+	if code := call(t, a, http.MethodPost, "/api/v1/nodes/gpu07/engines", `{"engine":"vllm","build":"cu129"}`, nil); code != http.StatusOK {
+		t.Fatalf("status %d", code)
+	}
+	if len(told) != 1 || told[0] != "gpu07" {
+		t.Errorf("the node was not told: %v", told)
+	}
+	var nodes []NodeView
+	call(t, a, http.MethodGet, "/api/v1/nodes", "", &nodes)
+	for _, n := range nodes {
+		switch n.ID {
+		case "gpu07":
+			if n.EngineBuilds["vllm"] != "cu129" {
+				t.Errorf("gpu07 shows %v", n.EngineBuilds)
+			}
+		case "cpu01":
+			if n.EngineBuilds != nil {
+				t.Errorf("a node with no GPU has no build to choose: %v", n.EngineBuilds)
+			}
+		}
+	}
+	for _, bad := range []string{`{"engine":"vllm","build":"cu200"}`, `{"engine":"llamacpp","build":"cu129"}`} {
+		if code := call(t, a, http.MethodPost, "/api/v1/nodes/gpu07/engines", bad, nil); code != http.StatusBadRequest {
+			t.Errorf("%s: status %d", bad, code)
+		}
+	}
+	if code := call(t, a, http.MethodPost, "/api/v1/nodes/gpu07/engines", `{"engine":"vllm","build":"auto"}`, nil); code != http.StatusOK ||
+		ch.For("gpu07")["vllm"] != "" {
+		t.Errorf("auto should leave it to the node again: %d %v", code, ch.For("gpu07"))
+	}
+}
@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"io"
 	"net/http"
+	"sort"
 	"strings"
 	"time"

@@ -72,38 +73,24 @@ func (a *API) registerQuotas(mux *http.ServeMux) {
 			writeError(w, http.StatusBadRequest, "the request is not valid JSON: "+err.Error())
 			return
 		}
-		scope, id, _, err := a.resolveScope(r.Context(), req.Scope)
+		key, err := a.SetQuota(r.Context(), req, actorOf(r))
 		if err != nil {
 			writeError(w, http.StatusBadRequest, err.Error())
 			return
 		}
-		if !req.Guarantee.Any() && !req.Limit.Any() {
-			writeError(w, http.StatusBadRequest, "a quota needs a guarantee or a limit; to remove one, DELETE it")
-			return
-		}
-		q := store.CapacityQuota{Scope: scope, ScopeID: id, Guarantee: req.Guarantee, Limit: req.Limit,
-			UpdatedBy: actorOf(r)}
-		if err := a.quotas.PutCapacityQuota(r.Context(), q); err != nil {
-			writeError(w, http.StatusBadRequest, err.Error())
-			return
-		}
-		writeJSON(w, http.StatusOK, map[string]string{"scope": quota.Key(scope, id)})
+		writeJSON(w, http.StatusOK, map[string]string{"scope": key})
 	})
 	mux.HandleFunc("DELETE /api/v1/quotas/{scope}", func(w http.ResponseWriter, r *http.Request) {
 		if a.quotas == nil {
 			writeError(w, http.StatusNotImplemented, "this control plane has no store")
 			return
 		}
-		scope, id, _, err := a.resolveScope(r.Context(), r.PathValue("scope"))
+		key, err := a.ClearQuota(r.Context(), r.PathValue("scope"))
 		if err != nil {
 			writeError(w, http.StatusBadRequest, err.Error())
 			return
 		}
-		if err := a.quotas.DeleteCapacityQuota(r.Context(), scope, id); err != nil {
-			writeError(w, http.StatusInternalServerError, err.Error())
-			return
-		}
-		writeJSON(w, http.StatusOK, map[string]string{"removed": quota.Key(scope, id)})
+		writeJSON(w, http.StatusOK, map[string]string{"removed": key})
 	})
 	mux.HandleFunc("POST /api/v1/models/{name}/owner", func(w http.ResponseWriter, r *http.Request) {
 		if a.models == nil || a.quotas == nil {
@@ -137,6 +124,9 @@ func (a *API) registerQuotas(mux *http.ServeMux) {

 // QuotaViews lists every quota with what its scope holds now.
 func (a *API) QuotaViews(ctx context.Context) ([]QuotaView, error) {
+	if a.quotas == nil {
+		return nil, nil
+	}
 	qs, err := a.quotas.CapacityQuotas(ctx)
 	if err != nil {
 		return nil, err
@@ -235,3 +225,61 @@ func (a *API) resolveScope(ctx context.Context, s string) (store.Scope, string,
 	}
 	return "", "", "", fmt.Errorf("%q could be any of %d %ss; give its ID", what, len(found), kind)
 }
+
+// SetQuota sets a scope's capacity quota, named by path or ID, and returns
+// the scope as org:ID.
+func (a *API) SetQuota(ctx context.Context, req QuotaRequest, by string) (string, error) {
+	if a.quotas == nil {
+		return "", fmt.Errorf("this control plane has no store")
+	}
+	scope, id, _, err := a.resolveScope(ctx, req.Scope)
+	if err != nil {
+		return "", err
+	}
+	if !req.Guarantee.Any() && !req.Limit.Any() {
+		return "", fmt.Errorf("a quota needs a guarantee or a limit; to remove one, clear it")
+	}
+	q := store.CapacityQuota{Scope: scope, ScopeID: id, Guarantee: req.Guarantee, Limit: req.Limit, UpdatedBy: by}
+	if err := a.quotas.PutCapacityQuota(ctx, q); err != nil {
+		return "", err
+	}
+	return quota.Key(scope, id), nil
+}
+
+// ClearQuota removes a scope's capacity quota.
+func (a *API) ClearQuota(ctx context.Context, scope string) (string, error) {
+	if a.quotas == nil {
+		return "", fmt.Errorf("this control plane has no store")
+	}
+	s, id, _, err := a.resolveScope(ctx, scope)
+	if err != nil {
+		return "", err
+	}
+	if err := a.quotas.DeleteCapacityQuota(ctx, s, id); err != nil {
+		return "", err
+	}
+	return quota.Key(s, id), nil
+}
+
+// ScopeName is a part of the tenancy tree: org:ID and its path.
+type ScopeName struct {
+	Key  string `json:"scope"`
+	Name string `json:"name"`
+}
+
+// Scopes lists every organization, team and project, by path.
+func (a *API) Scopes(ctx context.Context) ([]ScopeName, error) {
+	if a.quotas == nil {
+		return nil, nil
+	}
+	names, err := a.scopeNames(ctx)
+	if err != nil {
+		return nil, err
+	}
+	out := make([]ScopeName, 0, len(names))
+	for k, n := range names {
+		out = append(out, ScopeName{Key: k, Name: n})
+	}
+	sort.Slice(out, func(i, j int) bool { return out[i].Name < out[j].Name })
+	return out, nil
+}
@@ -24,6 +24,9 @@ type TargetView struct {
 	Room    int      `json:"room"`
 	Fits    bool     `json:"fits"`
 	Blocked []string `json:"blocked,omitempty"`
+	// Preempts is whose replicas would be stopped to make the room, when
+	// free room alone is not enough; Room counts that room too.
+	Preempts []string `json:"preempts,omitempty"`
 }

 // Target sizes a model for a throughput.
@@ -40,8 +43,11 @@ func (a *API) Target(model string, tokPerSec float64) (TargetView, error) {
 	}
 	v.Current = h.Current
 	v.More = max(v.Needed-h.Current, 0)
-	v.Room = h.More
-	v.Fits = v.More <= h.More || h.Limited
+	v.Room = h.More + h.ByPreempting
+	v.Fits = v.More <= v.Room || h.Limited
+	if v.More > h.More {
+		v.Preempts = h.Preempts
+	}
 	if !v.Fits {
 		v.Blocked = h.Blocked
 	}
@@ -31,6 +31,11 @@ func (a *API) Orphans() []OrphanView {
 	var wanted []string
 	for _, m := range a.ctl.Catalog().Models() {
 		wanted = append(wanted, m.Source)
+		// A model's adapters are its weights too; collecting them would
+		// leave replicas that cannot restart.
+		for _, ad := range m.Adapters {
+			wanted = append(wanted, ad.Source)
+		}
 	}

 	out := make([]OrphanView, 0)
@@ -189,6 +189,7 @@ func StartShardFromCommand(cmd reconciler.ShardCommand) *llmgwv1.StartShard {
 			GpusPerNode: int32(p.GPUsPerNode), Nodes: int32(p.Nodes),
 			Args: p.Args, Env: p.Env,
 			ModelPath: p.ModelPath, ServedModelName: p.ServedModelName,
+			Adapters: adaptersToWire(p.Adapters),
 		},
 		StartupTimeoutMs: cmd.StartupTimeout.Milliseconds(),
 		Pool:             cmd.Pool,
@@ -224,6 +225,7 @@ func ShardSpecFromStart(m *llmgwv1.StartShard) (engines.ShardSpec, time.Duration
 			GPUsPerNode: int(p.GetGpusPerNode()), Nodes: int(p.GetNodes()),
 			Args: p.GetArgs(), Env: p.GetEnv(),
 			ModelPath: p.GetModelPath(), ServedModelName: p.GetServedModelName(),
+			Adapters: adaptersFromWire(p.GetAdapters()),
 		},
 	}
 	return spec, time.Duration(m.GetStartupTimeoutMs()) * time.Millisecond
@@ -292,3 +294,19 @@ func fabricFrom(state, status, cluster, clique, bandwidth string) *discovery.GPU
 	}
 	return &discovery.GPUFabric{State: state, Status: status, ClusterUUID: cluster, CliqueID: clique, Bandwidth: bandwidth}
 }
+
+func adaptersToWire(as []engines.Adapter) []*llmgwv1.LoraAdapter {
+	var out []*llmgwv1.LoraAdapter
+	for _, a := range as {
+		out = append(out, &llmgwv1.LoraAdapter{Name: a.Name, Source: a.Path})
+	}
+	return out
+}
+
+func adaptersFromWire(as []*llmgwv1.LoraAdapter) []engines.Adapter {
+	var out []engines.Adapter
+	for _, a := range as {
+		out = append(out, engines.Adapter{Name: a.GetName(), Path: a.GetSource()})
+	}
+	return out
+}
@@ -0,0 +1,20 @@
+package agentsvc
+
+import (
+	"testing"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/engines"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/reconciler"
+)
+
+func TestAdaptersTravelToTheAgentInOrder(t *testing.T) {
+	cmd := reconciler.ShardCommand{ReplicaID: "r", ShardID: "s", Spec: engines.ShardSpec{
+		Profile: engines.Profile{Name: "cpu", Engine: "llamacpp", ModelPath: "/m.gguf",
+			Adapters: []engines.Adapter{{Name: "sql", Path: "http://mirror/sql.gguf"}, {Name: "docs", Path: "http://mirror/docs.gguf"}}},
+	}}
+	spec, _ := ShardSpecFromStart(StartShardFromCommand(cmd))
+	got := spec.Profile.Adapters
+	if len(got) != 2 || got[0].Name != "sql" || got[1].Path != "http://mirror/docs.gguf" {
+		t.Errorf("adapters %+v", got)
+	}
+}
@@ -6,6 +6,7 @@ import (

 	llmgwv1 "gitea.mynodes.xyz/ilker/llmgw/api/gen/llmgw/v1"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/provision"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/inventory"
 )

 func TestANodeIsAskedOnlyForEnginesItCanInstall(t *testing.T) {
@@ -25,3 +26,57 @@ func TestANodeIsAskedOnlyForEnginesItCanInstall(t *testing.T) {
 		t.Error("uv, a tool for the others, is offered as an engine")
 	}
 }
+
+func TestEachNodeIsToldTheEnginesDecidedForIt(t *testing.T) {
+	s := &Service{}
+	s.SetEnginesFor(func(n inventory.Node) map[string]string {
+		if n.ID == "volta" {
+			return map[string]string{"llamacpp": "b1", "vllm": "0.25.1"}
+		}
+		return map[string]string{"llamacpp": "b1", "vllm": "0.30.0"}
+	})
+	reg := &llmgwv1.Register{InstallableEngines: []string{"llamacpp", "vllm"}}
+	want := func(p *llmgwv1.Provision) map[string]string {
+		out := map[string]string{}
+		for _, e := range p.GetEngines() {
+			out[e.GetEngine()] = e.GetVersion()
+		}
+		return out
+	}
+	if got := want(s.provisionFor(inventory.Node{ID: "volta"}, installable(reg))); got["vllm"] != "0.25.1" {
+		t.Errorf("the Volta node: %v", got)
+	}
+	if got := want(s.provisionFor(inventory.Node{ID: "hopper"}, installable(reg))); got["vllm"] != "0.30.0" {
+		t.Errorf("the Hopper node: %v", got)
+	}
+	// A node that disabled vLLM does not offer it, and is not asked for it.
+	if got := want(s.provisionFor(inventory.Node{ID: "hopper"}, []string{"llamacpp"})); got["vllm"] != "" {
+		t.Errorf("asked a node for an engine it does not install: %v", got)
+	}
+}
+
+func TestABuildChosenForANodeIsWhatItIsTold(t *testing.T) {
+	s := &Service{}
+	s.SetEngines(map[string]string{"llamacpp": "b1", "vllm": "0.30.0"})
+	s.SetEngineChoices(func(node string) map[string]string {
+		return map[string]map[string]string{"gpu07": {"vllm": "cu129"}, "gpu08": {"vllm": "none"}}[node]
+	})
+	can := []string{"llamacpp", "vllm"}
+	variant := func(node string) (string, bool) {
+		for _, e := range s.provisionFor(inventory.Node{ID: node}, can).GetEngines() {
+			if e.GetEngine() == "vllm" {
+				return e.GetVariant(), true
+			}
+		}
+		return "", false
+	}
+	if v, ok := variant("gpu07"); !ok || v != "cu129" {
+		t.Errorf("gpu07: %q, %v", v, ok)
+	}
+	if _, ok := variant("gpu08"); ok {
+		t.Error("gpu08 was asked for a vLLM it was chosen to run none of")
+	}
+	if v, ok := variant("gpu09"); !ok || v != "" {
+		t.Errorf("gpu09, left to its hardware: %q, %v", v, ok)
+	}
+}
@@ -53,8 +53,13 @@ type Service struct {
 	mu    sync.RWMutex
 	conns map[string]*conn

-	// engines is what every node is told to install, set once at start.
-	engines map[string]string
+	// enginesFor is what a node is told to install, set once at start: the
+	// cluster's engines, with any override that picks the node laid over.
+	enginesFor func(node inventory.Node) map[string]string
+	// choices are the engine builds chosen per node, and installableBy what
+	// each connected node said it can install, for telling it again.
+	choices       func(nodeID string) map[string]string
+	installableBy map[string][]string

 	// logWaits are the log requests awaiting a reply, by request id.
 	logMu    sync.Mutex
@@ -68,7 +73,37 @@ type Service struct {
 }

 // SetEngines sets the engines every node installs, by name and version.
-func (s *Service) SetEngines(e map[string]string) { s.engines = maps.Clone(e) }
+func (s *Service) SetEngines(e map[string]string) {
+	e = maps.Clone(e)
+	s.enginesFor = func(inventory.Node) map[string]string { return e }
+}
+
+// SetEnginesFor sets what each node installs, decided per node: a cluster
+// of two GPU generations needs two vLLM builds, or vLLM on only one.
+func (s *Service) SetEnginesFor(f func(node inventory.Node) map[string]string) { s.enginesFor = f }
+
+// SetEngineChoices sets where the builds chosen for each node come from: by
+// engine, a build ("cu129") or "none". An engine not named is left to the
+// node's hardware.
+func (s *Service) SetEngineChoices(f func(nodeID string) map[string]string) { s.choices = f }
+
+// Reprovision tells a connected node again what to install, after a choice
+// for it changed; one not connected is told when it next connects.
+func (s *Service) Reprovision(nodeID string) error {
+	if s.enginesFor == nil {
+		return nil
+	}
+	s.mu.RLock()
+	can, ok := s.installableBy[nodeID]
+	s.mu.RUnlock()
+	node, known := s.inv.Get(nodeID)
+	if !ok || !known {
+		return nil
+	}
+	return s.send(nodeID, &llmgwv1.ControlMessage{
+		Msg: &llmgwv1.ControlMessage_Provision{Provision: s.provisionFor(node, can)},
+	})
+}

 // conn is one agent's open stream.
 type conn struct {
@@ -179,15 +214,15 @@ func (s *Service) Connect(
 	// What this node should have installed. Sent on every registration, so a
 	// node rebuilt from nothing is told the same as one that has been here a
 	// month, and a changed version reaches every node on its next connect.
-	if s.engines != nil {
-		p := &llmgwv1.Provision{}
+	if s.enginesFor != nil {
 		can := installable(reg)
-		for _, e := range slices.Sorted(maps.Keys(s.engines)) {
-			if !slices.Contains(can, e) {
-				continue
-			}
-			p.Engines = append(p.Engines, &llmgwv1.EngineWant{Engine: e, Version: s.engines[e]})
+		s.mu.Lock()
+		if s.installableBy == nil {
+			s.installableBy = map[string][]string{}
 		}
+		s.installableBy[nodeID] = can
+		s.mu.Unlock()
+		p := s.provisionFor(node, can)
 		if err := stream.Send(&llmgwv1.ControlMessage{
 			Msg: &llmgwv1.ControlMessage_Provision{Provision: p},
 		}); err != nil {
@@ -550,6 +585,9 @@ func loadFrom(l *llmgwv1.NodeLoad, at time.Time) *inventory.Load {
 		out.GPUs = append(out.GPUs, inventory.GPUSensor{
 			Index: int(g.GetIndex()), TempC: g.GetTempC(), MemoryTempC: g.GetMemoryTempC(),
 			PowerW: g.GetPowerW(), PowerLimitW: g.GetPowerLimitW(), Throttle: g.GetThrottle(),
+			UtilPercent: g.GetUtilizationPercent(), MemUtilPercent: g.GetMemoryUtilizationPercent(),
+			MemoryUsedMB: g.GetMemoryUsedMb(), MemoryTotalMB: g.GetMemoryTotalMb(),
+			SMClockMHz: g.GetSmClockMhz(), ECCUncorrected: g.GetEccUncorrected(),
 		})
 	}
 	return out
@@ -853,6 +891,24 @@ func rate(before, now uint64, secs float64) float64 {
 	return float64(now-before) / secs
 }

+// provisionFor is what a node is told to install: the engines decided for
+// it, those it can install.
+func (s *Service) provisionFor(node inventory.Node, can []string) *llmgwv1.Provision {
+	p := &llmgwv1.Provision{}
+	wants := s.enginesFor(node)
+	var chosen map[string]string
+	if s.choices != nil {
+		chosen = s.choices(node.ID)
+	}
+	for _, e := range slices.Sorted(maps.Keys(wants)) {
+		if !slices.Contains(can, e) || chosen[e] == "none" {
+			continue
+		}
+		p.Engines = append(p.Engines, &llmgwv1.EngineWant{Engine: e, Version: wants[e], Variant: chosen[e]})
+	}
+	return p
+}
+
 // installable are the engines a node can be asked to install. An agent older
 // than the list said nothing, and could install llama.cpp alone: asked for
 // vLLM too, it reported it unsupported, and a cluster half way through an
@@ -7,6 +7,7 @@ import (
 	"encoding/hex"
 	"encoding/json"
 	"fmt"
+	"net/url"
 	"slices"
 	"strings"
 	"time"
@@ -160,8 +161,20 @@ type Model struct {
 	// Owner is the part of the tenancy tree whose capacity quota this model
 	// counts against, written org:ID, team:ID or project:ID. Empty is shared:
 	// held by the cluster, counted against nobody.
-	Owner     string   `json:"owner,omitempty" koanf:"owner"`
-	Fallbacks []string `json:"fallbacks,omitempty" koanf:"fallbacks"`
+	Owner string `json:"owner,omitempty" koanf:"owner"`
+	// External is a model served by someone else — an OpenAI-compatible API
+	// outside the cluster — rather than by replicas placed here. Such a model
+	// has no profiles and nothing is placed for it; the router sends its
+	// requests to the provider. Most useful named in another model's
+	// fallbacks: used only when the cluster has no capacity left. Its
+	// prompts leave the cluster, so which keys may name it is worth an
+	// allowlist.
+	External *External `json:"external,omitempty" koanf:"external"`
+	// Adapters are LoRA adapters served on this model's replicas, each asked
+	// for by its own name as though it were a model. Every replica loads
+	// every adapter, applying one only to a request that names it.
+	Adapters  []Adapter `json:"adapters,omitempty" koanf:"adapters"`
+	Fallbacks []string  `json:"fallbacks,omitempty" koanf:"fallbacks"`
 	// RoutingProfile selects how the router scores endpoints for this model.
 	RoutingProfile string `json:"routing_profile,omitempty" koanf:"routing_profile"`
 	SLO            SLO    `json:"slo,omitempty" koanf:"slo"`
@@ -204,14 +217,17 @@ type Capabilities struct {
 type Catalog struct {
 	models  map[string]Model
 	aliases map[string]string
-	order   []string
+	// adapters maps an adapter's name to the model it runs on.
+	adapters map[string]string
+	order    []string
 }

 // New builds a catalog, rejecting definitions that would fail confusingly later.
 func New(models []Model) (*Catalog, error) {
 	c := &Catalog{
-		models:  make(map[string]Model, len(models)),
-		aliases: map[string]string{},
+		models:   make(map[string]Model, len(models)),
+		aliases:  map[string]string{},
+		adapters: map[string]string{},
 	}

 	for i := range models {
@@ -240,6 +256,18 @@ func New(models []Model) (*Catalog, error) {
 		}
 	}

+	// Adapters are named like models, so they share the namespace too.
+	for _, name := range c.order {
+		for _, a := range c.models[name].Adapters {
+			_, isModel := c.models[a.Name]
+			_, isAlias := c.aliases[a.Name]
+			if _, isAdapter := c.adapters[a.Name]; isModel || isAlias || isAdapter {
+				return nil, fmt.Errorf("model %q: adapter %q collides with a model, an alias or another adapter", name, a.Name)
+			}
+			c.adapters[a.Name] = name
+		}
+	}
+
 	// Fallbacks are resolved now rather than at request time: a typo in a
 	// fallback should stop a deployment, not surface during an outage when the
 	// fallback is finally needed.
@@ -277,6 +305,16 @@ func (m Model) normalize() Model {
 		if p.Nodes <= 0 {
 			p.Nodes = 1
 		}
+		// tp on a llama.cpp GPU profile, written before gpus_per_node meant
+		// anything for it, is read as the width it was meant as: N cards of
+		// one node. Refusing it would have dropped a catalog's model on the
+		// upgrade that made the width work. Only both given, disagreeing, is
+		// refused below, as ambiguous.
+		if p.Engine == "llamacpp" && p.GPUClass != "" && (p.TP > 1 || p.PP > 1) && p.Nodes == 1 {
+			if n := p.TP * max(p.PP, 1); p.GPUsPerNode == 0 || p.GPUsPerNode == n {
+				p.GPUsPerNode, p.TP, p.PP = n, 1, 0
+			}
+		}
 		switch {
 		case p.CPUOnly():
 			// Nothing to divide up: this profile asks for no cards at all.
@@ -295,6 +333,13 @@ func (m Model) normalize() Model {
 			p.Weight = 1
 		}
 	}
+	// A Hugging Face repository page for a model vLLM or SGLang runs is the
+	// repository it names: the engine downloads a repository id itself,
+	// where the page was fetched as one file of HTML and failed at start.
+	if repo, ok := repositoryPage(m.Source); ok && len(m.Profiles) > 0 &&
+		!slices.ContainsFunc(m.Profiles, func(p Profile) bool { return p.Engine == "llamacpp" }) {
+		m.Source = repo
+	}
 	return m
 }

@@ -305,7 +350,23 @@ func (m Model) validate() error {
 	if m.Source == "" {
 		return fmt.Errorf("model %q: source is required", m.Name)
 	}
-	if len(m.Profiles) == 0 {
+	if repo, ok := repositoryPage(m.Source); ok {
+		// A page, not weights: a node fetches the HTML as one file and the
+		// engine fails at start, on a path that looks right. For vLLM it was
+		// read as the repository id when the model was made; only a model
+		// llama.cpp runs is left here, which needs a file.
+		return fmt.Errorf("model %q: source %s is a Hugging Face repository page, not a file; "+
+			"llama.cpp needs a .gguf: mirror one and use its source (repository %s)",
+			m.Name, m.Source, repo)
+	}
+	if err := validateAdapters(m); err != nil {
+		return err
+	}
+	if m.External != nil {
+		if err := m.External.validate(m.Name, len(m.Profiles)); err != nil {
+			return err
+		}
+	} else if len(m.Profiles) == 0 {
 		return fmt.Errorf("model %q: at least one profile is required", m.Name)
 	}
 	if m.Owner != "" && !validOwner(m.Owner) {
@@ -346,6 +407,22 @@ func (m Model) validate() error {
 			}
 			continue
 		}
+		// llama.cpp on GPUs splits a model across one node's cards by layer,
+		// and how many is gpus_per_node. Not tp: llama.cpp has no tensor
+		// parallelism, and a tp it was given was a width nothing placed — the
+		// replica stayed unplaced, or ran on one card while four were meant.
+		if p.Engine == "llamacpp" {
+			switch {
+			case p.Nodes > 1:
+				return fmt.Errorf("model %q profile %q: llama.cpp does not span nodes; use vllm or sglang for a %d-node replica",
+					m.Name, p.Name, p.Nodes)
+			case p.TP > 1 || p.PP > 1:
+				return fmt.Errorf("model %q profile %q: llama.cpp splits a model across a node's GPUs by layer, and "+
+					"tp %d does not agree with gpus_per_node %d: give gpus_per_node alone", m.Name, p.Name,
+					max(p.TP, 1)*max(p.PP, 1), p.GPUsPerNode)
+			}
+			continue
+		}
 		// The arithmetic has to add up, or the engine is handed a shard spec it
 		// cannot satisfy and fails minutes later during model loading.
 		if want := p.GPUsPerNode * p.Nodes; want != p.TP*max(p.PP, 1) {
@@ -357,6 +434,23 @@ func (m Model) validate() error {
 }

 // Resolve maps a client-supplied name or alias onto a model.
+// ResolveAdapter finds an adapter by its name: the model it runs on, the
+// adapter, and its index among the model's adapters — which is the id the
+// engine knows it by.
+func (c *Catalog) ResolveAdapter(name string) (Model, Adapter, int, bool) {
+	base, ok := c.adapters[name]
+	if !ok {
+		return Model{}, Adapter{}, 0, false
+	}
+	m := c.models[base]
+	for i, a := range m.Adapters {
+		if a.Name == name {
+			return m, a, i, true
+		}
+	}
+	return Model{}, Adapter{}, 0, false
+}
+
 func (c *Catalog) Resolve(name string) (Model, bool) {
 	if m, ok := c.models[name]; ok {
 		return m, true
@@ -433,6 +527,7 @@ func (m Model) EngineProfile(name string) (engines.Profile, bool) {
 			Env:             p.Env,
 			ModelPath:       m.Source,
 			ServedModelName: m.Name,
+			Adapters:        m.engineAdapters(),
 		}, true
 	}
 	return engines.Profile{}, false
@@ -564,3 +659,108 @@ func validOwner(o string) bool {
 	}
 	return false
 }
+
+// External is a provider outside the cluster that serves a model.
+type External struct {
+	// URL is the provider's base address, without /v1: https://api.openai.com.
+	URL string `json:"url" koanf:"url"`
+	// Model is what the provider calls it; requests are rewritten to it.
+	Model string `json:"model" koanf:"model"`
+	// APIKeyFile holds the provider's key, on the control plane. A file
+	// rather than the key itself, so the key is in neither the catalog, the
+	// store nor the audit log; empty for a provider that asks for none.
+	APIKeyFile string `json:"api_key_file,omitempty" koanf:"api_key_file"`
+}
+
+func (e *External) validate(model string, profiles int) error {
+	if profiles > 0 {
+		return fmt.Errorf("model %q: an external model is served elsewhere and has no profiles", model)
+	}
+	u, err := url.Parse(e.URL)
+	if err != nil || (u.Scheme != "https" && u.Scheme != "http") || u.Host == "" {
+		return fmt.Errorf("model %q: external url %q is not an http(s) address", model, e.URL)
+	}
+	if strings.TrimSpace(e.Model) == "" {
+		return fmt.Errorf("model %q: external model needs the name the provider knows it by", model)
+	}
+	return nil
+}
+
+// IsExternal reports whether a model is served outside the cluster.
+func (m Model) IsExternal() bool { return m.External != nil }
+
+// Adapter is a LoRA adapter served on a model's replicas.
+type Adapter struct {
+	// Name is what clients ask for, like a model name.
+	Name string `json:"name" koanf:"name"`
+	// Source is the adapter file (GGUF for llama.cpp), fetched by each node
+	// as the model's weights are.
+	Source string `json:"source" koanf:"source"`
+	// Scale is how strongly it is applied; zero takes 1.
+	Scale float64 `json:"scale,omitempty" koanf:"scale"`
+}
+
+// EffectiveScale is the scale an adapter is applied at.
+func (a Adapter) EffectiveScale() float64 {
+	if a.Scale <= 0 {
+		return 1
+	}
+	return a.Scale
+}
+
+func validateAdapters(m Model) error {
+	if len(m.Adapters) == 0 {
+		return nil
+	}
+	if m.External != nil {
+		return fmt.Errorf("model %q: an external model cannot carry adapters", m.Name)
+	}
+	for _, p := range m.Profiles {
+		// llama.cpp loads adapters at start and applies them per request;
+		// the other engines are not driven that way yet.
+		if p.Engine != "llamacpp" && p.Engine != "sim" {
+			return fmt.Errorf("model %q profile %q: adapters are served by llamacpp profiles only", m.Name, p.Name)
+		}
+	}
+	seen := map[string]bool{}
+	for _, a := range m.Adapters {
+		switch {
+		case a.Name == "" || a.Source == "":
+			return fmt.Errorf("model %q: an adapter needs a name and a source", m.Name)
+		case a.Name == m.Name:
+			return fmt.Errorf("model %q: an adapter cannot have the model's own name", m.Name)
+		case seen[a.Name]:
+			return fmt.Errorf("model %q: adapter %q is declared twice", m.Name, a.Name)
+		case a.Scale < 0:
+			return fmt.Errorf("model %q: adapter %q has a negative scale", m.Name, a.Name)
+		}
+		seen[a.Name] = true
+	}
+	return nil
+}
+
+func (m Model) engineAdapters() []engines.Adapter {
+	var out []engines.Adapter
+	for _, a := range m.Adapters {
+		out = append(out, engines.Adapter{Name: a.Name, Path: a.Source})
+	}
+	return out
+}
+
+// repositoryPage reports whether source is a Hugging Face repository's page
+// (https://huggingface.co/org/repo, or one of its tree views) rather than a
+// file under /resolve/, and the repository's id.
+func repositoryPage(source string) (string, bool) {
+	for _, host := range []string{"https://huggingface.co/", "http://huggingface.co/", "https://hf.co/"} {
+		rest, ok := strings.CutPrefix(source, host)
+		if !ok {
+			continue
+		}
+		parts := strings.Split(strings.Trim(rest, "/"), "/")
+		if len(parts) < 2 || (len(parts) > 2 && parts[2] == "resolve") {
+			return "", false
+		}
+		return parts[0] + "/" + parts[1], true
+	}
+	return "", false
+}
@@ -321,3 +321,67 @@ func TestAProfileCannotSetWhatRunsCodeInTheEngine(t *testing.T) {
 		t.Errorf("ordinary tuning refused: %v", err)
 	}
 }
+
+func TestALlamaCppReplicaOnSeveralGPUsSaysHowManyAsGPUsPerNode(t *testing.T) {
+	four := model("big", Profile{Name: "t4", Engine: "llamacpp", GPUClass: "t4", GPUsPerNode: 4})
+	c, err := New([]Model{four})
+	if err != nil {
+		t.Fatalf("gpus_per_node 4 on llama.cpp: %v", err)
+	}
+	m, _ := c.Resolve("big")
+	if pl := m.PlacementProfiles(); len(pl) != 1 || pl[0].GPUsPerNode != 4 {
+		t.Errorf("placement was not asked for four GPUs: %+v", pl)
+	}
+	// tp, written before gpus_per_node meant anything for llama.cpp, is read
+	// as the same width rather than refused: refusing dropped the model on
+	// the upgrade that made the width work.
+	old := model("old", Profile{Name: "t4", Engine: "llamacpp", GPUClass: "t4", TP: 4})
+	c, err = New([]Model{old})
+	if err != nil {
+		t.Fatalf("tp 4 on llama.cpp: %v", err)
+	}
+	got, _ := c.Resolve("old")
+	if p := got.PlacementProfiles()[0]; p.GPUsPerNode != 4 {
+		t.Errorf("tp 4 should place on four GPUs: %+v", p)
+	}
+	// Both given and disagreeing is ambiguous, and refused; so is more than
+	// one node.
+	for _, p := range []Profile{
+		{Name: "both", Engine: "llamacpp", GPUClass: "t4", TP: 4, GPUsPerNode: 2},
+		{Name: "nodes", Engine: "llamacpp", GPUClass: "t4", Nodes: 2, GPUsPerNode: 1},
+	} {
+		if _, err := New([]Model{model("x", p)}); err == nil {
+			t.Errorf("%s: accepted", p.Name)
+		}
+	}
+}
+
+func TestARepositoryPageIsReadAsTheRepositoryForVLLM(t *testing.T) {
+	// The field: an AWQ model's page given as a vLLM source, fetched as one
+	// file of HTML. vLLM downloads a repository id itself.
+	for _, src := range []string{
+		"https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-AWQ",
+		"https://huggingface.co/Qwen/Qwen2.5-32B-Instruct-AWQ/tree/main",
+	} {
+		m := model("awq")
+		m.Source = src
+		c, err := New([]Model{m})
+		if err != nil {
+			t.Fatalf("%s: %v", src, err)
+		}
+		if got, _ := c.Resolve("awq"); got.Source != "Qwen/Qwen2.5-32B-Instruct-AWQ" {
+			t.Errorf("%s became %q", src, got.Source)
+		}
+	}
+	// llama.cpp needs a file, which a page is not.
+	g := model("gguf", Profile{Name: "cpu", Engine: "llamacpp", MinMemoryGB: 4})
+	g.Source = "https://huggingface.co/Qwen/X-GGUF"
+	if _, err := New([]Model{g}); err == nil || !strings.Contains(err.Error(), "needs a .gguf") {
+		t.Errorf("a page for llama.cpp: %v", err)
+	}
+	f := model("file")
+	f.Source = "https://huggingface.co/Qwen/X-GGUF/resolve/abc123/x.gguf"
+	if _, err := New([]Model{f}); err != nil {
+		t.Errorf("a file under /resolve/ is a source: %v", err)
+	}
+}
@@ -9,6 +9,7 @@ import (
 	"encoding/hex"
 	"fmt"
 	"log/slog"
+	"os"
 	"strings"
 	"sync"
 	"time"
@@ -74,15 +75,18 @@ func DefaultConfig() Config {

 // Controller owns the control plane's state.
 type Controller struct {
-	cfg   Config
-	log   *slog.Logger
-	inv   *inventory.Store
-	cat   *catalog.Live
-	rec   *reconciler.Reconciler
-	lead  *Leadership
-	disp  reconciler.Dispatcher
-	sink  EndpointSink
-	usage UsageSource
+	// engineChoices are the engine builds chosen per node; nil without a
+	// store to keep them in.
+	engineChoices *EngineChoices
+	cfg           Config
+	log           *slog.Logger
+	inv           *inventory.Store
+	cat           *catalog.Live
+	rec           *reconciler.Reconciler
+	lead          *Leadership
+	disp          reconciler.Dispatcher
+	sink          EndpointSink
+	usage         UsageSource

 	mu   sync.Mutex
 	last string // fingerprint of the last published list
@@ -338,6 +342,34 @@ func (c *Controller) endpoints() []endpoint.Info {
 			Priority: prof.Priority,
 		})
 	}
+	return append(out, c.externalEndpoints()...)
+}
+
+// externalEndpoints are the models served outside the cluster, as endpoints:
+// one each, with the provider's key read from its file every time, so a key
+// rotated on disk is used from the next publish on.
+func (c *Controller) externalEndpoints() []endpoint.Info {
+	var out []endpoint.Info
+	for _, m := range c.cat.Load().Models() {
+		if !m.IsExternal() {
+			continue
+		}
+		key := ""
+		if f := m.External.APIKeyFile; f != "" {
+			b, err := os.ReadFile(f) //nolint:gosec // the operator's own path
+			if err != nil {
+				c.log.Warn("an external model's key could not be read; not offered", "model", m.Name, "err", err)
+				continue
+			}
+			key = strings.TrimSpace(string(b))
+		}
+		out = append(out, endpoint.Info{
+			ID: "external:" + m.Name, URL: strings.TrimRight(m.External.URL, "/"),
+			Models: append([]string{m.Name}, m.Aliases...), APIKey: key,
+			Engine: endpoint.EngineExternal, Pool: "external", Weight: 1,
+			UpstreamModel: m.External.Model,
+		})
+	}
 	return out
 }

@@ -345,7 +377,8 @@ func (c *Controller) endpoints() []endpoint.Info {
 func fingerprint(infos []endpoint.Info) string {
 	var b strings.Builder
 	for _, i := range infos {
-		fmt.Fprintf(&b, "%s|%s|%s|%s|%v|%g|%d;", i.ID, i.URL, i.Engine, i.Pool, i.Models, i.Weight, i.Priority)
+		fmt.Fprintf(&b, "%s|%s|%s|%s|%v|%g|%d|%s|%x;", i.ID, i.URL, i.Engine, i.Pool, i.Models, i.Weight, i.Priority,
+			i.UpstreamModel, sha256.Sum256([]byte(i.APIKey)))
 	}
 	return b.String()
 }
@@ -378,3 +411,9 @@ func EngineKeys(secret []byte) func(replicaID string) string {

 // QuotaUse is every capacity quota beside what its scope holds.
 func (c *Controller) QuotaUse() []reconciler.QuotaUse { return c.rec.QuotaUse() }
+
+// SetEngineChoices gives the controller the engine builds chosen per node.
+func (c *Controller) SetEngineChoices(e *EngineChoices) { c.engineChoices = e }
+
+// EngineChoices are the engine builds chosen per node, or nil.
+func (c *Controller) EngineChoices() *EngineChoices { return c.engineChoices }
@@ -0,0 +1,108 @@
+package control
+
+import (
+	"context"
+	"fmt"
+	"log/slog"
+	"maps"
+	"slices"
+	"strings"
+	"sync"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/provision"
+)
+
+// EngineChoiceStore is where the engine builds chosen per node are kept.
+type EngineChoiceStore interface {
+	NodeEngines(ctx context.Context) (map[string]map[string]string, error)
+	SetNodeEngine(ctx context.Context, node, engine, variant, by string) error
+}
+
+// EngineChoices are the engine builds chosen for nodes, rather than left to
+// their hardware: vLLM's CUDA 12.9 build on a node its CUDA 13 build would
+// also install on, or no vLLM at all. Kept in the store, held here for the
+// agent service to ask on every registration, and pushed to a node the
+// moment its choice changes.
+type EngineChoices struct {
+	st  EngineChoiceStore
+	log *slog.Logger
+
+	mu     sync.RWMutex
+	byNode map[string]map[string]string
+	notify func(node string) error
+}
+
+// NewEngineChoices reads what was chosen before.
+func NewEngineChoices(ctx context.Context, st EngineChoiceStore, log *slog.Logger) *EngineChoices {
+	c := &EngineChoices{st: st, log: log, byNode: map[string]map[string]string{}}
+	if got, err := st.NodeEngines(ctx); err != nil {
+		log.Error("could not read the engine builds chosen per node; every node installs what its hardware runs", "err", err)
+	} else {
+		c.byNode = got
+	}
+	return c
+}
+
+// OnChange is called with a node whose choice changed, to tell it.
+func (c *EngineChoices) OnChange(f func(node string) error) {
+	c.mu.Lock()
+	c.notify = f
+	c.mu.Unlock()
+}
+
+// For is what was chosen for one node, by engine: a build, or "none".
+func (c *EngineChoices) For(node string) map[string]string {
+	c.mu.RLock()
+	defer c.mu.RUnlock()
+	return maps.Clone(c.byNode[node])
+}
+
+// EngineBuilds are what can be chosen for each engine that has a choice:
+// "auto" is the node's hardware deciding, "none" no engine at all.
+func EngineBuilds(engine string) []string {
+	switch engine {
+	case "vllm":
+		return append(append([]string{"auto"}, provision.VLLMVariants...), "none")
+	case "sglang":
+		return []string{"auto", "none"}
+	}
+	return nil
+}
+
+// Set chooses an engine's build for a node, keeps it, and tells the node.
+func (c *EngineChoices) Set(ctx context.Context, node, engine, build, by string) error {
+	builds := EngineBuilds(engine)
+	if builds == nil {
+		return fmt.Errorf("%s has no builds to choose between", engine)
+	}
+	if !slices.Contains(builds, build) {
+		return fmt.Errorf("%s has no build %q; choose %s", engine, build, strings.Join(builds, ", "))
+	}
+	variant := build
+	if build == "auto" {
+		variant = ""
+	}
+	if err := c.st.SetNodeEngine(ctx, node, engine, variant, by); err != nil {
+		return err
+	}
+	c.mu.Lock()
+	if variant == "" {
+		delete(c.byNode[node], engine)
+	} else {
+		if c.byNode[node] == nil {
+			c.byNode[node] = map[string]string{}
+		}
+		c.byNode[node][engine] = variant
+	}
+	notify := c.notify
+	c.mu.Unlock()
+	c.log.Info("engine build chosen for a node", "node", node, "engine", engine, "build", build, "by", by)
+	if notify != nil {
+		if err := notify(node); err != nil {
+			// Kept and applied on its next connection; said, since a node that
+			// is not connected is the usual reason.
+			c.log.Warn("the node was not told yet; it is when it next connects", "node", node, "err", err)
+		}
+	}
+	return nil
+}
@@ -0,0 +1,75 @@
+package control
+
+import (
+	"context"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"sync"
+	"testing"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/catalog"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/reconciler"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/router/endpoint"
+)
+
+type recordSink struct {
+	mu    sync.Mutex
+	infos []endpoint.Info
+	times int
+}
+
+func (s *recordSink) SetEndpoints(_ context.Context, infos []endpoint.Info) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.infos, s.times = infos, s.times+1
+}
+
+type noDispatch struct{}
+
+func (noDispatch) StartShard(context.Context, reconciler.ShardCommand) error { return nil }
+func (noDispatch) StopShard(context.Context, string, string) error           { return nil }
+func (noDispatch) DrainShard(context.Context, string, string) error          { return nil }
+
+func TestAnExternalModelIsPublishedWithItsKeyAndNothingIsPlacedForIt(t *testing.T) {
+	dir := t.TempDir()
+	keyFile := filepath.Join(dir, "provider.key")
+	if err := os.WriteFile(keyFile, []byte("sk-provider-1\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	models := []catalog.Model{
+		{Name: "cloud", Source: "external", Aliases: []string{"gpt"}, Scaling: catalog.Scaling{Min: 3, Max: 3},
+			External: &catalog.External{URL: "https://api.example.com/", Model: "gpt-4o-mini", APIKeyFile: keyFile}},
+		{Name: "gone", Source: "external",
+			External: &catalog.External{URL: "https://api.example.com", Model: "m", APIKeyFile: filepath.Join(dir, "missing")}},
+	}
+	cat, err := catalog.New(models)
+	if err != nil {
+		t.Fatal(err)
+	}
+	sink := &recordSink{}
+	c := New(DefaultConfig(), catalog.NewLive(cat), noDispatch{}, sink, slog.New(slog.DiscardHandler))
+	c.Publish(t.Context())
+
+	if len(sink.infos) != 1 {
+		t.Fatalf("published %+v; the model whose key is missing must not be offered", sink.infos)
+	}
+	e := sink.infos[0]
+	if e.ID != "external:cloud" || e.URL != "https://api.example.com" || e.APIKey != "sk-provider-1" ||
+		e.UpstreamModel != "gpt-4o-mini" || e.Engine != endpoint.EngineExternal || len(e.Models) != 2 {
+		t.Errorf("endpoint %+v", e)
+	}
+	// A key rotated on disk is published on the next pass.
+	_ = os.WriteFile(keyFile, []byte("sk-provider-2"), 0o600)
+	c.Publish(t.Context())
+	if sink.times != 2 || sink.infos[0].APIKey != "sk-provider-2" {
+		t.Errorf("the rotated key was not published: %d pushes, key %q", sink.times, sink.infos[0].APIKey)
+	}
+	// Scaling of 3 or not, nothing is wanted of the cluster for a model
+	// served elsewhere, and so it never reads as short or unplaceable.
+	for _, h := range c.Reconciler().ModelHealth() {
+		if h.Model == "cloud" && (h.Desired != 0 || h.Unplaceable) {
+			t.Errorf("an external model wants replicas: %+v", h)
+		}
+	}
+}
@@ -73,6 +73,8 @@ type Collector struct {
 	gpuMemoryTotal *prometheus.Desc
 	gpuMemoryUsed  *prometheus.Desc
 	gpuUtil        *prometheus.Desc
+	gpuMemUtil     *prometheus.Desc
+	gpuSMClock     *prometheus.Desc
 	gpuTemp        *prometheus.Desc
 	gpuMemTemp     *prometheus.Desc
 	gpuPower       *prometheus.Desc
@@ -149,6 +151,11 @@ func New(src Source) *Collector {
 			"How busy a GPU was over the last sample period. Absent when the driver does not "+
 				"report it, which is not the same as idle. Memory used says a model is loaded; "+
 				"this says whether it is doing anything.", gpu, nil),
+		gpuMemUtil: prometheus.NewDesc("llmgw_gpu_memory_utilization_percent",
+			"How much of the sample period a GPU's memory was being read or written: "+
+				"decode is bound by it, so a card busy here and idle above is waiting on memory.", gpu, nil),
+		gpuSMClock: prometheus.NewDesc("llmgw_gpu_sm_clock_mhz",
+			"A GPU's streaming multiprocessor clock now; one held low under load is the card being throttled.", gpu, nil),
 		gpuMemTemp: prometheus.NewDesc("llmgw_gpu_memory_temperature_celsius",
 			"A GPU's memory (HBM) temperature, where the driver reports it.", gpu, nil),
 		gpuPower: prometheus.NewDesc("llmgw_gpu_power_watts",
@@ -271,7 +278,7 @@ func (c *Collector) Describe(ch chan<- *prometheus.Desc) {
 		c.nodeCPU, c.nodeCPUs, c.nodeLoad, c.nodeMemUsed, c.nodeMemAvail, c.shardCPU, c.shardMemory,
 		c.ibActive, c.ibRate, c.ibCounter, c.fabricSMs, c.fabricMaster, c.fabricProblems,
 		c.gpuNVLJoined, c.imexUp,
-		c.gpuMemoryTotal, c.gpuMemoryUsed, c.gpuUtil, c.gpuTemp, c.gpuHealthy,
+		c.gpuMemoryTotal, c.gpuMemoryUsed, c.gpuUtil, c.gpuMemUtil, c.gpuSMClock, c.gpuTemp, c.gpuHealthy,
 		c.gpuMemTemp, c.gpuPower, c.gpuPowerLimit, c.gpuThrottled, c.nodeCPUTemp, c.nodeCPUPower, c.nodeInletTemp,
 		c.gpuNVLink, c.gpuECC,
 		c.replicaState, c.replicaReady, c.replicaIdle,
@@ -388,20 +395,28 @@ func (c *Collector) Collect(ch chan<- prometheus.Metric) {
 			byClass[class]++
 			labels := []string{n.ID, strconv.Itoa(g.Index), class}

-			ch <- prometheus.MustNewConstMetric(c.gpuMemoryTotal, prometheus.GaugeValue,
-				float64(g.MemoryTotalMB)*1024*1024, labels...)
-			ch <- prometheus.MustNewConstMetric(c.gpuMemoryUsed, prometheus.GaugeValue,
-				float64(g.MemoryUsedMB)*1024*1024, labels...)
-			if g.UtilizationPercent > 0 {
-				// Only when reported. A driver too old to give it would
-				// otherwise show every card at a flat zero, which reads as an
-				// idle fleet rather than a missing measurement.
-				ch <- prometheus.MustNewConstMetric(c.gpuUtil, prometheus.GaugeValue,
-					float64(g.UtilizationPercent), labels...)
+			s, fresh := sensorFor(n.Load, g.Index, now)
+			if fresh && s.Busy() {
+				// What the card is doing now, sampled every heartbeat.
+				c.gpuBusy(ch, s, labels)
+			} else {
+				// An agent too old to sample it: what it read at
+				// registration, which is all there is.
+				ch <- prometheus.MustNewConstMetric(c.gpuMemoryTotal, prometheus.GaugeValue,
+					float64(g.MemoryTotalMB)*1024*1024, labels...)
+				ch <- prometheus.MustNewConstMetric(c.gpuMemoryUsed, prometheus.GaugeValue,
+					float64(g.MemoryUsedMB)*1024*1024, labels...)
+				if g.UtilizationPercent > 0 {
+					// Only when reported. A driver too old to give it would
+					// otherwise show every card at a flat zero, which reads as
+					// an idle fleet rather than a missing measurement.
+					ch <- prometheus.MustNewConstMetric(c.gpuUtil, prometheus.GaugeValue,
+						float64(g.UtilizationPercent), labels...)
+				}
 			}
 			// The sensors' fresh reading when there is one; the temperature
 			// read at registration otherwise, which is all an older agent sends.
-			if s, ok := sensorFor(n.Load, g.Index, now); ok {
+			if fresh {
 				c.gpuSensor(ch, s, labels)
 			} else if g.Health.TemperatureC > 0 {
 				ch <- prometheus.MustNewConstMetric(c.gpuTemp, prometheus.GaugeValue,
@@ -414,8 +429,11 @@ func (c *Collector) Collect(ch chan<- prometheus.Metric) {
 			ch <- prometheus.MustNewConstMetric(c.gpuHealthy, prometheus.GaugeValue, healthy, labels...)
 			ch <- prometheus.MustNewConstMetric(c.gpuNVLink, prometheus.GaugeValue,
 				float64(len(g.NVLinkPeers)), labels...)
-			ch <- prometheus.MustNewConstMetric(c.gpuECC, prometheus.GaugeValue,
-				float64(g.Health.ECCErrors), labels...)
+			ecc := float64(g.Health.ECCErrors)
+			if fresh && s.Busy() && s.ECCUncorrected >= 0 {
+				ecc = s.ECCUncorrected
+			}
+			ch <- prometheus.MustNewConstMetric(c.gpuECC, prometheus.GaugeValue, ecc, labels...)
 		}
 		for class, n2 := range byClass {
 			ch <- prometheus.MustNewConstMetric(c.nodeGPUs, prometheus.GaugeValue,
@@ -569,6 +587,23 @@ func sensorFor(l *inventory.Load, index int, now time.Time) (inventory.GPUSensor
 	return inventory.GPUSensor{}, false
 }

+// gpuBusy is how busy a card is, from a sample that carries it.
+func (c *Collector) gpuBusy(ch chan<- prometheus.Metric, s inventory.GPUSensor, labels []string) {
+	ch <- prometheus.MustNewConstMetric(c.gpuMemoryTotal, prometheus.GaugeValue, s.MemoryTotalMB*1024*1024, labels...)
+	ch <- prometheus.MustNewConstMetric(c.gpuMemoryUsed, prometheus.GaugeValue, s.MemoryUsedMB*1024*1024, labels...)
+	// Zero is a real reading here — an idle card — and only -1, the driver
+	// not saying, leaves the series out.
+	if s.UtilPercent >= 0 {
+		ch <- prometheus.MustNewConstMetric(c.gpuUtil, prometheus.GaugeValue, s.UtilPercent, labels...)
+	}
+	if s.MemUtilPercent >= 0 {
+		ch <- prometheus.MustNewConstMetric(c.gpuMemUtil, prometheus.GaugeValue, s.MemUtilPercent, labels...)
+	}
+	if s.SMClockMHz > 0 {
+		ch <- prometheus.MustNewConstMetric(c.gpuSMClock, prometheus.GaugeValue, s.SMClockMHz, labels...)
+	}
+}
+
 func (c *Collector) gpuSensor(ch chan<- prometheus.Metric, s inventory.GPUSensor, labels []string) {
 	for _, v := range []struct {
 		d   *prometheus.Desc
@@ -39,3 +39,41 @@ func TestSensorsArePublishedFreshAndOnlyWhenMeasured(t *testing.T) {
 		}
 	}
 }
+
+func TestHowBusyAGPUIsComesFromTheLatestSampleNotRegistration(t *testing.T) {
+	// Discovery read 87% and 70 GB used when the agent came up; the card has
+	// been idle since. The sample says so, zero included.
+	reg := discovery.GPU{Index: 0, Name: "NVIDIA H100 80GB HBM3", MemoryTotalMB: 81559,
+		MemoryUsedMB: 70000, UtilizationPercent: 87}
+	n := inventory.Node{ID: "g1", State: inventory.Ready, GPUs: []discovery.GPU{reg},
+		Load: &inventory.Load{At: time.Now(), GPUs: []inventory.GPUSensor{{Index: 0, TempC: 35,
+			UtilPercent: 0, MemUtilPercent: 0, MemoryUsedMB: 512, MemoryTotalMB: 81559,
+			SMClockMHz: 345, ECCUncorrected: 2}}}}
+	out := gather(t, fakeSource{nodes: []inventory.Node{n}})
+	for _, want := range []string{
+		"llmgw_gpu_utilization_percent class=h100-80g gpu=0 node=g1 -> 0\n",
+		"llmgw_gpu_memory_utilization_percent class=h100-80g gpu=0 node=g1 -> 0\n",
+		"llmgw_gpu_memory_used_bytes class=h100-80g gpu=0 node=g1 -> 536870912\n",
+		"llmgw_gpu_sm_clock_mhz class=h100-80g gpu=0 node=g1 -> 345\n",
+		"llmgw_gpu_ecc_errors class=h100-80g gpu=0 node=g1 -> 2\n",
+	} {
+		if !strings.Contains(out, want) {
+			t.Errorf("missing %s in:\n%s", want, out)
+		}
+	}
+	if strings.Contains(out, "-> 87\n") {
+		t.Error("the utilization read at registration was published over the sample's")
+	}
+
+	// An agent too old to sample it: registration's figures, and no zero
+	// standing in for a utilization the driver did not give.
+	n.Load.GPUs = []inventory.GPUSensor{{Index: 0, TempC: 35}}
+	out = gather(t, fakeSource{nodes: []inventory.Node{n}})
+	if !strings.Contains(out, "llmgw_gpu_utilization_percent class=h100-80g gpu=0 node=g1 -> 87") {
+		t.Errorf("an older agent's registration figure was dropped:\n%s", out)
+	}
+	n.Load.GPUs = []inventory.GPUSensor{{Index: 0, TempC: 35, UtilPercent: -1, MemUtilPercent: -1, MemoryTotalMB: 81559}}
+	if out := gather(t, fakeSource{nodes: []inventory.Node{n}}); strings.Contains(out, "llmgw_gpu_utilization_percent ") {
+		t.Errorf("a utilization the driver did not report was published:\n%s", out)
+	}
+}
@@ -319,8 +319,20 @@ type GPUSensor struct {
 	PowerW      float64  `json:"power_w,omitempty"`
 	PowerLimitW float64  `json:"power_limit_w,omitempty"`
 	Throttle    []string `json:"throttle,omitempty"`
+	// How busy the card is, as of this sample; -1 is not reported. Zero
+	// MemoryTotalMB is an agent too old to send these, whose figures are
+	// the ones it read at registration.
+	UtilPercent    float64 `json:"util_percent,omitempty"`
+	MemUtilPercent float64 `json:"mem_util_percent,omitempty"`
+	MemoryUsedMB   float64 `json:"memory_used_mb,omitempty"`
+	MemoryTotalMB  float64 `json:"memory_total_mb,omitempty"`
+	SMClockMHz     float64 `json:"sm_clock_mhz,omitempty"`
+	ECCUncorrected float64 `json:"ecc_uncorrected,omitempty"`
 }

+// Busy reports whether the sample carries how busy the card is.
+func (s GPUSensor) Busy() bool { return s.MemoryTotalMB > 0 }
+
 // ShardLoad is one engine's use: cores busy, and resident memory.
 type ShardLoad struct {
 	CPUCores float64
@@ -87,6 +87,9 @@ func (s *Service) Sources(ctx context.Context) (map[string][]catalog.Model, erro
 		if m.Source != "" {
 			out[m.Source] = append(out[m.Source], m)
 		}
+		for _, ad := range m.Adapters {
+			out[ad.Source] = append(out[ad.Source], m)
+		}
 	}
 	return out, nil
 }
@@ -268,6 +271,23 @@ func (s *Service) SetOwner(ctx context.Context, name, owner, by string) (catalog
 	return final, err
 }

+// SetAdapters replaces the LoRA adapters a model serves. Replicas load the
+// new set when they next start.
+func (s *Service) SetAdapters(ctx context.Context, name string, adapters []catalog.Adapter, by string) (catalog.Model, error) {
+	final, err := s.update(ctx, name, by, "change its adapters", func(m *catalog.Model) error {
+		m.Adapters = adapters
+		return nil
+	})
+	if err == nil {
+		names := make([]string, 0, len(adapters))
+		for _, a := range adapters {
+			names = append(names, a.Name)
+		}
+		s.log.Info("model adapters set", "model", name, "adapters", names, "by", by)
+	}
+	return final, err
+}
+
 // update changes one model added at runtime and keeps the change.
 //
 // Built before anything is written, like Add: catalog.New is what applies the
@@ -94,6 +94,9 @@ func (r *Reconciler) desiredFor(m catalog.Model) int {
 // baseFor is what a model runs without load: its minimum, or for an on-demand
 // model one while somebody is asking and none once they stop.
 func (r *Reconciler) baseFor(m catalog.Model) int {
+	if m.IsExternal() {
+		return 0 // served elsewhere: nothing to place, whatever its scaling says
+	}
 	if !m.Scaling.OnDemand() {
 		return m.Scaling.Min
 	}
@@ -28,6 +28,11 @@ type Headroom struct {
 	// TokPerSec is what one more replica was measured to serve, on the
 	// profile placement would give it; zero when it was never measured.
 	TokPerSec float64
+	// ByPreempting is how many more, beyond More, would fit by stopping
+	// replicas the model's class may take, and Preempts the models those
+	// replicas belong to: what "room" means for a model that outranks others.
+	ByPreempting int
+	Preempts     []string
 }

 // Headroom tries placing up to limit more replicas of a model on a copy of
@@ -61,12 +66,38 @@ func (r *Reconciler) Headroom(model string, limit int) (Headroom, bool) {
 	if pol != nil {
 		usage = pol.Usage(r.Holdings())
 	}
-	for h.More < limit {
+	place := func() placement.Result {
 		res := sched.Place(r.orderedProfiles(m), avoid)
 		if res.Placement == nil && len(avoid) > 0 {
 			// As scaleUp does: spreading is a preference, not a rule.
 			res = sched.Place(r.orderedProfiles(m), nil)
 		}
+		return res
+	}
+	// What the model's class may take, in the order preemption would, each
+	// released on the copy only when free room has run out.
+	victims, nextVictim, preempting := r.victims(m), 0, false
+	for h.More+h.ByPreempting < limit {
+		res := place()
+		// Named only once they have made room: candidates tried on the way
+		// to finding there is none would otherwise read as whose replicas a
+		// placement would stop.
+		var tried []string
+		for res.Placement == nil && nextVictim < len(victims) {
+			v := victims[nextVictim]
+			nextVictim++
+			inv.ReleaseReplica(v.ID)
+			tried = append(tried, v.Model)
+			res = place()
+		}
+		if res.Placement != nil && len(tried) > 0 {
+			preempting = true
+			for _, m := range tried {
+				if !slices.Contains(h.Preempts, m) {
+					h.Preempts = append(h.Preempts, m)
+				}
+			}
+		}
 		if res.Placement == nil {
 			for _, rej := range res.Rejections {
 				h.Blocked = append(h.Blocked, rej.String())
@@ -103,8 +134,14 @@ func (r *Reconciler) Headroom(model string, limit int) (Headroom, bool) {
 			nodes = append(nodes, a.NodeID)
 		}
 		avoid = append(avoid, nodes...)
-		h.Nodes = append(h.Nodes, nodes)
-		h.More++
+		// Once anything was stopped, what fits after is counted as fitting
+		// by preempting: it may be sitting in the room that was freed.
+		if preempting {
+			h.ByPreempting++
+		} else {
+			h.Nodes = append(h.Nodes, nodes)
+			h.More++
+		}
 	}
 	h.Limited = true
 	return h, true
@@ -312,3 +312,107 @@ func TestOfTwoThatWouldEachMakeRoomTheSmallerIsStopped(t *testing.T) {
 		}
 	}
 }
+
+func TestHeadroomCountsWhatAHigherClassCouldTake(t *testing.T) {
+	h := fillThenWant(t, gpuModel("vital", catalog.ClassCritical, 1), gpuModel("chat", "", 1), gpuNode("1", 8))
+	got, _ := h.r.Headroom("vital", 3)
+	if got.More != 0 || got.ByPreempting != 1 || len(got.Preempts) != 1 || got.Preempts[0] != "chat" {
+		t.Errorf("a critical model should see one more by stopping chat: %+v", got)
+	}
+	// Tried on a copy: nothing was stopped.
+	if replicasOf(h, "chat") != 1 {
+		t.Fatal("asking about room stopped a replica")
+	}
+	// A model of chat's own class may not take its minimum, so has no such room.
+	peer := newFleetHarness(t, []catalog.Model{gpuModel("chat", "", 1), withMin(gpuModel("peer", "", 1), 0)}, gpuNode("1", 8))
+	peer.settle(t, 3)
+	peer.advance(peer.r.cfg.MinLifetime + time.Minute)
+	if got, _ := peer.r.Headroom("peer", 3); got.ByPreempting != 0 {
+		t.Errorf("a standard model counted room in another's minimum: %+v", got)
+	}
+}
+
+func TestARestartBorrowsALowerClassesRoomAndGivesItBack(t *testing.T) {
+	// chat fills node 1; the only other node holds an opportunistic replica.
+	chat := gpuModel("chat", "", 1)
+	extra := gpuModel("extra", catalog.ClassOpportunistic, 1)
+	h := newFleetHarness(t, []catalog.Model{chat, extra}, gpuNode("1", 8), gpuNode("2", 8))
+	h.settle(t, 3)
+	h.advance(h.r.cfg.MinLifetime + time.Minute)
+	var old string
+	for _, rep := range h.r.Replicas() {
+		if rep.Model == "chat" {
+			old = rep.ID
+		}
+	}
+	if old == "" || replicasOf(h, "extra") != 1 {
+		t.Fatalf("setup: %+v", h.r.Replicas())
+	}
+
+	if err := h.r.Restart(old); err != nil {
+		t.Fatal(err)
+	}
+	// Replacement first, beside the old one, in the room extra gives up;
+	// the old one keeps serving until then.
+	servedThroughout := true
+	for range 8 {
+		h.r.Reconcile(t.Context())
+		h.markReady()
+		ready := 0
+		for _, rep := range h.r.Replicas() {
+			if rep.Model == "chat" && rep.State == Ready {
+				ready++
+			}
+		}
+		if ready == 0 {
+			servedThroughout = false
+		}
+		h.advance(h.r.cfg.PreemptGrace + time.Second)
+	}
+	if !servedThroughout {
+		t.Error("chat had no ready replica during its restart")
+	}
+	var newChat bool
+	for _, rep := range h.r.Replicas() {
+		if rep.Model == "chat" && rep.ID != old && rep.State == Ready {
+			newChat = true
+		}
+		if rep.ID == old {
+			t.Errorf("the old replica is still there: %+v", rep)
+		}
+	}
+	if !newChat {
+		t.Fatalf("no new chat replica: %+v", h.r.Replicas())
+	}
+	// And extra has its replica back, in the room the old chat left.
+	h.settle(t, 4)
+	if replicasOf(h, "extra") != 1 {
+		t.Errorf("the opportunistic model did not get its room back: %+v", h.r.Replicas())
+	}
+}
+
+func TestHeadroomNamesOnlyTheReplicasThatWouldMakeRoom(t *testing.T) {
+	// vital needs a whole node and chat is on the only one; a second
+	// critical can never fit. Asking for 3: one by stopping chat, then no
+	// candidates left — nothing else may be named.
+	h := fillThenWant(t, gpuModel("vital", catalog.ClassCritical, 1), gpuModel("chat", "", 1), gpuNode("1", 8))
+	got, _ := h.r.Headroom("vital", 3)
+	if got.ByPreempting != 1 || len(got.Preempts) != 1 {
+		t.Errorf("%+v", got)
+	}
+	// Free room for one, and candidates that cannot make room for a second:
+	// none of them may be listed.
+	small := catalog.Model{Name: "small", Source: "/m", Class: catalog.ClassOpportunistic,
+		Scaling:  catalog.Scaling{Min: 1, Max: 1},
+		Profiles: []catalog.Profile{{Name: "tp2", Engine: "vllm", GPUClass: "h100-80g", TP: 2}}}
+	h2 := fillThenWant(t, gpuModel("big", "", 1), small, gpuNode("1", 8), gpuNode("2", 8))
+	got2, _ := h2.r.Headroom("big", 3)
+	if got2.More != 1 || got2.ByPreempting != 1 || len(got2.Preempts) != 1 {
+		t.Errorf("one free, one by stopping small: %+v", got2)
+	}
+	h3 := fillThenWant(t, gpuModel("big", "", 1), small, gpuNode("1", 8), gpuNode("2", 6))
+	got3, _ := h3.r.Headroom("big", 3)
+	if got3.ByPreempting != 0 || len(got3.Preempts) != 0 {
+		t.Errorf("stopping small frees a 6-GPU node, too small for big; nobody may be named: %+v", got3)
+	}
+}
@@ -255,6 +255,11 @@ type Reconciler struct {
 	// restarted and not yet begun. One is replaced at a time: replacing them
 	// all at once would need the memory of every replica twice over.
 	restartQueue map[string][]string
+	// restartInPlace are the replicas asked to restart by stopping first.
+	restartInPlace map[string]bool
+	// restartRefused says, by replica, why a restart was given up rather
+	// than leave the model without it.
+	restartRefused map[string]string
 	// parked holds models that failed ParkAfter times in a row.
 	parked map[string]parkRecord
 	// extra is how many replicas load has added above a model's base, and
@@ -304,6 +309,8 @@ func New(cfg Config, cat *catalog.Live, inv *inventory.Store, disp Dispatcher, l
 		reservedPorts:   map[int]struct{}{},
 		wanted:          map[string]time.Time{},
 		restartQueue:    map[string][]string{},
+		restartInPlace:  map[string]bool{},
+		restartRefused:  map[string]string{},
 		parked:          map[string]parkRecord{},
 		extra:           map[string]int{},
 		busySince:       map[string]time.Time{},
@@ -729,6 +736,7 @@ func (r *Reconciler) reconcileModel(ctx context.Context, m catalog.Model) {
 	r.mu.Unlock()

 	desired := r.desiredFor(m)
+	outcome := notTried
 	switch {
 	case len(liveIDs) < desired && isParked:
 		// Not placed. Its failure is on every status the operator reads.
@@ -742,7 +750,7 @@ func (r *Reconciler) reconcileModel(ctx context.Context, m catalog.Model) {
 			// the real error in noise.
 			return
 		}
-		r.scaleUp(ctx, m, avoid)
+		outcome = r.scaleUp(ctx, m, avoid)
 	case len(liveIDs) > desired:
 		r.scaleDown(ctx, liveIDs[desired:])
 	}
@@ -781,7 +789,7 @@ func (r *Reconciler) reconcileModel(ctx context.Context, m catalog.Model) {
 		}
 	}

-	r.retire(ctx, retiringIDs, readyLive, len(liveIDs), desired)
+	r.retire(ctx, m, retiringIDs, readyLive, len(liveIDs), desired, outcome, isParked)
 }

 // scaleUp places and starts one replica per pass.
@@ -789,7 +797,7 @@ func (r *Reconciler) reconcileModel(ctx context.Context, m catalog.Model) {
 // One at a time on purpose: placement decisions depend on what is already
 // allocated, and starting several at once from the same inventory snapshot
 // would hand them the same GPUs.
-func (r *Reconciler) scaleUp(ctx context.Context, m catalog.Model, avoid []string) {
+func (r *Reconciler) scaleUp(ctx context.Context, m catalog.Model, avoid []string) scaleOutcome {
 	profiles := r.orderedProfiles(m)
 	res := r.sched.Place(profiles, avoid)
 	if res.Placement == nil && len(avoid) > 0 {
@@ -813,13 +821,13 @@ func (r *Reconciler) scaleUp(ctx context.Context, m catalog.Model, avoid []strin
 		// the pass that reaps it, and the next pass places this model — which
 		// keeps the stop-then-start ordering in one place rather than two.
 		if m.Scaling.OnDemand() && r.makeRoomFor(ctx, m) {
-			return
+			return makingRoom
 		}
 		if r.awaitingPreemption(m.Name) || r.preemptFor(ctx, m, avoid) {
-			return
+			return makingRoom
 		}
 		r.notePlacementFailure(m.Name, res.Explain())
-		return
+		return noRoom
 	}
 	if why := r.overLimit(m, res.Placement); why != "" {
 		// Recorded as the placement's answer, so the model reads as not
@@ -829,13 +837,26 @@ func (r *Reconciler) scaleUp(ctx context.Context, m catalog.Model, avoid []strin
 			{Profile: res.Placement.Profile.Name, Reason: why}}}
 		r.mu.Unlock()
 		r.notePlacementFailure(m.Name, why)
-		return
+		return overLimit
 	}
 	if err := r.startReplica(ctx, m, res.Placement); err != nil {
 		r.log.Error("starting replica failed", "model", m.Name, "err", err)
+		return noRoom
 	}
+	return placed
 }

+// scaleOutcome is what one scaleUp did.
+type scaleOutcome int
+
+const (
+	notTried   scaleOutcome = iota // nothing was asked of placement this pass
+	placed                         // a replica was started
+	makingRoom                     // room is being made: a preemption, an idle replica stopping
+	noRoom                         // placement found nowhere, and nothing is making room
+	overLimit                      // there is room, but the owner's quota does not allow it
+)
+
 func (r *Reconciler) startReplica(ctx context.Context, m catalog.Model, pl *placement.Placement) error {
 	prof, ok := m.ProfileByName(pl.Profile.Name)
 	if !ok {
@@ -114,6 +114,17 @@ type harness struct {
 	inv   *inventory.Store
 	disp  *fakeDispatcher
 	clock *clock
+	live  *catalog.Live
+}
+
+// setModels replaces the catalog the reconciler works from, as an edit does.
+func (h *harness) setModels(t *testing.T, models ...catalog.Model) {
+	t.Helper()
+	cat, err := catalog.New(models)
+	if err != nil {
+		t.Fatal(err)
+	}
+	h.live.Store(cat)
 }

 func newHarness(t *testing.T, m catalog.Model, nodes ...inventory.Node) *harness {
@@ -135,7 +146,8 @@ func newHarness(t *testing.T, m catalog.Model, nodes ...inventory.Node) *harness
 	cfg.UnreachableAfter = 15 * time.Second
 	cfg.LostAfter = 60 * time.Second

-	r := New(cfg, catalog.NewLive(cat), inv, disp, slog.New(slog.DiscardHandler))
+	live := catalog.NewLive(cat)
+	r := New(cfg, live, inv, disp, slog.New(slog.DiscardHandler))
 	r.SetClock(c.now)
 	// Nodes were just registered, so their LastSeen is wall-clock now, which is
 	// far from the fake clock's epoch. Re-stamp them onto the fake timeline.
@@ -150,7 +162,7 @@ func newHarness(t *testing.T, m catalog.Model, nodes ...inventory.Node) *harness
 	for _, n := range nodes {
 		touchAt(inv, n.ID, c.now())
 	}
-	return &harness{r: r, inv: inv, disp: disp, clock: c}
+	return &harness{r: r, inv: inv, disp: disp, clock: c, live: live}
 }

 // touchAt rewrites a node's LastSeen onto the fake clock's timeline.
@@ -1787,7 +1799,9 @@ func TestARestartedReplicaServesUntilItsReplacementIsReady(t *testing.T) {

 func TestARestartWithNoRoomBesideItHappensInPlace(t *testing.T) {
 	// One node, room for one replica: the replacement can only start once the
-	// old one has gone. Waiting forever would mean never restarting at all.
+	// old one has gone. It is the only thing in the way, so it goes now —
+	// waiting out a grace period first changed nothing but the wait, which
+	// in the field read as a restart that never happened.
 	h := newHarness(t, testModel(1), gpuNode("1", 4))
 	h.r.Reconcile(t.Context())
 	h.markReady()
@@ -1796,14 +1810,8 @@ func TestARestartWithNoRoomBesideItHappensInPlace(t *testing.T) {
 		t.Fatal(err)
 	}
 	h.r.Reconcile(t.Context())
-	if rep, _ := replicaByID(h, old); rep.State != Ready {
-		t.Fatalf("stopped before the grace period: %s", rep.State)
-	}
-
-	h.advance(h.r.cfg.DrainGrace + time.Second)
-	h.r.Reconcile(t.Context())
 	if rep, ok := replicaByID(h, old); ok && rep.State == Ready {
-		t.Fatal("still serving after the grace period with no replacement possible")
+		t.Fatal("still serving with no replacement possible beside it")
 	}
 	for i := 0; i < 3; i++ {
 		h.r.Reconcile(t.Context())
@@ -1815,6 +1823,52 @@ func TestARestartWithNoRoomBesideItHappensInPlace(t *testing.T) {
 	}
 }

+func TestARestartWhoseReplacementFitsNowhereIsGivenUp(t *testing.T) {
+	// The profile was edited to need more cards than the node has: stopping
+	// the old replica would leave the model with nothing at all.
+	m := testModel(1)
+	h := newHarness(t, m, gpuNode("1", 4))
+	h.r.Reconcile(t.Context())
+	h.markReady()
+	old := h.r.ReadyReplicas()[0].ID
+
+	m.Profiles[0].TP, m.Profiles[0].GPUsPerNode = 8, 8
+	h.setModels(t, m)
+	if err := h.r.Restart(old); err != nil {
+		t.Fatal(err)
+	}
+	for range 3 {
+		h.r.Reconcile(t.Context())
+		h.advance(h.r.cfg.DrainGrace + time.Second)
+	}
+	if rep, ok := replicaByID(h, old); !ok || rep.State != Ready {
+		t.Fatalf("the only replica was stopped for a replacement that fits nowhere: %+v", rep)
+	}
+	if st := h.r.RestartState(old); !strings.HasPrefix(st, "refused: ") || !strings.Contains(st, "fits on no node") {
+		t.Errorf("restart state %q should say it was given up, and why", st)
+	}
+	// Asked again, it is tried again.
+	if err := h.r.Restart(old); err != nil || h.r.RestartState(old) != "queued" {
+		t.Errorf("a new restart did not clear the refusal: %q, %v", h.r.RestartState(old), err)
+	}
+}
+
+func TestARestartAskedInPlaceStopsFirst(t *testing.T) {
+	// Room beside it, but in place was asked for: an engine that must not
+	// run twice, or memory somebody knows is short.
+	h := newHarness(t, testModel(1), gpuNode("1", 4), gpuNode("2", 4))
+	h.r.Reconcile(t.Context())
+	h.markReady()
+	old := h.r.ReadyReplicas()[0].ID
+	if err := h.r.RestartInPlace(old); err != nil {
+		t.Fatal(err)
+	}
+	h.r.Reconcile(t.Context())
+	if rep, ok := replicaByID(h, old); ok && rep.State == Ready {
+		t.Fatalf("asked in place, it kept serving beside its replacement: %+v", rep)
+	}
+}
+
 func TestReplicasOfAModelAreRestartedOneAtATime(t *testing.T) {
 	h := newHarness(t, testModel(2), gpuNode("1", 8), gpuNode("2", 8), gpuNode("3", 8))
 	for i := 0; i < 3; i++ {
@@ -6,17 +6,34 @@ import (
 	"fmt"
 	"slices"
 	"time"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/catalog"
 )

 // retire finishes restarting replicas.
 //
 // A retiring replica keeps serving until its replacement is ready, and is then
-// drained like any other. When no replacement could be placed at all —
-// nowhere else has the memory, which on a small CPU fleet is the usual case —
-// it is stopped after DrainGrace so its replacement can take its place: a
-// restart in place, with the model unserved while it loads if it had no other
-// replica. Said in the log, since that is the one case with an outage.
-func (r *Reconciler) retire(ctx context.Context, ids []string, readyLive, live, desired int) {
+// drained like any other. When no replacement could be placed beside it, what
+// happens depends on what is in the way:
+//
+//   - the replica itself — one node, or every card that fits held: it is
+//     stopped now, and its replacement takes its place. Waiting would change
+//     nothing; it used to wait DrainGrace anyway, five minutes reading only
+//     "restart replacing", which looked like a restart that never would.
+//   - more than it: the replacement fits nowhere even with it stopped — its
+//     profile was edited to need more than any node has free. Stopping it
+//     would leave the model with nothing, so the restart is given up, and
+//     says why where the replica is shown.
+//   - something being moved out of the way (a preemption, an idle model
+//     stopping): waited for.
+//   - nothing tried this pass (a backoff): the old rule, stopped after
+//     DrainGrace with no replacement in sight.
+//
+// A restart asked for in place (ctl restart --in-place) stops first, always.
+// Each stop is said in the log, since it is the one case with an outage.
+func (r *Reconciler) retire(ctx context.Context, m catalog.Model, ids []string, readyLive, live, desired int,
+	outcome scaleOutcome, parked bool,
+) {
 	for _, id := range ids {
 		r.mu.Lock()
 		rep, ok := r.replicas[id]
@@ -24,6 +41,7 @@ func (r *Reconciler) retire(ctx context.Context, ids []string, readyLive, live,
 		if ok {
 			since = rep.RetiringAt
 		}
+		inPlace := r.restartInPlace[id]
 		r.mu.Unlock()
 		if !ok {
 			continue
@@ -31,15 +49,48 @@ func (r *Reconciler) retire(ctx context.Context, ids []string, readyLive, live,
 		switch {
 		case readyLive >= desired:
 			r.drainForRestart(ctx, id)
-		case live < desired && r.now().Sub(since) > r.cfg.DrainGrace:
-			r.log.Warn("restarting a replica in place: no room for its replacement beside it",
+		case inPlace:
+			r.log.Warn("restarting a replica in place, as asked: stopped before its replacement starts", "replica", id)
+			r.stopForRestart(ctx, id)
+		case live >= desired:
+			// A replacement is placed and loading; it is waited for.
+		case parked:
+			r.refuseRestart(id, "its replacements fail to start; the model is parked (llmgw ctl retry "+m.Name+")")
+		case outcome == noRoom && r.fitsWithout(m, nil, []victim{{ID: id, Model: m.Name}}):
+			r.log.Warn("restarting a replica in place: it is the only thing in its replacement's way",
+				"replica", id, "model", m.Name)
+			r.stopForRestart(ctx, id)
+		case outcome == noRoom || outcome == overLimit:
+			r.refuseRestart(id, "its replacement fits on no node even with this replica stopped "+
+				"(llmgw ctl explain-placement "+m.Name+")")
+		case outcome == notTried && r.now().Sub(since) > r.cfg.DrainGrace:
+			r.log.Warn("restarting a replica in place: no replacement beside it",
 				"replica", id, "after", r.cfg.DrainGrace)
-			r.drainForRestart(ctx, id)
-			r.stopReplica(ctx, id)
+			r.stopForRestart(ctx, id)
 		}
 	}
 }

+// stopForRestart drains a replica and stops it at once, for its replacement
+// to take its place.
+func (r *Reconciler) stopForRestart(ctx context.Context, id string) {
+	r.drainForRestart(ctx, id)
+	r.stopReplica(ctx, id)
+}
+
+// refuseRestart gives up a replica's restart, leaving it serving, and keeps
+// why for whoever looks at the replica.
+func (r *Reconciler) refuseRestart(id, why string) {
+	r.mu.Lock()
+	if rep, ok := r.replicas[id]; ok {
+		rep.RetiringAt = time.Time{}
+	}
+	delete(r.restartInPlace, id)
+	r.restartRefused[id] = why
+	r.mu.Unlock()
+	r.log.Warn("not restarting a replica: it keeps serving", "replica", id, "because", why)
+}
+
 func (r *Reconciler) drainForRestart(ctx context.Context, id string) {
 	r.mu.Lock()
 	rep, ok := r.replicas[id]
@@ -48,6 +99,7 @@ func (r *Reconciler) drainForRestart(ctx context.Context, id string) {
 		rep.State, rep.DrainingAt, rep.RetiringAt = Draining, r.now(), time.Time{}
 		shards = slices.Clone(rep.Shards)
 	}
+	delete(r.restartInPlace, id)
 	r.mu.Unlock()
 	for _, s := range shards {
 		if err := r.disp.DrainShard(ctx, s.NodeID, s.ID); err != nil {
@@ -78,7 +130,14 @@ var ErrNoSuchReplica = errors.New("no ready replica with that id")
 // Restart queues replicas to be replaced, each by a new one started beside it
 // before it is stopped — to pick up an edited profile, an engine key, a new
 // engine version. Replicas of one model are replaced one at a time.
-func (r *Reconciler) Restart(ids ...string) error {
+func (r *Reconciler) Restart(ids ...string) error { return r.restart(false, ids) }
+
+// RestartInPlace queues replicas to be replaced by stopping each first and
+// then starting its replacement: for when there is knowingly no room beside
+// it, and the outage while it loads is accepted.
+func (r *Reconciler) RestartInPlace(ids ...string) error { return r.restart(true, ids) }
+
+func (r *Reconciler) restart(inPlace bool, ids []string) error {
 	r.mu.Lock()
 	for _, id := range ids {
 		rep, ok := r.replicas[id]
@@ -86,6 +145,10 @@ func (r *Reconciler) Restart(ids ...string) error {
 			r.mu.Unlock()
 			return fmt.Errorf("%w: %s", ErrNoSuchReplica, id)
 		}
+		delete(r.restartRefused, id)
+		if inPlace {
+			r.restartInPlace[id] = true
+		}
 		if !slices.Contains(r.restartQueue[rep.Model], id) && rep.RetiringAt.IsZero() {
 			r.restartQueue[rep.Model] = append(r.restartQueue[rep.Model], id)
 		}
@@ -95,8 +158,8 @@ func (r *Reconciler) Restart(ids ...string) error {
 	return nil
 }

-// RestartState says where a replica's restart stands: "queued", "replacing",
-// or "" when none was asked for.
+// RestartState says where a replica's restart stands: "queued",
+// "replacing", "in place", "refused: why", or "" when none was asked for.
 func (r *Reconciler) RestartState(id string) string {
 	r.mu.Lock()
 	defer r.mu.Unlock()
@@ -104,10 +167,14 @@ func (r *Reconciler) RestartState(id string) string {
 	switch {
 	case !ok:
 		return ""
+	case !rep.RetiringAt.IsZero() && r.restartInPlace[id]:
+		return "in place"
 	case !rep.RetiringAt.IsZero():
 		return "replacing"
 	case slices.Contains(r.restartQueue[rep.Model], id):
 		return "queued"
+	case r.restartRefused[id] != "":
+		return "refused: " + r.restartRefused[id]
 	}
 	return ""
 }
@@ -10,6 +10,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"maps"
 	"net"
 	"os"
 	"os/exec"
@@ -86,10 +87,12 @@ type Facts struct {

 	// What installing vLLM here would meet: the version this llmgw installs,
 	// where, the space there, and whether it already has.
-	VLLMVersion   string  `json:"vllm_version,omitempty"`
-	EnginesDir    string  `json:"engines_dir,omitempty"`
-	EnginesFreeGB float64 `json:"engines_free_gb,omitempty"`
-	VLLMInstalled bool    `json:"vllm_installed"`
+	VLLMVersion string `json:"vllm_version,omitempty"`
+	// DisabledEngines are those agent.yaml says this node never runs.
+	DisabledEngines []string `json:"disabled_engines,omitempty"`
+	EnginesDir      string   `json:"engines_dir,omitempty"`
+	EnginesFreeGB   float64  `json:"engines_free_gb,omitempty"`
+	VLLMInstalled   bool     `json:"vllm_installed"`
 }

 // Modules checked, and why each matters.
@@ -103,6 +106,10 @@ type Options struct {
 	// VLLMVersion is the vLLM this llmgw installs; its prerequisites are
 	// checked when set.
 	VLLMVersion string
+	// Fabrics are fabrics about to be declared, by name, with their
+	// adapters: what `node add --fabric` will write into agent.yaml, which
+	// does not exist yet when doctor runs before it.
+	Fabrics map[string][]string
 }

 // Collect reads the node.
@@ -183,6 +190,14 @@ func Collect(ctx context.Context, o Options) Facts {
 	}

 	f.readConfig(o)
+	if len(o.Fabrics) > 0 {
+		declared := map[string][]string{}
+		for _, c := range f.Fabrics {
+			declared[c.Fabric] = c.Declared
+		}
+		maps.Copy(declared, o.Fabrics)
+		f.Fabrics = discovery.CheckFabrics(declared, f.HCAs)
+	}
 	f.Host.MinComputeCapability = discovery.MinComputeCapability(f.GPUs)
 	if f.VLLMVersion = o.VLLMVersion; f.VLLMVersion != "" {
 		if f.EnginesDir == "" {
@@ -238,6 +253,7 @@ func (f *Facts) readConfig(o Options) {
 	_, err = net.InterfaceByName(f.MgmtIface)
 	f.IfaceExists = err == nil
 	f.EnginesDir = cfg.EnginesDir
+	f.DisabledEngines = cfg.Engines.Disabled
 	f.LaunchMode = cfg.Engines.VLLM.LaunchMode
 	if f.LaunchMode == "" {
 		f.LaunchMode = "mp"
@@ -419,8 +435,15 @@ func Evaluate(f Facts) []Finding {
 			}
 			switch {
 			case p.Active() && p.LinkLayer == discovery.LinkEthernet && p.RoCEv2GID < 0:
-				add(Fail, "fabric", where+" is RoCE with no RoCE v2 GID entry: NCCL would use a v1 entry and hang",
-					"give the interface an IPv4 address, or check the adapter's RoCE mode")
+				// Fatal only for a port a declared fabric uses. An adapter
+				// nothing is configured to use — a RoCE NIC on a node added
+				// for single-node inference — kept nodes out for a fabric
+				// nobody asked for.
+				msg := where + " is RoCE with no RoCE v2 GID entry: NCCL would use a v1 entry and hang"
+				if declared[h.Name] == "" {
+					msg += "; no declared fabric uses it, so nothing will until one does"
+				}
+				add(level, "fabric", msg, "give the interface an IPv4 address, or check the adapter's RoCE mode")
 			case p.Active():
 				on := ""
 				switch {
@@ -430,17 +453,24 @@ func Evaluate(f Facts) []Finding {
 					on = ", network " + p.Network
 				}
 				add(OK, "fabric", fmt.Sprintf("%s %s at %d Gb/s (%s)%s", where, p.State, p.RateGbps, p.LinkLayer, on), "")
+				// Questions about which fabric a port is on matter to the
+				// fabrics that use it; for a port none does they are noted,
+				// not warned of.
+				quiet := Warn
+				if declared[h.Name] == "" {
+					quiet = Info
+				}
 				if p.LinkLayer == discovery.LinkInfiniBand && p.SM == nil {
 					why := f.SMErrors[where]
 					fix := ""
 					if strings.Contains(why, "infiniband-diags") {
 						fix = "install infiniband-diags"
 					}
-					add(Warn, "fabric", where+": which fabric it is on cannot be told — "+orUnknown(why)+
+					add(quiet, "fabric", where+": which fabric it is on cannot be told — "+orUnknown(why)+
 						"; without it, separate fabrics given one name cannot be caught", fix)
 				}
 				if p.LinkLayer == discovery.LinkInfiniBand && p.SMLID == 0 {
-					add(Warn, "fabric", where+" is active but reports no subnet manager LID", "")
+					add(quiet, "fabric", where+" is active but reports no subnet manager LID", "")
 				}
 			case strings.EqualFold(p.PhysState, "LinkUp"):
 				add(level, "fabric", where+" has a link but is "+p.State+": no subnet manager has configured it",
@@ -479,7 +509,14 @@ func Evaluate(f Facts) []Finding {
 	}
 	if len(f.HCAs) > 0 {
 		if !f.Modules["ib_uverbs"] {
-			add(Fail, "fabric", "ib_uverbs is not loaded: nothing in user space can use RDMA", "modprobe ib_uverbs")
+			// RDMA is for replicas spanning nodes; a node that declares no
+			// fabric runs none, and is a node without it.
+			if len(declared) > 0 {
+				add(Fail, "fabric", "ib_uverbs is not loaded: nothing in user space can use RDMA", "modprobe ib_uverbs")
+			} else {
+				add(Warn, "fabric", "ib_uverbs is not loaded: nothing in user space can use RDMA, which a fabric declared "+
+					"later would need; this node runs single-node replicas without it", "modprobe ib_uverbs")
+			}
 		}
 		if len(f.GPUs) > 0 && !f.Modules["nvidia_peermem"] && !f.Modules["nv_peer_mem"] {
 			add(Warn, "fabric", "nvidia_peermem is not loaded: GPUDirect RDMA is off, and NCCL copies every transfer through host memory",
@@ -525,7 +562,9 @@ func Evaluate(f Facts) []Finding {
 		}
 	}
 	// --- engines ---
-	if f.VLLMVersion != "" {
+	// An engine disabled on this node: whether it could be installed is not
+	// a question anybody is asking.
+	if f.VLLMVersion != "" && !slices.Contains(f.DisabledEngines, "vllm") {
 		vllmFindings(f, add)
 	}

@@ -542,28 +581,44 @@ func Evaluate(f Facts) []Finding {
 // the same decision the agent makes, so the report and the node agree.
 func vllmFindings(f Facts, add func(Level, string, string, string)) {
 	v := f.VLLMVersion
-	b, err := provision.Recipes["vllm"].Choose(v, f.Host)
+	builds, err := provision.Recipes["vllm"].Candidates(v, "", f.Host)
+	var b provision.Build
+	if err == nil {
+		// The first the agent would try; it moves to the next by itself if
+		// that one does not work here.
+		b = builds[0]
+		if b.Version != "" {
+			v = b.Version
+		}
+	}
 	var nh provision.NotHere
 	switch {
 	case errors.As(err, &nh):
 		add(Info, "engines", "vLLM will not be installed: "+nh.Reason, "")
 		return
 	case err != nil:
-		add(Fail, "engines", "vLLM "+v+" cannot be installed: "+err.Error(),
-			"upgrade the NVIDIA driver, or set the control plane's engines.vllm to a version for this one")
+		// A warning, not a failure: the node is still a node, and serves
+		// llama.cpp. Refusing it here kept older GPU generations out of a
+		// cluster for an engine nobody meant to run on them.
+		add(Warn, "engines", "vLLM "+v+" will not run here, and the node joins without it: "+err.Error(),
+			"add it with --disable-engine vllm if it is meant for llama.cpp alone")
 		return
 	case f.VLLMInstalled:
 		add(OK, "engines", "vLLM "+v+" is installed under "+f.EnginesDir, "")
 		return
 	}
 	if f.EnginesFreeGB > 0 && f.EnginesFreeGB < float64(b.FreeGB) {
-		add(Fail, "engines", fmt.Sprintf("vLLM %s needs about %d GB free under %s and there are %.1f GB",
+		add(Warn, "engines", fmt.Sprintf("vLLM %s needs about %d GB free under %s and there are %.1f GB; the node joins without it until there is",
 			v, b.FreeGB, f.EnginesDir, f.EnginesFreeGB),
 			"free space there, or point engines_dir in agent.yaml at a larger disk")
 		return
 	}
-	add(OK, "engines", fmt.Sprintf("vLLM %s (%s) can be installed here: driver %s, %.1f GB free under %s",
-		v, b.Variant, f.Host.NVIDIADriver, f.EnginesFreeGB, f.EnginesDir), "")
+	note := ""
+	if b.Note != "" {
+		note = "; " + b.Note
+	}
+	add(OK, "engines", fmt.Sprintf("vLLM %s (%s) can be installed here: driver %s, %.1f GB free under %s%s",
+		v, b.Variant, f.Host.NVIDIADriver, f.EnginesFreeGB, f.EnginesDir, note), "")
 }

 func rank(l Level) int {
@@ -151,8 +151,21 @@ func TestDoctorSaysWhetherVLLMCanBeInstalled(t *testing.T) {
 		says  string
 	}{
 		{"CPU node", Facts{Host: discovery.Host{Arch: "amd64"}, VLLMVersion: "0.30.0"}, Info, "has none"},
-		{"old driver", Facts{Host: discovery.Host{Arch: "amd64", NVIDIADriver: "550.54.15"}, VLLMVersion: "0.30.0"}, Fail, "driver 580"},
-		{"small disk", Facts{Host: gpu, VLLMVersion: "0.30.0", EnginesDir: "/opt/llmgw/engines", EnginesFreeGB: 9}, Fail, "9.0 GB"},
+		// What vLLM cannot do keeps vLLM off the node, not the node out of
+		// the cluster: it still serves llama.cpp.
+		{"a driver too old for either build", Facts{Host: discovery.Host{Arch: "amd64", NVIDIADriver: "470.256.02"}, VLLMVersion: "0.30.0"}, Warn, "driver 525"},
+		// A driver too old for CUDA 13 runs the same version's CUDA 12.9 build.
+		{"turing, old driver", Facts{Host: discovery.Host{Arch: "amd64", NVIDIADriver: "550.54.15", MinComputeCapability: 7.5},
+			VLLMVersion: "0.30.0", EnginesDir: "/opt/llmgw/engines", EnginesFreeGB: 200}, OK, "(cu129) can be installed"},
+		// A V100 has no build of vLLM 0.20 or later; it gets the newest
+		// that has Volta kernels, and doctor says which.
+		{"volta", Facts{Host: discovery.Host{Arch: "amd64", NVIDIADriver: "550.54.15", MinComputeCapability: 7.0},
+			VLLMVersion: "0.30.0", EnginesDir: "/opt/llmgw/engines", EnginesFreeGB: 200}, OK, "vLLM 0.19.1 (cu12) can be installed here"},
+		{"pascal", Facts{Host: discovery.Host{Arch: "amd64", NVIDIADriver: "580.65.06", MinComputeCapability: 6.1},
+			VLLMVersion: "0.30.0"}, Warn, "no vLLM is built for GPUs of compute capability 6.1"},
+		{"volta, the last vLLM built for it", Facts{Host: discovery.Host{Arch: "amd64", NVIDIADriver: "550.54.15", MinComputeCapability: 7.0},
+			VLLMVersion: "0.19.1", EnginesDir: "/opt/llmgw/engines", EnginesFreeGB: 200}, OK, "(cu12) can be installed"},
+		{"small disk", Facts{Host: gpu, VLLMVersion: "0.30.0", EnginesDir: "/opt/llmgw/engines", EnginesFreeGB: 9}, Warn, "9.0 GB"},
 		{"ready to install", Facts{Host: gpu, VLLMVersion: "0.30.0", EnginesDir: "/opt/llmgw/engines", EnginesFreeGB: 200}, OK, "can be installed"},
 		{"installed", Facts{Host: gpu, VLLMVersion: "0.30.0", EnginesDir: "/opt/llmgw/engines", VLLMInstalled: true}, OK, "is installed"},
 	}
@@ -200,3 +213,38 @@ func TestDoctorSaysWhenTheCPUHidesAVX2(t *testing.T) {
 		t.Errorf("ARM: %+v", f)
 	}
 }
+
+func TestAnEngineDisabledOnTheNodeIsNotChecked(t *testing.T) {
+	// The field: a V100 node meant for llama.cpp, refused for a vLLM it was
+	// never going to run.
+	f := Facts{Host: discovery.Host{Arch: "amd64", NVIDIADriver: "580.65.06", MinComputeCapability: 7.0},
+		VLLMVersion: "0.30.0", DisabledEngines: []string{"vllm"}}
+	for _, fd := range Evaluate(f) {
+		if fd.Area == "engines" {
+			t.Errorf("a disabled vLLM was reported on: %s %q", fd.Level, fd.Message)
+		}
+	}
+}
+
+func TestAnAdapterNoFabricUsesDoesNotKeepTheNodeOut(t *testing.T) {
+	// The field: a RoCE NIC with no v2 GID on a node added for single-node
+	// inference, no fabric given, and the node refused for it.
+	roce := discovery.HCA{Name: "mlx5_2", Ports: []discovery.Port{{Num: 1, State: "ACTIVE",
+		PhysState: "LinkUp", RateGbps: 100, LinkLayer: discovery.LinkEthernet, RoCEv2GID: -1}}}
+	f := healthy()
+	f.HCAs = append(f.HCAs, roce)
+	f.Modules["ib_uverbs"] = false
+	f.Fabrics = nil // nothing declared
+	fs := Evaluate(f)
+	if fails := levels(fs, Fail); len(fails) > 0 {
+		t.Errorf("an undeclared adapter failed the node: %v", fails)
+	}
+	has(t, fs, Warn, "no declared fabric uses it")
+	has(t, fs, Warn, "runs single-node replicas without it")
+
+	// Declared, the same adapter and module are what the fabric needs.
+	f.Fabrics = discovery.CheckFabrics(map[string][]string{"roce-a": {"mlx5_2"}}, f.HCAs)
+	fs = Evaluate(f)
+	has(t, fs, Fail, "mlx5_2/1 is RoCE with no RoCE v2 GID entry")
+	has(t, fs, Fail, "ib_uverbs is not loaded")
+}
@@ -25,6 +25,10 @@ type Header struct {
 	Version      uint32
 	Architecture string
 	Name         string
+	// Type is general.type: "model", or "adapter" for a LoRA adapter, which
+	// is applied to a model rather than run on its own. Empty in files made
+	// before the key existed, which are models.
+	Type string

 	// The shape the KV cache is sized from, from the architecture's own keys
 	// (llama.block_count, qwen2.attention.head_count_kv, ...). Zero when the
@@ -235,6 +239,8 @@ func Read(r io.Reader) (Header, error) {
 			h.Architecture = v
 		case "general.name":
 			h.Name = v
+		case "general.type":
+			h.Type = v
 		}
 		if h.Architecture != "" {
 			switch strings.TrimPrefix(key, h.Architecture+".") {
@@ -290,6 +296,41 @@ func FromURL(ctx context.Context, client *http.Client, url string) (Header, erro
 	return Read(io.LimitReader(resp.Body, headerBytes))
 }

+// IsAdapter reports whether the file is a LoRA adapter rather than a model.
+func (h Header) IsAdapter() bool { return h.Type == "adapter" }
+
+// adapterProbe is how much of a file is read to tell an adapter from a model:
+// general.type comes before the tokenizer, which is what makes a model's
+// header megabytes long.
+const adapterProbe = 256 << 10
+
+// KindFromURL reads just enough of a GGUF file to say what it is — its type
+// and architecture — over HTTP. A header cut short by the probe is fine: the
+// keys asked for come first.
+func KindFromURL(ctx context.Context, client *http.Client, url string) (Header, error) {
+	if client == nil {
+		client = &http.Client{Timeout: 30 * time.Second}
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
+	if err != nil {
+		return Header{}, err
+	}
+	req.Header.Set("Range", fmt.Sprintf("bytes=0-%d", adapterProbe-1))
+	resp, err := client.Do(req)
+	if err != nil {
+		return Header{}, err
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK && resp.StatusCode != http.StatusPartialContent {
+		return Header{}, fmt.Errorf("%s answered %s", url, resp.Status)
+	}
+	h, err := Read(io.LimitReader(resp.Body, adapterProbe))
+	if h.Architecture != "" {
+		return h, nil // what was asked for was read, whatever came after
+	}
+	return h, err
+}
+
 // IsGGUF reports whether a source names a GGUF file, or a directory of them.
 func IsGGUF(source string) bool {
 	s := strings.ToLower(strings.TrimSuffix(source, "/"))
@@ -0,0 +1,31 @@
+package gguf
+
+import (
+	"os"
+	"testing"
+)
+
+// Run by hand against the real adapter this was written for:
+//
+//	LLMGW_LIVE_GGUF=1 go test ./internal/gguf -run LiveAdapter -v
+func TestLiveAdapterIsAnAdapter(t *testing.T) {
+	if os.Getenv("LLMGW_LIVE_GGUF") == "" {
+		t.Skip("set LLMGW_LIVE_GGUF to read a real adapter from Hugging Face")
+	}
+	h, err := KindFromURL(t.Context(), nil,
+		"https://huggingface.co/ggml-org/LoRA-Qwen2.5-Coder-7B-Instruct-F16-GGUF/resolve/main/LoRA-Qwen2.5-Coder-7B-Instruct-f16.gguf")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !h.IsAdapter() || h.Architecture != "qwen2" {
+		t.Errorf("%+v", h)
+	}
+	m, err := KindFromURL(t.Context(), nil,
+		"https://huggingface.co/bartowski/Qwen2.5-0.5B-Instruct-GGUF/resolve/main/Qwen2.5-0.5B-Instruct-Q4_K_M.gguf")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if m.IsAdapter() || m.Architecture != "qwen2" {
+		t.Errorf("a model read as %+v", m)
+	}
+}
@@ -0,0 +1,58 @@
+package gguf
+
+import (
+	"bytes"
+	"encoding/binary"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+)
+
+// adapterHeader is the start of a LoRA adapter as llama.cpp's converter
+// writes it, followed by padding that stands in for its tensors.
+func adapterHeader(kind string, pad int) []byte {
+	var b bytes.Buffer
+	w := func(v any) { _ = binary.Write(&b, binary.LittleEndian, v) }
+	str := func(s string) { w(uint64(len(s))); b.WriteString(s) }
+	b.WriteString("GGUF")
+	w(uint32(3))
+	w(uint64(0))
+	w(uint64(3))
+	str("general.architecture")
+	w(uint32(tString))
+	str("qwen2")
+	str("general.type")
+	w(uint32(tString))
+	str(kind)
+	str("adapter.type")
+	w(uint32(tString))
+	str("lora")
+	b.Write(make([]byte, pad))
+	return b.Bytes()
+}
+
+func TestAnAdapterIsToldFromAModelByItsFirstBytes(t *testing.T) {
+	var asked string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		asked = r.Header.Get("Range")
+		w.WriteHeader(http.StatusPartialContent)
+		if r.URL.Path == "/adapter.gguf" {
+			_, _ = w.Write(adapterHeader("adapter", 1<<20))
+			return
+		}
+		_, _ = w.Write(header("qwen2"))
+	}))
+	defer srv.Close()
+
+	h, err := KindFromURL(t.Context(), srv.Client(), srv.URL+"/adapter.gguf")
+	if err != nil || !h.IsAdapter() || h.Architecture != "qwen2" {
+		t.Errorf("an adapter read as %+v, %v", h, err)
+	}
+	if asked != "bytes=0-262143" {
+		t.Errorf("asked for %q: the probe should read the start only", asked)
+	}
+	m, err := KindFromURL(t.Context(), srv.Client(), srv.URL+"/model.gguf")
+	if err != nil || m.IsAdapter() || m.Architecture != "qwen2" {
+		t.Errorf("a model read as %+v, %v", m, err)
+	}
+}
@@ -156,6 +156,10 @@ type Query struct {
 	// applies it, so a page of small models is a full page.
 	MaxParams int64
 	Limit     int
+	// Filters are further registry tags every result must carry, such as
+	// "lora", or "base_model:adapter:Qwen/Qwen2.5-7B-Instruct" for the
+	// adapters made for one model.
+	Filters []string
 }

 // exploreFields is everything a listing shows. Asking for any field with
@@ -188,7 +192,10 @@ func (c *Client) Explore(ctx context.Context, q Query) ([]Model, error) {
 		v.Set("pipeline_tag", q.Task)
 	}
 	if q.GGUF {
-		v.Set("filter", "gguf")
+		v.Add("filter", "gguf")
+	}
+	for _, f := range q.Filters {
+		v.Add("filter", f)
 	}
 	if q.MaxParams > 0 {
 		v.Set("num_parameters", fmt.Sprintf("min:0,max:%d", q.MaxParams))
@@ -251,6 +258,41 @@ func (f File) IsWeights() bool {
 	return false
 }

+// companionNames are the files beside safetensors weights that an engine
+// loading the directory reads: the model's shape, its tokenizer, how it was
+// quantised, and which shard holds which tensor.
+var companionNames = map[string]bool{
+	"config.json": true, "generation_config.json": true,
+	"tokenizer.json": true, "tokenizer_config.json": true, "tokenizer.model": true,
+	"special_tokens_map.json": true, "added_tokens.json": true,
+	"vocab.json": true, "vocab.txt": true, "merges.txt": true,
+	"chat_template.jinja": true, "chat_template.json": true,
+	"preprocessor_config.json": true, "processor_config.json": true,
+	"quantize_config.json": true, "quant_config.json": true,
+}
+
+// Companions are the files of a repository an engine reads beside a
+// safetensors set: config.json and the tokenizer, the shard index. Only
+// those at the top: the directory an engine is given is one level deep.
+//
+// Never code (a .py): a model that needs its own Python to load asks the
+// engine to run what the repository ships, which is a decision for whoever
+// runs the engine, made on purpose, not a file swept in with the tokenizer.
+func Companions(files []File) []File {
+	var out []File
+	for _, f := range files {
+		name := f.Path
+		if strings.Contains(name, "/") {
+			continue
+		}
+		if companionNames[name] || strings.HasSuffix(name, ".safetensors.index.json") ||
+			strings.HasSuffix(name, ".tiktoken") {
+			out = append(out, f)
+		}
+	}
+	return out
+}
+
 // Repo is a repository's current revision and what is in it.
 type Repo struct {
 	ID string
@@ -94,6 +94,13 @@ func addWith(ctx context.Context, client *http.Client, root, source string, e En
 		_ = os.Remove(part)
 		return Added{}, err
 	}
+	// The checksum beside the file, before the file is in place: a manifest
+	// listing the file then always has its checksum to hand on, and a node
+	// can tell a file damaged on the way from the one that was mirrored.
+	if err := writeChecksum(final, sum); err != nil {
+		_ = os.Remove(part)
+		return Added{}, err
+	}
 	if err := os.Rename(part, final); err != nil {
 		return Added{}, err
 	}
@@ -146,6 +153,9 @@ func Remove(root string, e Entry) error {
 	if err := os.Remove(path); err != nil {
 		return err
 	}
+	// Its checksum goes with it; left, it would keep the directory from
+	// being empty, and so from going.
+	_ = os.Remove(checksumPath(path))
 	versionDir := filepath.Dir(path)
 	if empty, _ := isEmpty(versionDir); empty {
 		_ = os.Remove(versionDir)
@@ -16,12 +16,15 @@ package mirror
 import (
 	"crypto/sha256"
 	"encoding/hex"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
 	"os"
 	"path/filepath"
+	"regexp"
 	"sort"
+	"strconv"
 	"strings"
 )

@@ -91,7 +94,7 @@ func List(root string) ([]Entry, error) {
 				continue
 			}
 			for _, f := range files {
-				if f.IsDir() || strings.HasSuffix(f.Name(), partSuffix) {
+				if f.IsDir() || strings.HasSuffix(f.Name(), partSuffix) || strings.HasPrefix(f.Name(), ".") {
 					continue
 				}
 				info, err := f.Info()
@@ -169,6 +172,31 @@ func normalizeURL(u string) string {

 const partSuffix = ".part"

+// checksumPath is where a file's sha256 is kept: beside it, hidden, so no
+// listing takes it for a file of the model and no name a person gives a
+// file (which may not start with a dot) can be the same.
+func checksumPath(file string) string {
+	return filepath.Join(filepath.Dir(file), "."+filepath.Base(file)+".sha256")
+}
+
+func writeChecksum(file, sum string) error {
+	tmp := checksumPath(file) + partSuffix
+	if err := os.WriteFile(tmp, []byte(sum+"\n"), 0o644); err != nil { //nolint:gosec // a checksum, for anyone to read
+		return err
+	}
+	return os.Rename(tmp, checksumPath(file))
+}
+
+// readChecksum is a file's recorded sha256, or "" for a file put in place by
+// hand, which has none.
+func readChecksum(file string) string {
+	b, err := os.ReadFile(checksumPath(file)) //nolint:gosec // beside a file in the mirror
+	if err != nil {
+		return ""
+	}
+	return strings.TrimSpace(string(b))
+}
+
 // sha256File is the checksum of a file on disk.
 //
 // Reported after a download so an operator has something to compare against
@@ -205,8 +233,16 @@ type Manifest struct {
 type ManifestFile struct {
 	Name  string `json:"name"`
 	Bytes int64  `json:"bytes"`
+	// SHA256 is the file's checksum when the mirror fetched it, for a node
+	// to check what it downloaded; empty for a file put there by hand.
+	SHA256 string `json:"sha256,omitempty"`
 }

+// ErrIncomplete is a version not yet whole: a download still running, or a
+// shard its own index names missing. Its manifest is not given out, so no
+// node starts a model from part of it.
+var ErrIncomplete = errors.New("this version is not complete")
+
 // ManifestFor lists the files of one model version.
 //
 // Built from the directory rather than from a file written alongside it. A
@@ -219,15 +255,79 @@ func ManifestFor(root, model, version string) (Manifest, error) {
 		return Manifest{}, err
 	}
 	m := Manifest{Model: model, Version: version}
+	dir := filepath.Join(root, model, version)
 	for _, e := range entries {
 		if e.Model != model || e.Version != version {
 			continue
 		}
-		m.Files = append(m.Files, ManifestFile{Name: e.File, Bytes: e.Bytes})
+		m.Files = append(m.Files, ManifestFile{Name: e.File, Bytes: e.Bytes,
+			SHA256: readChecksum(filepath.Join(dir, e.File))})
 		m.Bytes += e.Bytes
 	}
 	if len(m.Files) == 0 {
 		return Manifest{}, fmt.Errorf("%s/%s: %w", model, version, os.ErrNotExist)
 	}
+	if err := complete(dir, m.Files); err != nil {
+		return Manifest{}, fmt.Errorf("%s/%s: %w: %v", model, version, ErrIncomplete, err)
+	}
 	return m, nil
 }
+
+// complete says whether a version's files are the whole of it: nothing still
+// downloading, every shard a safetensors index maps a tensor to, and every
+// part of a split ("-00002-of-00004").
+//
+// From what the files say of each other rather than from a list written when
+// the fetch began, so a version filled by hand is held to the same.
+func complete(dir string, files []ManifestFile) error {
+	if parts, _ := filepath.Glob(filepath.Join(dir, "*"+partSuffix)); len(parts) > 0 {
+		var names []string
+		for _, p := range parts {
+			if b := filepath.Base(p); !strings.HasPrefix(b, ".") {
+				names = append(names, strings.TrimSuffix(b, partSuffix))
+			}
+		}
+		if len(names) > 0 {
+			return fmt.Errorf("still downloading %s", strings.Join(names, ", "))
+		}
+	}
+	have := map[string]bool{}
+	for _, f := range files {
+		have[f.Name] = true
+	}
+	var missing []string
+	want := map[string]bool{}
+	for _, f := range files {
+		if strings.HasSuffix(f.Name, ".safetensors.index.json") {
+			var idx struct {
+				WeightMap map[string]string `json:"weight_map"`
+			}
+			b, err := os.ReadFile(filepath.Join(dir, f.Name)) //nolint:gosec // a file in the mirror
+			if err != nil || json.Unmarshal(b, &idx) != nil {
+				return fmt.Errorf("%s cannot be read", f.Name)
+			}
+			for _, shard := range idx.WeightMap {
+				want[shard] = true
+			}
+		}
+		if m := splitPart.FindStringSubmatch(f.Name); m != nil {
+			total, _ := strconv.Atoi(m[3])
+			for i := 1; i <= total; i++ {
+				want[fmt.Sprintf("%s-%0*d-of-%s%s", m[1], len(m[2]), i, m[3], m[4])] = true
+			}
+		}
+	}
+	for name := range want {
+		if !have[name] {
+			missing = append(missing, name)
+		}
+	}
+	if len(missing) > 0 {
+		sort.Strings(missing)
+		return fmt.Errorf("missing %s", strings.Join(missing, ", "))
+	}
+	return nil
+}
+
+// splitPart matches one part of a split: prefix, part, total, extension.
+var splitPart = regexp.MustCompile(`^(.+)-(\d{5})-of-(\d{5})(\.[A-Za-z0-9]+)$`)
@@ -177,3 +177,63 @@ func TestANameThatLeavesTheMirrorIsRefused(t *testing.T) {
 		}
 	}
 }
+
+func TestAVersionIsGivenOutOnlyWhole(t *testing.T) {
+	root := t.TempDir()
+	dir := filepath.Join(root, "m", "v1")
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	write := func(name, body string) {
+		t.Helper()
+		if err := os.WriteFile(filepath.Join(dir, name), []byte(body), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	write("config.json", "{}")
+	write("model.safetensors.index.json",
+		`{"weight_map":{"a":"model-00001-of-00002.safetensors","b":"model-00002-of-00002.safetensors"}}`)
+	write("model-00001-of-00002.safetensors", "aa")
+	write("model-00002-of-00002.safetensors.part", "b")
+
+	if _, err := ManifestFor(root, "m", "v1"); !errors.Is(err, ErrIncomplete) || !strings.Contains(err.Error(), "still downloading") {
+		t.Fatalf("a version with a shard downloading was given out: %v", err)
+	}
+	// The download failed and its .part went: the index still names it.
+	_ = os.Remove(filepath.Join(dir, "model-00002-of-00002.safetensors.part"))
+	if _, err := ManifestFor(root, "m", "v1"); !errors.Is(err, ErrIncomplete) ||
+		!strings.Contains(err.Error(), "missing model-00002-of-00002.safetensors") {
+		t.Fatalf("a version missing a shard its index names was given out: %v", err)
+	}
+	write("model-00002-of-00002.safetensors", "bb")
+	m, err := ManifestFor(root, "m", "v1")
+	if err != nil || len(m.Files) != 4 {
+		t.Fatalf("the whole version: %+v, %v", m, err)
+	}
+
+	// A split GGUF with no index is held to its own numbering.
+	gg := filepath.Join(root, "g", "v1")
+	_ = os.MkdirAll(gg, 0o755)
+	_ = os.WriteFile(filepath.Join(gg, "q-00001-of-00003.gguf"), []byte("x"), 0o644)
+	_ = os.WriteFile(filepath.Join(gg, "q-00003-of-00003.gguf"), []byte("x"), 0o644)
+	if _, err := ManifestFor(root, "g", "v1"); !errors.Is(err, ErrIncomplete) || !strings.Contains(err.Error(), "q-00002-of-00003.gguf") {
+		t.Fatalf("a split missing its second part was given out: %v", err)
+	}
+}
+
+func TestAMirroredFilesChecksumIsHandedOnAndHiddenFromListings(t *testing.T) {
+	root := t.TempDir()
+	e := Entry{Model: "m", Version: "v1", File: "f.gguf"}
+	added, err := Add(context.Background(), root, serving(t, "weights"), e, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	entries, _ := List(root)
+	if len(entries) != 1 {
+		t.Errorf("the checksum was listed as a file of the model: %+v", entries)
+	}
+	m, err := ManifestFor(root, "m", "v1")
+	if err != nil || m.Files[0].SHA256 != added.SHA256 || added.SHA256 == "" {
+		t.Errorf("manifest %+v, checksum %q: %v", m, added.SHA256, err)
+	}
+}
@@ -18,7 +18,6 @@ import (
 	"encoding/pem"
 	"errors"
 	"fmt"
-	"gitea.mynodes.xyz/ilker/llmgw/internal/release"
 	"net"
 	"os"
 	"path/filepath"
@@ -27,6 +26,8 @@ import (
 	"strings"
 	"time"

+	"gitea.mynodes.xyz/ilker/llmgw/internal/release"
+
 	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/discovery"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/doctor"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/tlsconf"
@@ -86,6 +87,15 @@ type Options struct {
 	Releases string
 	// IgnoreChecks goes on when doctor finds a problem.
 	IgnoreChecks bool
+	// DisableEngines are engines this node never installs or advertises,
+	// written into its agent.yaml, and not checked by doctor: vLLM on a node
+	// of cards it does not run on, meant for llama.cpp alone. Narrower than
+	// IgnoreChecks, which also passes over everything else.
+	DisableEngines []string
+	// VLLMVersion is the vLLM the control plane has nodes install, for
+	// doctor to check against; "none" when it has them install none, empty
+	// for doctor's own default.
+	VLLMVersion string
 	// SSHWait is how long an unreachable machine is waited for.
 	SSHWait time.Duration

@@ -390,6 +400,18 @@ func (r *Runner) check(ctx context.Context, res *Result) error {
 	if r.Opts.IBRoot != "" {
 		args += " --ib-root " + shellQuote(r.Opts.IBRoot)
 	}
+	// The fabrics given, checked as the ones they will be; any adapter they
+	// do not name is checked as one nothing uses.
+	for _, fab := range r.Opts.Fabrics {
+		args += " --fabric " + shellQuote(fab.Name+":"+strings.Join(fab.HCAs, ","))
+	}
+	switch {
+	case slices.Contains(r.Opts.DisableEngines, "vllm"):
+		args += " --vllm none"
+		r.emit(StepCheck, Info, "vLLM is disabled on this node: it joins with llama.cpp alone", "")
+	case r.Opts.VLLMVersion != "":
+		args += " --vllm " + shellQuote(r.Opts.VLLMVersion)
+	}
 	out, err := r.Remote.Run(ctx, args, nil)
 	var rep struct {
 		Findings []doctor.Finding `json:"findings"`
@@ -596,8 +618,11 @@ func AgentYAML(o Options, res *Result) string {
 			fmt.Fprintf(&b, "  ib_root: %q\n", o.IBRoot)
 		}
 	}
-	if o.LlamaCppBin != "" || o.VLLMBin != "" || o.SGLangPython != "" {
+	if o.LlamaCppBin != "" || o.VLLMBin != "" || o.SGLangPython != "" || len(o.DisableEngines) > 0 {
 		b.WriteString("engines:\n")
+		if len(o.DisableEngines) > 0 {
+			fmt.Fprintf(&b, "  disabled: [%s]\n", strings.Join(o.DisableEngines, ", "))
+		}
 		if o.LlamaCppBin != "" {
 			fmt.Fprintf(&b, "  llamacpp:\n    bin: %q\n", o.LlamaCppBin)
 		}
@@ -280,3 +280,44 @@ func said(events []Event, level, s string) bool {
 	}
 	return false
 }
+
+func TestAnEngineDisabledOnANodeIsWrittenAndNotChecked(t *testing.T) {
+	m := newFake()
+	r, events := runner(t, m, Options{Version: "dev", DisableEngines: []string{"vllm"}, VLLMVersion: "0.30.0"})
+	if _, err := r.Run(t.Context()); err != nil {
+		t.Fatal(err)
+	}
+	if yaml := string(m.files["/etc/llmgw/agent.yaml"]); !strings.Contains(yaml, "engines:\n  disabled: [vllm]") {
+		t.Errorf("agent.yaml:\n%s", yaml)
+	}
+	if !m.ran("sudo llmgw doctor --json --vllm none") {
+		t.Errorf("doctor was asked about a vLLM this node will never run: %v", m.scripts)
+	}
+	if !said(*events, Info, "joins with llama.cpp alone") {
+		t.Error("the reduced engine set went unsaid")
+	}
+
+	// Not disabled: doctor checks the vLLM the control plane installs, not
+	// its own default.
+	m = newFake()
+	r, _ = runner(t, m, Options{Version: "dev", VLLMVersion: "0.25.1"})
+	if _, err := r.Run(t.Context()); err != nil {
+		t.Fatal(err)
+	}
+	if !m.ran("sudo llmgw doctor --json --vllm '0.25.1'") && !m.ran("sudo llmgw doctor --json --vllm 0.25.1") {
+		t.Errorf("doctor was not told the control plane's vLLM: %v", m.scripts)
+	}
+}
+
+func TestTheFabricsGivenAreCheckedAsDeclared(t *testing.T) {
+	// agent.yaml is written after doctor runs, so the fabrics on the command
+	// line are handed to doctor: checked strictly, and nothing else is.
+	m := newFake()
+	r, _ := runner(t, m, Options{Version: "dev", Fabrics: []Fabric{{Name: "ib-a", HCAs: []string{"mlx5_0", "mlx5_1"}}}})
+	if _, err := r.Run(t.Context()); err != nil {
+		t.Fatal(err)
+	}
+	if !m.ran("--fabric 'ib-a:mlx5_0,mlx5_1'") {
+		t.Errorf("doctor was not told the fabric: %v", m.scripts)
+	}
+}
@@ -9,6 +9,7 @@ package release

 import (
 	"bufio"
+	"bytes"
 	"context"
 	"crypto/sha256"
 	"encoding/hex"
@@ -51,7 +52,7 @@ func Install(ctx context.Context, client *http.Client, base, version, goos, goar
 	asset := AssetName(version, goos, goarch)
 	root := strings.TrimSuffix(base, "/") + "/" + version + "/"

-	want, err := checksum(ctx, client, root+"checksums.txt", asset)
+	want, err := checksum(ctx, client, root, asset)
 	if err != nil {
 		return "", err
 	}
@@ -85,13 +86,35 @@ func Install(ctx context.Context, client *http.Client, base, version, goos, goar
 	return backup, nil
 }

-func checksum(ctx context.Context, client *http.Client, url, asset string) (string, error) {
-	body, err := get(ctx, client, url)
+// checksum reads the release's checksum for asset — after checking the
+// checksums' signature, when this build trusts a signing key. A release with
+// no signature is then refused rather than taken on its checksums: a missing
+// signature is exactly what replacing a release would look like.
+func checksum(ctx context.Context, client *http.Client, root, asset string) (string, error) {
+	body, err := get(ctx, client, root+"checksums.txt")
 	if err != nil {
 		return "", err
 	}
 	defer body.Close()
-	sc := bufio.NewScanner(io.LimitReader(body, 1<<20))
+	sums, err := io.ReadAll(io.LimitReader(body, 1<<20))
+	if err != nil {
+		return "", err
+	}
+	if key := TrustedKey(); key != "" {
+		sigBody, err := get(ctx, client, root+SignatureName)
+		if err != nil {
+			return "", fmt.Errorf("this build only installs signed releases, and %s could not be read: %w", SignatureName, err)
+		}
+		sig, err := io.ReadAll(io.LimitReader(sigBody, 4096))
+		sigBody.Close()
+		if err != nil {
+			return "", err
+		}
+		if err := Verify(key, sums, string(sig)); err != nil {
+			return "", err
+		}
+	}
+	sc := bufio.NewScanner(bytes.NewReader(sums))
 	for sc.Scan() {
 		f := strings.Fields(sc.Text())
 		if len(f) == 2 && f[1] == asset && len(f[0]) == 64 {
@@ -0,0 +1,72 @@
+package release
+
+import (
+	"crypto/ed25519"
+	"crypto/rand"
+	_ "embed"
+	"encoding/base64"
+	"errors"
+	"fmt"
+	"os"
+	"strings"
+)
+
+// trustedPub is the public key releases are signed with, base64, built into
+// every binary from the repository. Empty — as it is until a key is made —
+// trusts a release on its checksums alone, as every release before this did.
+//
+// Built in rather than fetched, because the point of a signature is that the
+// server the release comes from cannot vouch for it: a machine that already
+// runs llmgw checks the next release against the key the release it runs was
+// built with, so replacing both the binaries and the checksums on the server
+// is no longer enough.
+//
+//go:embed trusted.pub
+var trustedPub string
+
+// TrustedKey is the key releases are verified with: LLMGW_RELEASE_KEY when set
+// (for rotating a key, or testing), the built-in one otherwise; empty is none.
+func TrustedKey() string {
+	if k := strings.TrimSpace(os.Getenv("LLMGW_RELEASE_KEY")); k != "" {
+		return k
+	}
+	return strings.TrimSpace(trustedPub)
+}
+
+// SignatureName is the file a release's checksums are signed in.
+const SignatureName = "checksums.txt.sig"
+
+// GenerateKey makes a signing key pair, both base64: the private key is the
+// 32-byte seed, which is what goes into the CI secret.
+func GenerateKey() (public, private string, err error) {
+	pub, priv, err := ed25519.GenerateKey(rand.Reader)
+	if err != nil {
+		return "", "", err
+	}
+	return base64.StdEncoding.EncodeToString(pub), base64.StdEncoding.EncodeToString(priv.Seed()), nil
+}
+
+// Sign signs data with a base64 seed, returning the signature in base64.
+func Sign(privateSeed string, data []byte) (string, error) {
+	seed, err := base64.StdEncoding.DecodeString(strings.TrimSpace(privateSeed))
+	if err != nil || len(seed) != ed25519.SeedSize {
+		return "", errors.New("the signing key is not a base64 ed25519 seed")
+	}
+	return base64.StdEncoding.EncodeToString(ed25519.Sign(ed25519.NewKeyFromSeed(seed), data)), nil
+}
+
+// Verify checks a base64 signature of data against a base64 public key.
+func Verify(public string, data []byte, signature string) error {
+	pub, err := base64.StdEncoding.DecodeString(strings.TrimSpace(public))
+	if err != nil || len(pub) != ed25519.PublicKeySize {
+		return errors.New("the trusted release key is not a base64 ed25519 public key")
+	}
+	sig, err := base64.StdEncoding.DecodeString(strings.TrimSpace(signature))
+	if err != nil {
+		return fmt.Errorf("the release signature is not base64: %w", err)
+	}
+	if !ed25519.Verify(pub, data, sig) {
+		return errors.New("the release's checksums are not signed by the trusted key")
+	}
+	return nil
+}
@@ -0,0 +1,99 @@
+package release
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// signedServer publishes a release whose checksums are signed with seed,
+// or not signed at all when seed is empty; tamper rewrites the checksums
+// after signing.
+func signedServer(t *testing.T, seed string, body []byte, tamper bool) (*httptest.Server, string) {
+	t.Helper()
+	asset := AssetName("v2.0.0", "linux", "amd64")
+	sum := sha256.Sum256(body)
+	sums := hex.EncodeToString(sum[:]) + "  " + asset + "\n"
+	sig := ""
+	if seed != "" {
+		var err error
+		if sig, err = Sign(seed, []byte(sums)); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if tamper {
+		sums += strings.Repeat("0", 64) + "  extra-file\n"
+	}
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/v2.0.0/checksums.txt":
+			_, _ = w.Write([]byte(sums))
+		case "/v2.0.0/" + SignatureName:
+			if sig == "" {
+				http.NotFound(w, r)
+				return
+			}
+			_, _ = w.Write([]byte(sig + "\n"))
+		case "/v2.0.0/" + asset:
+			_, _ = w.Write(body)
+		default:
+			http.NotFound(w, r)
+		}
+	}))
+	t.Cleanup(srv.Close)
+	return srv, asset
+}
+
+func TestABuildThatTrustsAKeyInstallsOnlyReleasesSignedWithIt(t *testing.T) {
+	pub, priv, err := GenerateKey()
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, otherPriv, _ := GenerateKey()
+	t.Setenv("LLMGW_RELEASE_KEY", pub)
+	install := func(srv *httptest.Server) error {
+		target := filepath.Join(t.TempDir(), "llmgw")
+		_ = os.WriteFile(target, []byte("old"), 0o755)
+		_, err := Install(t.Context(), srv.Client(), srv.URL, "v2.0.0", "linux", "amd64", target)
+		return err
+	}
+	body := []byte("new binary")
+
+	good, _ := signedServer(t, priv, body, false)
+	if err := install(good); err != nil {
+		t.Fatalf("a release signed with the trusted key was refused: %v", err)
+	}
+	for name, srv := range map[string]*httptest.Server{
+		"unsigned":            func() *httptest.Server { s, _ := signedServer(t, "", body, false); return s }(),
+		"signed by another":   func() *httptest.Server { s, _ := signedServer(t, otherPriv, body, false); return s }(),
+		"checksums rewritten": func() *httptest.Server { s, _ := signedServer(t, priv, body, true); return s }(),
+	} {
+		if err := install(srv); err == nil {
+			t.Errorf("%s: installed", name)
+		}
+	}
+
+	// A build that trusts no key installs on checksums, as before.
+	t.Setenv("LLMGW_RELEASE_KEY", "")
+	if TrustedKey() != "" {
+		t.Skip("this build embeds a trusted key; the unsigned path cannot be tested here")
+	}
+	unsigned, _ := signedServer(t, "", body, false)
+	if err := install(unsigned); err != nil {
+		t.Errorf("without a trusted key an unsigned release was refused: %v", err)
+	}
+}
+
+func TestAKeyThatIsNotOneIsSaid(t *testing.T) {
+	if _, err := Sign("not-a-key", []byte("x")); err == nil {
+		t.Error("signed with a malformed key")
+	}
+	if err := Verify("bm9wZQ==", []byte("x"), "c2ln"); err == nil || !strings.Contains(err.Error(), "public key") {
+		t.Errorf("a malformed trusted key: %v", err)
+	}
+}
@@ -0,0 +1,98 @@
+package router
+
+import (
+	"encoding/json"
+	"net/http"
+	"testing"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/auth"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/config"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/catalog"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/store"
+)
+
+func adapterModel() catalog.Model {
+	m := catModel("coder", nil)
+	m.Profiles = []catalog.Profile{{Name: "cpu", Engine: "llamacpp", MinMemoryGB: 4}}
+	m.Adapters = []catalog.Adapter{
+		{Name: "coder-sql", Source: "http://mirror/sql.gguf"},
+		{Name: "coder-docs", Source: "http://mirror/docs.gguf", Scale: 0.5},
+	}
+	return m
+}
+
+func TestAnAdapterIsAskedForByNameAndAppliedOnItsModel(t *testing.T) {
+	e := newFakeEngine(t, "coder")
+	cfg := baseConfig(config.Endpoint{ID: "a", URL: e.srv.URL, Models: []string{"coder"}})
+	cfg.Models = []catalog.Model{adapterModel()}
+	rt := newTestRouter(t, cfg)
+
+	sent := func(model string) map[string]any {
+		t.Helper()
+		rec := post(t, rt.Handler(), map[string]any{"model": model, "messages": []any{}}, nil)
+		if rec.Code != http.StatusOK {
+			t.Fatalf("%s: status %d: %s", model, rec.Code, rec.Body.String())
+		}
+		var body map[string]any
+		_ = json.Unmarshal(e.last.Load().([]byte), &body)
+		return body
+	}
+	body := sent("coder-docs")
+	if body["model"] != "coder" {
+		t.Errorf("the engine was asked for %v, not the model the adapter runs on", body["model"])
+	}
+	lora, _ := json.Marshal(body["lora"])
+	if string(lora) != `[{"id":1,"scale":0.5}]` {
+		t.Errorf("lora %s, expected the second adapter at its scale", lora)
+	}
+	if body := sent("coder-sql"); func() string { b, _ := json.Marshal(body["lora"]); return string(b) }() != `[{"id":0,"scale":1}]` {
+		t.Errorf("the first adapter, at the default scale: %v", body["lora"])
+	}
+	if body := sent("coder"); body["lora"] != nil {
+		t.Errorf("the base model was sent an adapter: %v", body["lora"])
+	}
+}
+
+func TestAKeyAllowedAnAdapterReachesThatAdapterOnly(t *testing.T) {
+	cfg := baseConfig()
+	cfg.Models = []catalog.Model{adapterModel()}
+	rt := newTestRouter(t, cfg)
+	key := func(allowed ...string) auth.Principal {
+		return auth.Principal{Limits: store.Effective{AllowedModels: allowed}}
+	}
+	for _, c := range []struct {
+		allowed   []string
+		requested string
+		want      bool
+	}{
+		{[]string{"coder-sql"}, "coder-sql", true},
+		{[]string{"coder-sql"}, "coder", false},
+		{[]string{"coder-sql"}, "coder-docs", false},
+		{[]string{"coder"}, "coder-sql", true}, // the model allows its adapters
+		{[]string{"coder"}, "coder", true},
+	} {
+		if got := rt.permits(key(c.allowed...), c.requested); got != c.want {
+			t.Errorf("allowed %v, asked for %s: %v", c.allowed, c.requested, got)
+		}
+	}
+}
+
+func TestAdaptersShareTheNamespaceOfModels(t *testing.T) {
+	a := adapterModel()
+	b := catModel("coder-sql", nil)
+	if _, err := catalog.New([]catalog.Model{a, b}); err == nil {
+		t.Error("an adapter named like another model was accepted")
+	}
+	bad := adapterModel()
+	bad.Profiles = []catalog.Profile{{Name: "gpu", Engine: "vllm", GPUClass: "h100-80g", TP: 1}}
+	if _, err := catalog.New([]catalog.Model{bad}); err == nil {
+		t.Error("adapters on a vLLM profile were accepted")
+	}
+	c, err := catalog.New([]catalog.Model{adapterModel()})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if m, ad, i, ok := c.ResolveAdapter("coder-docs"); !ok || m.Name != "coder" || i != 1 || ad.EffectiveScale() != 0.5 {
+		t.Errorf("resolve: %v %+v %d %v", m.Name, ad, i, ok)
+	}
+}
@@ -34,6 +34,12 @@ type Config struct {
 	// QueueCapacityPerBand bounds each band's queue. Rejecting immediately
 	// beats accumulating silently.
 	QueueCapacityPerBand int `koanf:"queue_capacity_per_band"`
+	// QueueCapacityTotal bounds everything queued, across bands. When it is
+	// reached, a request is not refused while a lower band has requests
+	// queued: the newest of the lowest band is turned away instead
+	// (displacement), since the oldest have waited longest and are nearest
+	// their TTL. Zero takes half of every band's capacity together.
+	QueueCapacityTotal int `koanf:"queue_capacity_total"`
 	// Interval is how often the dispatcher re-checks saturation. Releases also
 	// wake it, so this is only the backstop for saturation falling on its own.
 	Interval time.Duration `koanf:"interval"`
@@ -94,6 +100,9 @@ func (c *Config) normalize() error {
 	if c.QueueCapacityPerBand <= 0 {
 		c.QueueCapacityPerBand = 2000
 	}
+	if c.QueueCapacityTotal <= 0 {
+		c.QueueCapacityTotal = max(c.QueueCapacityPerBand*len(c.Bands)/2, 1)
+	}
 	if c.Interval <= 0 {
 		c.Interval = 10 * time.Millisecond
 	}
@@ -123,6 +132,10 @@ var (
 	ErrQueueTimeout = errors.New("admission queue timeout")
 	// ErrClosed means the controller is shutting down.
 	ErrClosed = errors.New("admission controller closed")
+	// ErrDisplaced means a queued request was turned away to make room for
+	// one of a higher band. Worth retrying: the queue is full now, not for
+	// good.
+	ErrDisplaced = errors.New("displaced from the admission queue by higher-priority requests")
 )

 // Stats is a snapshot of one band's queue, for metrics.
@@ -259,6 +272,12 @@ func (c *Controller) Acquire(ctx context.Context, r Request) (*Lease, error) {
 		c.mu.Unlock()
 		return nil, fmt.Errorf("%w: band %q", ErrQueueFull, band.Name)
 	}
+	if c.queuedLocked() >= c.cfg.QueueCapacityTotal && !c.displaceBelowLocked(idx) {
+		c.rejected[band.Name]++
+		c.mu.Unlock()
+		return nil, fmt.Errorf("%w: %d queued in all, and nothing of a lower band to make room",
+			ErrQueueFull, c.cfg.QueueCapacityTotal)
+	}

 	w := &waiter{
 		tenant:   tenantKey(r),
@@ -279,6 +298,10 @@ func (c *Controller) Acquire(ctx context.Context, r Request) (*Lease, error) {
 			c.mu.Unlock()
 			return nil, ErrClosed
 		}
+		if w.err != nil {
+			c.mu.Unlock()
+			return nil, w.err
+		}
 		c.waitTotal[band.Name] += time.Since(w.enqueued)
 		c.mu.Unlock()
 		return &Lease{c: c, band: band.Name, model: r.Model}, nil
@@ -297,8 +320,9 @@ func (c *Controller) Acquire(ctx context.Context, r Request) (*Lease, error) {
 func (c *Controller) remove(idx int, w *waiter, reason string) {
 	c.mu.Lock()
 	defer c.mu.Unlock()
-	if w.admitted {
-		// It won the race; the caller will see ready closed next time round.
+	if w.admitted || w.err != nil {
+		// It won the race, or was displaced; the caller will see ready
+		// closed next time round.
 		return
 	}
 	c.queues[idx].remove(w)
@@ -410,3 +434,31 @@ func tenantKey(r Request) string {
 	}
 	return "anonymous"
 }
+
+// queuedLocked is how many requests wait in every band. Called with mu held.
+func (c *Controller) queuedLocked() int {
+	n := 0
+	for _, q := range c.queues {
+		n += q.len()
+	}
+	return n
+}
+
+// displaceBelowLocked turns away the newest request of the lowest band below
+// idx that has any queued, and says whether there was one. Called with mu held.
+func (c *Controller) displaceBelowLocked(idx int) bool {
+	// Bands are sorted highest priority first, so lower bands have larger
+	// indices, and the lowest is last.
+	for j := len(c.queues) - 1; j > idx; j-- {
+		w := c.queues[j].newest()
+		if w == nil {
+			continue
+		}
+		c.queues[j].remove(w)
+		w.err = fmt.Errorf("%w (band %q)", ErrDisplaced, c.cfg.Bands[j].Name)
+		close(w.ready)
+		c.rejected[c.cfg.Bands[j].Name+"/displaced"]++
+		return true
+	}
+	return false
+}
@@ -0,0 +1,89 @@
+package admission
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+)
+
+// queueUntil starts an Acquire in the background and waits until the band
+// holds depth requests, returning where its result will arrive.
+func queueUntil(t *testing.T, c *Controller, band string, bandIdx, depth int) <-chan error {
+	t.Helper()
+	out := make(chan error, 1)
+	go func() {
+		l, err := c.Acquire(context.Background(), Request{Model: "m", Band: band})
+		if err == nil {
+			l.Release()
+		}
+		out <- err
+	}()
+	deadline := time.Now().Add(2 * time.Second)
+	for c.Depths()[bandIdx].Depth < depth && time.Now().Before(deadline) {
+		time.Sleep(time.Millisecond)
+	}
+	if got := c.Depths()[bandIdx].Depth; got != depth {
+		t.Fatalf("band %s holds %d, expected %d", band, got, depth)
+	}
+	return out
+}
+
+func TestAHigherBandDisplacesTheNewestOfTheLowest(t *testing.T) {
+	s := &fixedSat{}
+	s.set(5) // past every ceiling: everything queues
+	cfg := testConfig()
+	cfg.QueueCapacityTotal = 3
+	c := newController(t, cfg, s.fn)
+	// Bands, highest first: platinum 0, gold 1, silver 2, bronze 3.
+	oldest := queueUntil(t, c, "bronze", 3, 1)
+	time.Sleep(2 * time.Millisecond)
+	silver := queueUntil(t, c, "silver", 2, 1)
+	time.Sleep(2 * time.Millisecond)
+	newest := queueUntil(t, c, "bronze", 3, 2)
+
+	// Full at three. Gold arrives: the newest bronze goes, not the oldest,
+	// and not silver, which outranks bronze.
+	gold := queueUntil(t, c, "gold", 1, 1)
+	select {
+	case err := <-newest:
+		if !errors.Is(err, ErrDisplaced) {
+			t.Fatalf("the newest bronze got %v", err)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("nothing was displaced")
+	}
+	if d := c.Depths(); d[3].Depth != 1 || d[2].Depth != 1 || d[1].Depth != 1 {
+		t.Errorf("depths %+v: one bronze, the silver and the gold should remain", d)
+	}
+
+	// Bronze cannot displace anything: nothing is below it.
+	_, err := c.Acquire(t.Context(), Request{Model: "m", Band: "bronze"})
+	if !errors.Is(err, ErrQueueFull) {
+		t.Errorf("a bronze request at a full queue got %v", err)
+	}
+	// Another gold takes the remaining bronze, then silver.
+	queueUntil(t, c, "gold", 1, 2)
+	if err := <-oldest; !errors.Is(err, ErrDisplaced) {
+		t.Errorf("the last bronze got %v", err)
+	}
+	queueUntil(t, c, "gold", 1, 3)
+	if err := <-silver; !errors.Is(err, ErrDisplaced) {
+		t.Errorf("silver got %v", err)
+	}
+	// All gold now: a fourth has nobody lower to displace.
+	if _, err := c.Acquire(t.Context(), Request{Model: "m", Band: "gold"}); !errors.Is(err, ErrQueueFull) {
+		t.Errorf("a gold request with only gold queued got %v", err)
+	}
+
+	// And they are served once there is room.
+	s.set(0)
+	select {
+	case err := <-gold:
+		if err != nil {
+			t.Errorf("the first gold was not served: %v", err)
+		}
+	case <-time.After(2 * time.Second):
+		t.Error("the queue did not drain")
+	}
+}
@@ -12,6 +12,9 @@ type waiter struct {
 	// admitted guards the race between the dispatcher admitting a waiter and
 	// the waiter giving up on its own.
 	admitted bool
+	// err is why the waiter was turned away while queued — displaced by a
+	// higher band — set before ready is closed.
+	err error
 }

 // bandQueue holds one band's waiters, grouped by tenant.
@@ -166,3 +169,17 @@ func (q *bandQueue) dropExpired(now time.Time, onDrop func()) {
 		q.byTenant[tenant] = kept
 	}
 }
+
+// newest is the most recently queued waiter, across every tenant: the one
+// displacement takes, since it has invested the least waiting.
+func (q *bandQueue) newest() *waiter {
+	var out *waiter
+	for _, ws := range q.byTenant {
+		for _, w := range ws {
+			if out == nil || w.enqueued.After(out.enqueued) {
+				out = w
+			}
+		}
+	}
+	return out
+}
@@ -37,8 +37,16 @@ type Info struct {
 	Weight float64
 	// Priority expresses pool preference; higher pools are tried first.
 	Priority int
+	// UpstreamModel is what the endpoint calls the model, when that differs
+	// from the catalog's name: an external provider's. Requests are rewritten
+	// to it. Empty sends the name as the client did.
+	UpstreamModel string
 }

+// EngineExternal marks an endpoint that is a provider outside the cluster.
+// It publishes no metrics, so it is never scraped and never reads as stale.
+const EngineExternal = "external"
+
 // Serves reports whether the replica serves the given model.
 func (i Info) Serves(model string) bool {
 	return slices.Contains(i.Models, model)
@@ -242,7 +250,7 @@ func (r *Registry) Snapshots(model string) []Snapshot {
 			continue
 		}
 		var age time.Duration
-		stale := true
+		stale := l.info.Engine != EngineExternal
 		if !l.scrapedAt.IsZero() {
 			age = now.Sub(l.scrapedAt)
 			stale = age > r.cfg.StaleAfter
@@ -0,0 +1,89 @@
+package router
+
+import (
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"sync/atomic"
+	"testing"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/config"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/control/catalog"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/router/endpoint"
+)
+
+func TestALocalModelWithNoCapacityFallsBackToAnExternalProvider(t *testing.T) {
+	local := newFakeEngine(t, "local")
+	local.status.Store(http.StatusServiceUnavailable)
+
+	var calls atomic.Int64
+	var gotModel, gotAuth atomic.Value
+	provider := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/metrics" {
+			t.Error("the provider was scraped for metrics")
+		}
+		calls.Add(1)
+		gotAuth.Store(r.Header.Get("Authorization"))
+		var body map[string]any
+		b, _ := io.ReadAll(r.Body)
+		_ = json.Unmarshal(b, &body)
+		gotModel.Store(fmt.Sprint(body["model"]))
+		w.Header().Set("Content-Type", "application/json")
+		fmt.Fprint(w, `{"model":"gpt-4o-mini","choices":[{"message":{"content":"from outside"}}],"usage":{"prompt_tokens":3,"completion_tokens":2}}`)
+	}))
+	t.Cleanup(provider.Close)
+
+	cfg := baseConfig(config.Endpoint{ID: "l", URL: local.srv.URL, Models: []string{"local"}})
+	cloud := catalog.Model{Name: "cloud", Source: "external",
+		External: &catalog.External{URL: provider.URL, Model: "gpt-4o-mini"}}
+	cfg.Models = []catalog.Model{catModel("local", nil, "cloud"), cloud}
+	rt := newTestRouter(t, cfg)
+	rt.SetEndpoints(t.Context(), []endpoint.Info{
+		{ID: "l", URL: local.srv.URL, Models: []string{"local"}},
+		{ID: "external:cloud", URL: provider.URL, Models: []string{"cloud"}, APIKey: "provider-key",
+			Engine: endpoint.EngineExternal, Pool: "external", Weight: 1, UpstreamModel: "gpt-4o-mini"},
+	})
+
+	rec := post(t, rt.Handler(), map[string]any{"model": "local", "messages": []any{}}, nil)
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status %d: %s", rec.Code, rec.Body.String())
+	}
+	if calls.Load() != 1 {
+		t.Fatalf("the provider was called %d times", calls.Load())
+	}
+	if gotModel.Load() != "gpt-4o-mini" {
+		t.Errorf("the provider was asked for %v, not its own name for the model", gotModel.Load())
+	}
+	if gotAuth.Load() != "Bearer provider-key" {
+		t.Errorf("the provider got %v, not its key", gotAuth.Load())
+	}
+	if got := rec.Header().Get("x-llmgw-served-model"); got != "cloud" {
+		t.Errorf("the client should be told it was served by cloud, got %q", got)
+	}
+	// Never scraped, and never stale for it.
+	for _, s := range rt.reg.Snapshots("cloud") {
+		if s.Stale {
+			t.Error("an external endpoint reads as stale")
+		}
+	}
+}
+
+func TestAnExternalModelIsNotAModelWithProfiles(t *testing.T) {
+	good := catalog.Model{Name: "cloud", Source: "external",
+		External: &catalog.External{URL: "https://api.example.com", Model: "gpt-4o-mini"}}
+	if _, err := catalog.New([]catalog.Model{good}); err != nil {
+		t.Fatalf("a valid external model was refused: %v", err)
+	}
+	for name, m := range map[string]catalog.Model{
+		"profiles": {Name: "x", Source: "external", External: &catalog.External{URL: "https://a.b", Model: "m"},
+			Profiles: []catalog.Profile{{Name: "p", Engine: "llamacpp"}}},
+		"no url":   {Name: "x", Source: "external", External: &catalog.External{URL: "api.example.com", Model: "m"}},
+		"no model": {Name: "x", Source: "external", External: &catalog.External{URL: "https://a.b"}},
+	} {
+		if _, err := catalog.New([]catalog.Model{m}); err == nil {
+			t.Errorf("%s: accepted", name)
+		}
+	}
+}
@@ -88,6 +88,9 @@ func TestARefusalBeforeDispatchSaysWhetherWaitingHelps(t *testing.T) {
 		{"queued too long", func(w http.ResponseWriter) {
 			rt.rejectAdmission(w, &parsedRequest{Model: "m"}, admission.ErrQueueTimeout, time.Now())
 		}, 504, ""},
+		{"displaced for a higher band", func(w http.ResponseWriter) {
+			rt.rejectAdmission(w, &parsedRequest{Model: "m"}, admission.ErrDisplaced, time.Now())
+		}, 503, "5"},
 		{"shutting down", func(w http.ResponseWriter) {
 			rt.rejectAdmission(w, &parsedRequest{Model: "m"}, admission.ErrClosed, time.Now())
 		}, 503, ""},
@@ -60,9 +60,15 @@ func (r *Router) rejectAdmission(w http.ResponseWriter, req *parsedRequest, err
 		return
 	case errors.Is(err, admission.ErrQueueTimeout):
 		status = http.StatusGatewayTimeout
+	case errors.Is(err, admission.ErrDisplaced):
+		// Turned away for higher-priority work, not for good: said so the
+		// client backs off and comes back rather than giving up.
+		w.Header().Set("Retry-After", "5")
 	}
 	reason := "unavailable"
 	switch {
+	case errors.Is(err, admission.ErrDisplaced):
+		reason = "displaced"
 	case errors.Is(err, admission.ErrQueueFull):
 		reason = "queue_full"
 	case errors.Is(err, admission.ErrQueueTimeout):
@@ -185,6 +191,16 @@ func (r *Router) Handler() http.Handler {
 			pages.RegisterAdminUsers(mux, r.store)
 			// Quotas are written on the same tree the accounts page builds, so
 			// they are mounted together.
+			// Capacity quotas are the control plane's, reached through the
+			// same admin API the CLI uses.
+			if c, ok := r.clusterView.(web.CapacityAdmin); ok {
+				pages.SetCapacity(c)
+			}
+			// The engine build each GPU node installs, chosen on the cluster
+			// page.
+			if e, ok := r.clusterView.(web.NodeEngineAdmin); ok {
+				pages.RegisterNodeEngines(mux, e)
+			}
 			pages.RegisterAdminQuotas(mux, r.store, r.acct)
 			// And the report of what that tree actually used.
 			pages.RegisterAdminUsage(mux, r.store, r.acct)
@@ -415,6 +431,12 @@ func (r *Router) handleInference(path string) http.HandlerFunc {
 // around the allowlist; matching only the canonical name would silently ignore
 // an allowlist written in the names clients actually use.
 func (r *Router) permits(p auth.Principal, requested string) bool {
+	// An adapter named in the allowlist allows that adapter: not its model,
+	// and not the model's other adapters, which the checks below would let
+	// through if the adapter's own name were not asked about first.
+	if _, ok := r.adapter(requested); ok && p.Allows(requested) {
+		return true
+	}
 	canonical, m := r.resolve(requested)
 	if p.Allows(canonical) {
 		return true
@@ -973,3 +995,44 @@ func (r *Router) rateLimitHeaders(w http.ResponseWriter, p auth.Principal) {
 		w.Header().Set("x-ratelimit-reset-tokens", reset)
 	}
 }
+
+// withModel is body with its model set to name, for an endpoint that knows the
+// model by another name — an external provider. The body unchanged when it is
+// not a JSON object, which the endpoint then refuses as it would have anyway.
+func withModel(body []byte, name string) []byte {
+	var m map[string]json.RawMessage
+	if err := json.Unmarshal(body, &m); err != nil {
+		return body
+	}
+	encoded, err := json.Marshal(name)
+	if err != nil {
+		return body
+	}
+	m["model"] = encoded
+	out, err := json.Marshal(m)
+	if err != nil {
+		return body
+	}
+	return out
+}
+
+// withAdapter is body asking for model with a LoRA adapter applied, in
+// llama.cpp's form: "lora": [{"id": index, "scale": scale}]. Adapters not
+// listed stay at zero, the scale the engine loads each one at.
+func withAdapter(body []byte, model string, index int, scale float64) []byte {
+	var m map[string]json.RawMessage
+	if err := json.Unmarshal(body, &m); err != nil {
+		return body
+	}
+	name, err1 := json.Marshal(model)
+	lora, err2 := json.Marshal([]map[string]any{{"id": index, "scale": scale}})
+	if err1 != nil || err2 != nil {
+		return body
+	}
+	m["model"], m["lora"] = name, lora
+	out, err := json.Marshal(m)
+	if err != nil {
+		return body
+	}
+	return out
+}
@@ -61,6 +61,8 @@ type Router struct {
 	// router's half of not having to. A request takes the map once and works
 	// from it, so a swap mid-request is invisible rather than a race.
 	models atomic.Pointer[map[string]catalog.Model]
+	// adapters are the LoRA adapters by the name clients ask for them by.
+	adapters atomic.Pointer[map[string]adapterRef]

 	// extraMetrics are collectors other components publish on /metrics.
 	extraMetrics []prometheus.Collector
@@ -315,14 +317,7 @@ func New(cfg config.Server, log *slog.Logger) (*Router, error) {
 	// all. A control plane pushes over the top of this at startup and on every
 	// change after.
 	if len(cfg.Models) > 0 {
-		seed := map[string]catalog.Model{}
-		for _, m := range cfg.Models {
-			seed[m.Name] = m
-			for _, a := range m.Aliases {
-				seed[a] = m
-			}
-		}
-		r.models.Store(&seed)
+		r.storeModels(cfg.Models)
 	}
 	return r, nil
 }
@@ -347,6 +342,7 @@ func admissionConfig(rc config.Router) admission.Config {
 		},
 		QueueTTL:             rc.Admission.QueueTTL,
 		QueueCapacityPerBand: rc.Admission.QueueCapacityPerBand,
+		QueueCapacityTotal:   rc.Admission.QueueCapacityTotal,
 		Interval:             rc.Admission.Interval,
 	}
 }
@@ -504,15 +500,44 @@ func (r *Router) modelMap() map[string]catalog.Model {
 // The catalog's counterpart to SetEndpoints, and pushed for the same reason:
 // the control plane owns what should exist, the router needs to know, and
 // neither has to be in the other's process for that to work.
-func (r *Router) SetCatalog(c *catalog.Catalog) {
+func (r *Router) SetCatalog(c *catalog.Catalog) { r.storeModels(c.Models()) }
+
+// storeModels makes models what names resolve to: each model by its name, its
+// aliases and its adapters. One place for it, whether the models came from
+// configuration at start or from the control plane since.
+func (r *Router) storeModels(models []catalog.Model) {
 	m := map[string]catalog.Model{}
-	for _, entry := range c.Models() {
+	ads := map[string]adapterRef{}
+	for _, entry := range models {
 		m[entry.Name] = entry
 		for _, a := range entry.Aliases {
 			m[a] = entry
 		}
+		// An adapter resolves to the model it runs on, as an alias does, and
+		// is applied to the request on the way there.
+		for i, a := range entry.Adapters {
+			m[a.Name] = entry
+			ads[a.Name] = adapterRef{base: entry.Name, index: i, scale: a.EffectiveScale()}
+		}
 	}
 	r.models.Store(&m)
+	r.adapters.Store(&ads)
+}
+
+// adapterRef is where an adapter runs and how it is applied.
+type adapterRef struct {
+	base  string
+	index int
+	scale float64
+}
+
+// adapter finds one by the name a client asked for.
+func (r *Router) adapter(name string) (adapterRef, bool) {
+	if p := r.adapters.Load(); p != nil {
+		a, ok := (*p)[name]
+		return a, ok
+	}
+	return adapterRef{}, false
 }

 // LastUsed reports when each endpoint was last sent a request, satisfying
@@ -714,10 +739,20 @@ func (r *Router) dispatchOne(ctx context.Context, w http.ResponseWriter, req *pa
 		uctx, uspan := tracer().Start(ctx, "upstream", trace.WithSpanKind(trace.SpanKindClient),
 			trace.WithAttributes(attribute.String("llmgw.endpoint", dec.Endpoint.ID),
 				attribute.String("url.full", dec.Endpoint.URL+req.Path)))
+		body := req.UpstreamBody
+		if dec.Endpoint.UpstreamModel != "" {
+			body = withModel(body, dec.Endpoint.UpstreamModel)
+		}
+		// An adapter asked for by name, served by its model's replicas: the
+		// engine is asked for the model, with the adapter applied. Not on a
+		// fallback to another model, which does not have it.
+		if a, ok := r.adapter(req.Model); ok && a.base == model {
+			body = withAdapter(body, model, a.index, a.scale)
+		}
 		out, perr := r.px.Do(uctx, w, proxy.Request{
 			URL:                dec.Endpoint.URL + req.Path,
 			APIKey:             dec.Endpoint.APIKey,
-			Body:               req.UpstreamBody,
+			Body:               body,
 			Header:             withTraceparent(uctx, req.Header),
 			Stream:             req.Stream,
 			SuppressUsageChunk: req.SuppressUsageChunk,
@@ -1,6 +1,9 @@
 package scheduler

 import (
+	"sort"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/router/admission"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/router/breaker"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/router/endpoint"
 )
@@ -51,34 +54,41 @@ func KVSaturated(ceiling float64) Filter {
 	}
 }

-// PoolPriority keeps only the endpoints of the highest-priority pool
-// (plan §7.3, the two-stage decision: pool first, then replica).
+// PoolPriority keeps the endpoints of one pool: the highest-priority one,
+// unless it is saturated and a lower one is not (plan §7.3, the two-stage
+// decision — pool first, then replica).
 //
-// Spillover arrives in phase 5, at which point this filter will fall through
-// to the next pool once the preferred one is saturated.
-func PoolPriority() Filter {
-	return poolPriorityFilter{}
+// Saturated is what admission means by it, the same formula over the same
+// thresholds, so the two layers agree on when a pool is full: the H100 pool is
+// used until it is, and only then the A100 one — never the A100 because it
+// happens to be emptier. When every pool is saturated the preferred one is
+// kept: moving to a pool that is just as full buys nothing and loses whatever
+// made the first preferred.
+func PoolPriority(targetWaiting, kvCeiling float64) Filter {
+	return poolPriorityFilter{t: admission.Thresholds{TargetWaitingPerReplica: targetWaiting, KVCeiling: kvCeiling}}
 }

-type poolPriorityFilter struct{}
+type poolPriorityFilter struct{ t admission.Thresholds }

 func (poolPriorityFilter) Name() string { return "pool-priority" }

-func (poolPriorityFilter) Filter(_ *Ctx, eps []endpoint.Snapshot) []endpoint.Snapshot {
+func (f poolPriorityFilter) Filter(_ *Ctx, eps []endpoint.Snapshot) []endpoint.Snapshot {
 	if len(eps) == 0 {
 		return eps
 	}
-	best := eps[0].Priority
-	for _, e := range eps[1:] {
-		if e.Priority > best {
-			best = e.Priority
-		}
-	}
-	out := make([]endpoint.Snapshot, 0, len(eps))
+	byPriority := map[int][]endpoint.Snapshot{}
+	var order []int
 	for _, e := range eps {
-		if e.Priority == best {
-			out = append(out, e)
+		if _, ok := byPriority[e.Priority]; !ok {
+			order = append(order, e.Priority)
+		}
+		byPriority[e.Priority] = append(byPriority[e.Priority], e)
+	}
+	sort.Sort(sort.Reverse(sort.IntSlice(order)))
+	for _, p := range order {
+		if admission.Saturation(byPriority[p], 0, f.t) < 1 {
+			return byPriority[p]
 		}
 	}
-	return out
+	return byPriority[order[0]]
 }
@@ -50,7 +50,7 @@ func Chat(c ProfileConfig) *Profile {
 	c.normalize()
 	return &Profile{
 		Name:    "chat",
-		Filters: []Filter{BreakerOpen(), PoolPriority()},
+		Filters: []Filter{BreakerOpen(), PoolPriority(c.TargetWaiting, c.KVCeiling)},
 		Scorers: c.withPrefix([]WeightedScorer{
 			{Session(c.SessionLoadFactor), 3.0},
 			{Queue(c.TargetWaiting), 1.0},
@@ -81,7 +81,7 @@ func Batch(c ProfileConfig) *Profile {
 	c.normalize()
 	return &Profile{
 		Name:    "batch",
-		Filters: []Filter{BreakerOpen(), PoolPriority()},
+		Filters: []Filter{BreakerOpen(), PoolPriority(c.TargetWaiting, c.KVCeiling)},
 		Scorers: []WeightedScorer{
 			{Queue(c.TargetWaiting), 3.0},
 			{KV(c.KVCeiling), 1.0},
@@ -99,7 +99,7 @@ func LongContext(c ProfileConfig) *Profile {
 	c.normalize()
 	return &Profile{
 		Name:    "long-context",
-		Filters: []Filter{BreakerOpen(), PoolPriority()},
+		Filters: []Filter{BreakerOpen(), PoolPriority(c.TargetWaiting, c.KVCeiling)},
 		// Prefix matching earns more here than anywhere else: a long prompt is
 		// mostly prefill, and prefill is what a cache hit skips.
 		Scorers: c.withPrefix([]WeightedScorer{
@@ -177,7 +177,7 @@ func TestStaleEndpointIsPenalizedNotDropped(t *testing.T) {
 func TestPoolPriority(t *testing.T) {
 	p := &Profile{
 		Name:    "t",
-		Filters: []Filter{PoolPriority()},
+		Filters: []Filter{PoolPriority(4, 0.9)},
 		Scorers: []WeightedScorer{{Queue(4), 1}},
 		Picker:  MaxScore(),
 	}
@@ -189,7 +189,8 @@ func TestPoolPriority(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	// The lower-priority pool must not win despite being emptier (spillover is phase 5).
+	// The lower-priority pool must not win for being emptier: the preferred
+	// one is not saturated (3 waiting of a healthy 4).
 	if d.Endpoint.ID != "h100" {
 		t.Errorf("pool priority was ignored: %s", d.Endpoint.ID)
 	}
@@ -383,3 +384,36 @@ func TestBuiltinProfilesSchedule(t *testing.T) {
 		}
 	}
 }
+
+func TestASaturatedPoolSpillsToTheNext(t *testing.T) {
+	p := &Profile{Name: "t", Filters: []Filter{PoolPriority(4, 0.9)},
+		Scorers: []WeightedScorer{{Queue(4), 1}}, Picker: MaxScore()}
+	h100 := func(w int) endpoint.Snapshot {
+		return ep("h100", waiting(w), func(s *endpoint.Snapshot) { s.Priority = 100; s.Pool = "ib-a/h100" })
+	}
+	a100 := func(w int) endpoint.Snapshot {
+		return ep("a100", waiting(w), func(s *endpoint.Snapshot) { s.Priority = 50; s.Pool = "ib-b/a100" })
+	}
+	pick := func(eps ...endpoint.Snapshot) string {
+		d, err := p.Schedule(newCtx(Request{Model: "m"}), eps)
+		if err != nil {
+			t.Fatal(err)
+		}
+		return d.Endpoint.ID
+	}
+	if got := pick(h100(8), a100(0)); got != "a100" {
+		t.Errorf("the preferred pool at twice its healthy queue did not spill: %s", got)
+	}
+	if got := pick(h100(8), a100(6)); got != "h100" {
+		t.Errorf("both pools saturated: the preferred one should be kept, got %s", got)
+	}
+	if got := pick(h100(4), a100(0)); got != "a100" {
+		t.Errorf("at exactly its healthy limit a pool is full: %s", got)
+	}
+	// KV pressure saturates a pool as queue depth does.
+	full := h100(0)
+	full.Stats.KVCacheUsage = 0.95
+	if got := pick(full, a100(0)); got != "a100" {
+		t.Errorf("a pool out of KV cache did not spill: %s", got)
+	}
+}
@@ -96,6 +96,9 @@ func (p *Poller) Sync(ctx context.Context) {

 	want := make(map[string]endpoint.Info, len(infos))
 	for _, in := range infos {
+		if in.Engine == endpoint.EngineExternal {
+			continue // a provider elsewhere publishes no /metrics
+		}
 		want[in.ID] = in
 	}

@@ -45,6 +45,7 @@ var copyOrder = []string{
 	// derived from this, and a new secret would lock the router out of all
 	// of them at the moment of the move.
 	"cluster_secrets",
+	"node_engines",
 }

 // notCopied are tables deliberately left behind.
@@ -0,0 +1,11 @@
+-- The build of an engine chosen for a node, where it is not left to the
+-- node's hardware: vLLM's CUDA 12.9 build on a node that could run CUDA 13's,
+-- or no vLLM at all. A node with no row here installs what its hardware runs.
+CREATE TABLE node_engines (
+  node_id    TEXT NOT NULL,
+  engine     TEXT NOT NULL,
+  variant    TEXT NOT NULL,
+  updated_at BIGINT NOT NULL,
+  updated_by TEXT NOT NULL,
+  PRIMARY KEY (node_id, engine)
+);
@@ -0,0 +1,11 @@
+-- The build of an engine chosen for a node, where it is not left to the
+-- node's hardware: vLLM's CUDA 12.9 build on a node that could run CUDA 13's,
+-- or no vLLM at all. A node with no row here installs what its hardware runs.
+CREATE TABLE node_engines (
+  node_id    TEXT NOT NULL,
+  engine     TEXT NOT NULL,
+  variant    TEXT NOT NULL,
+  updated_at BIGINT NOT NULL,
+  updated_by TEXT NOT NULL,
+  PRIMARY KEY (node_id, engine)
+);
@@ -0,0 +1,44 @@
+package store
+
+import (
+	"context"
+	"time"
+)
+
+// NodeEngines are the engine builds chosen per node: node id, then engine,
+// then the build ("cu129") or "none". A node or engine not here is left to
+// the node's hardware.
+func (s *DB) NodeEngines(ctx context.Context) (map[string]map[string]string, error) {
+	rows, err := s.db.QueryContext(ctx, `SELECT node_id, engine, variant FROM node_engines`)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	out := map[string]map[string]string{}
+	for rows.Next() {
+		var node, engine, variant string
+		if err := rows.Scan(&node, &engine, &variant); err != nil {
+			return nil, err
+		}
+		if out[node] == nil {
+			out[node] = map[string]string{}
+		}
+		out[node][engine] = variant
+	}
+	return out, rows.Err()
+}
+
+// SetNodeEngine chooses an engine's build for a node; an empty variant
+// leaves it to the node's hardware again.
+func (s *DB) SetNodeEngine(ctx context.Context, node, engine, variant, by string) error {
+	if variant == "" {
+		_, err := s.db.ExecContext(ctx, `DELETE FROM node_engines WHERE node_id = ? AND engine = ?`, node, engine)
+		return err
+	}
+	_, err := s.db.ExecContext(ctx,
+		`INSERT INTO node_engines (node_id, engine, variant, updated_at, updated_by) VALUES (?, ?, ?, ?, ?)
+		 ON CONFLICT (node_id, engine) DO UPDATE SET variant = excluded.variant,
+		   updated_at = excluded.updated_at, updated_by = excluded.updated_by`,
+		node, engine, variant, time.Now().UnixMilli(), by)
+	return err
+}
@@ -0,0 +1,24 @@
+package store
+
+import "testing"
+
+func TestAnEngineBuildIsChosenPerNodeAndCanBeLeftToItAgain(t *testing.T) {
+	db := capacityStore(t)
+	ctx := t.Context()
+	if err := db.SetNodeEngine(ctx, "gpu07", "vllm", "cu129", "ilker"); err != nil {
+		t.Fatal(err)
+	}
+	if err := db.SetNodeEngine(ctx, "gpu07", "vllm", "none", "ilker"); err != nil {
+		t.Fatal(err)
+	}
+	got, err := db.NodeEngines(ctx)
+	if err != nil || got["gpu07"]["vllm"] != "none" {
+		t.Fatalf("got %v, %v; the second choice replaces the first", got, err)
+	}
+	if err := db.SetNodeEngine(ctx, "gpu07", "vllm", "", "ilker"); err != nil {
+		t.Fatal(err)
+	}
+	if got, _ := db.NodeEngines(ctx); len(got["gpu07"]) != 0 {
+		t.Errorf("left to the node again, still %v", got)
+	}
+}
Author	SHA1	Message	Date
Ilker ManapandClaude Opus 5.5	bd36798b15	Sample GPU utilisation and memory every heartbeat, and chart them Release / build (push) Successful in 1m12s Details Release / notes (push) Successful in 4s Details CI / test (push) Successful in 3m52s Details 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>	2026-10-01 15:42:17 +02:00
Ilker ManapandClaude Opus 5.5	dc483d8c68	Choose each engine's build from the node's hardware, for every GPU generation Release / build (push) Successful in 1m12s Details CI / test (push) Successful in 3m52s Details Release / notes (push) Successful in 3s Details 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>	2026-10-01 14:35:58 +02:00
Ilker ManapandClaude Opus 5.5	43d7c4eac1	Let the upgrade repair what the field's catalog already holds Release / build (push) Successful in 1m11s Details Release / notes (push) Successful in 3s Details CI / test (push) Successful in 3m51s Details 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>	2026-10-01 14:20:11 +02:00
Ilker ManapandClaude Opus 5.5	86e9a1cb83	Give V100s the last vLLM built for them, not a build without their kernels Release / build (push) Successful in 1m13s Details Release / notes (push) Successful in 4s Details CI / test (push) Successful in 3m47s Details 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>	2026-10-01 14:04:45 +02:00
Ilker ManapandClaude Opus 5.5	4f280a3546	Give each GPU node a vLLM build it can run, chosen on the cluster page Release / build (push) Successful in 1m12s Details Release / notes (push) Successful in 3s Details CI / test (push) Successful in 3m55s Details 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>	2026-10-01 12:39:14 +02:00
Ilker ManapandClaude Opus 5.5	2735c5a318	Mirror a many-file model whole, with its checksums, for vLLM CI / test (push) Successful in 3m58s Details Release / build (push) Successful in 1m12s Details Release / notes (push) Successful in 3s Details 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>	2026-10-01 12:21:04 +02:00
Ilker ManapandClaude Opus 5.5	8957db30a2	Recognise a duplicate sign-up on PostgreSQL, and say what to do 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>	2026-10-01 12:15:37 +02:00
Ilker ManapandClaude Opus 5.5	2ef4220643	Fail a node only for the fabrics it declares 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>	2026-10-01 12:13:18 +02:00
Ilker ManapandClaude Opus 5.5	6014d1fee8	Say how a restart with no room beside it goes Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>	2026-10-01 12:11:46 +02:00
Ilker ManapandClaude Opus 5.5	a3ee48aced	Restart in place at once when only the old replica is in the way 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>	2026-10-01 12:11:37 +02:00
Ilker ManapandClaude Opus 5.5	e313f32696	Decide engines per node, and let a node join without one it cannot run 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>	2026-10-01 12:07:25 +02:00
Ilker ManapandClaude Opus 5.5	cba3e329fc	Let the mirror's serve form place a model on a GPU 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>	2026-10-01 11:59:22 +02:00
Ilker ManapandClaude Opus 5.5	cbd5521f4f	Place llama.cpp across several GPUs of a node with gpus_per_node 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>	2026-10-01 11:56:34 +02:00
Ilker ManapandClaude Opus 5.5	057be16a92	Do not call a GPU shard ready while it computes on the CPU 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>	2026-10-01 11:54:39 +02:00
Ilker ManapandClaude Opus 5.5	0c7934ba96	v0.10.6 Release / build (push) Successful in 1m11s Details Release / notes (push) Successful in 4s Details CI / test (push) Successful in 3m51s Details Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>	2026-10-01 08:52:01 +02:00
Ilker ManapandClaude Opus 5.5	e40b9ac202	Load LoRA adapters at scale zero Release / build (push) Successful in 1m17s Details CI / test (push) Successful in 4m0s Details Release / notes (push) Successful in 5s Details --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>	2026-10-01 07:39:12 +02:00
Ilker ManapandClaude Opus 5.5	323d172538	Attach LoRA adapters from the command line Release / build (push) Successful in 1m14s Details Release / notes (push) Successful in 4s Details CI / test (push) Successful in 3m52s Details 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>	2026-10-01 07:30:13 +02:00
Ilker ManapandClaude Opus 5.5	6b1fa65f59	Format kind_live_test.go Release / build (push) Successful in 1m12s Details Release / notes (push) Successful in 3s Details CI / test (push) Successful in 4m3s Details Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>	2026-10-01 07:20:35 +02:00
Ilker ManapandClaude Opus 5.5	663770226e	Find LoRA adapters on Hugging Face and attach them from the mirror page CI / test (push) Failing after 33s Details Release / build (push) Successful in 1m3s Details Release / notes (push) Successful in 3s Details 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>	2026-10-01 07:19:08 +02:00
Ilker ManapandClaude Opus 5.5	cf53904cc5	Serve LoRA adapters on llama.cpp models, asked for by name Release / notes (push) Successful in 3s Details Release / build (push) Successful in 1m11s Details CI / test (push) Successful in 3m47s Details 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>	2026-09-30 23:40:47 +02:00
Ilker ManapandClaude Opus 5.5	94eec557ad	Serve a model from an external OpenAI-compatible provider CI / test (push) Successful in 3m47s Details Release / build (push) Successful in 1m11s Details Release / notes (push) Successful in 3s Details 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>	2026-09-30 23:29:01 +02:00
Ilker ManapandClaude Opus 5.5	d5f02c0a4b	Spill a saturated pool's requests to the next pool CI / test (push) Successful in 3m40s Details 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>	2026-09-30 23:23:32 +02:00
Ilker ManapandClaude Opus 5.5	3d3c96e422	Displace the newest lowest-band request when the gateway queue is full Release / notes (push) Successful in 3s Details Release / build (push) Successful in 1m9s Details CI / test (push) Successful in 3m50s Details 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>	2026-09-30 23:07:53 +02:00
Ilker ManapandClaude Opus 5.5	623cdb5ba3	Sign releases, and install only signed ones once a key is trusted CI / test (push) Successful in 4m6s Details 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>	2026-09-30 23:03:48 +02:00
Ilker ManapandClaude Opus 5.5	561667c9be	Name only the replicas that would make room in headroom CI / test (push) Successful in 3m58s Details 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>	2026-09-30 23:01:40 +02:00
Ilker ManapandClaude Opus 5.5	cf0178bf05	Test that a restart borrows a lower class's room and returns it CI / test (push) Successful in 3m41s Details Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>	2026-09-30 22:47:12 +02:00
Ilker ManapandClaude Opus 5.5	f288f3c423	Set capacity quotas, class, owner and reservation from the admin pages Release / build (push) Successful in 1m9s Details Release / notes (push) Successful in 3s Details CI / test (push) Successful in 3m49s Details 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>	2026-09-30 22:40:45 +02:00