Mirror a many-file model whole, with its checksums, for vLLM

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>

CHANGELOG.md

@@ -20,6 +20,166 @@ 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.
+
+### 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.
+- **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 is refused as a source**, saying to mirror
+  it or to give the repository id for vLLM to fetch itself.
+
+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.
+
+### 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: a CUDA 12 vLLM on
+  Volta beside a CUDA 13 one on newer cards.
+- **The CUDA requirement follows the version.** vLLM before 0.26 and SGLang
+  before 0.5 are built for CUDA 12 and are installed on driver 525 and Volta;
+  the threshold had been CUDA 13's for every version, so the doctor's advice
+  to set an older version could not work. A refusal now says which versions
+  would.
+- **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 — and is now refused
+when the model is added, saying to write `gpus_per_node` instead; so is
+`nodes` above one.
+
 ## v0.10.6 — 2026-10-01
 
 Published first as release candidates v0.10.6-rc.1 to rc.8; rc.5 failed its

docs/install.md

@@ -376,8 +376,10 @@ 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;
+- **driver 580 or newer and a Turing GPU or newer** (compute capability 7.5)
+  for vLLM 0.26 and later, which are built for CUDA 13. A V100 cannot run
+  them; vLLM before 0.26 is built for CUDA 12, and runs on driver 525 and
+  Volta (7.0);
 - **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 +396,38 @@ 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.
+
+#### 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: v100-32g       # Volta, which CUDA 13 dropped
+    engines: { vllm: "0.25.1" }  # the last CUDA 12 vLLM
+  - 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:
@@ -1032,7 +1064,12 @@ 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
 
@@ -1216,9 +1253,9 @@ 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.
+- **vLLM 0.26 and later need CUDA 13**, driver 580 and a Turing GPU or newer.
+  Give older hardware a CUDA 12 vLLM (before 0.26) with `engine_overrides`
+  (step 4), or install one built for your CUDA yourself.
 - **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

docs/reference.md

@@ -81,6 +81,10 @@ 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) |
@@ -183,6 +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) |
@@ -645,6 +650,32 @@ 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`
@@ -1412,6 +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`

internal/agent/agent.go

@@ -450,7 +450,7 @@ 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()
@@ -931,6 +931,14 @@ 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

internal/agent/disabled_test.go

@@ -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)
+	}
+}

internal/agent/discovery/gpuprocs.go

@@ -0,0 +1,65 @@
+package discovery
+
+import (
+	"context"
+	"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 CUDA
+// 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.
+//
+// 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"
+	}
+	gpus, err := exec.CommandContext(ctx, binary, "--query-gpu=index,uuid", "--format=csv,noheader").Output()
+	if err != nil {
+		return nil, fmt.Errorf("running %s: %w", binary, err)
+	}
+	apps, err := exec.CommandContext(ctx, binary, "--query-compute-apps=gpu_uuid,pid", "--format=csv,noheader").Output()
+	if err != nil {
+		return nil, fmt.Errorf("running %s: %w", binary, err)
+	}
+	return parseGPUProcesses(string(gpus), string(apps)), nil
+}
+
+// parseGPUProcesses joins nvidia-smi's GPU list (index, uuid) to its compute
+// apps (gpu_uuid, pid): the apps are listed by UUID, placement speaks index.
+func parseGPUProcesses(gpus, apps string) map[int][]int {
+	index := map[string]int{}
+	for _, line := range strings.Split(gpus, "\n") {
+		i, uuid, ok := strings.Cut(line, ",")
+		if !ok {
+			continue
+		}
+		n, err := strconv.Atoi(strings.TrimSpace(i))
+		if err != nil {
+			continue
+		}
+		index[strings.TrimSpace(uuid)] = n
+	}
+	out := map[int][]int{}
+	for _, line := range strings.Split(apps, "\n") {
+		uuid, p, ok := strings.Cut(line, ",")
+		if !ok {
+			continue
+		}
+		pid, err := strconv.Atoi(strings.TrimSpace(p))
+		i, known := index[strings.TrimSpace(uuid)]
+		if err != nil || !known {
+			continue
+		}
+		out[i] = append(out[i], pid)
+	}
+	return out
+}

internal/agent/discovery/gpuprocs_test.go

@@ -0,0 +1,21 @@
+package discovery
+
+import (
+	"reflect"
+	"testing"
+)
+
+func TestGPUProcessesAreListedByIndexNotUUID(t *testing.T) {
+	gpus := "0, GPU-aaa\n1, GPU-bbb\n2, GPU-ccc\n"
+	// Two processes on GPU 1, one on GPU 2, and a GPU this list has never
+	// heard of, which is dropped rather than guessed at.
+	apps := "GPU-bbb, 4100\nGPU-ccc, 4200\nGPU-bbb, 4300\nGPU-zzz, 9999\n"
+	got := parseGPUProcesses(gpus, apps)
+	want := map[int][]int{1: {4100, 4300}, 2: {4200}}
+	if !reflect.DeepEqual(got, want) {
+		t.Errorf("got %v, want %v", got, want)
+	}
+	if got := parseGPUProcesses(gpus, ""); len(got) != 0 {
+		t.Errorf("no compute apps should be no GPU in use: %v", got)
+	}
+}

internal/agent/engines/engine.go

@@ -9,6 +9,7 @@
 package engines
 
 import (
+	"context"
 	"fmt"
 	"os"
 	"slices"
@@ -146,6 +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
@@ -203,6 +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"`
@@ -232,6 +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 {

internal/agent/engines/llamacpp.go

@@ -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.
@@ -234,3 +243,38 @@ 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
+}

internal/agent/engines/llamacpp_test.go

@@ -323,3 +323,41 @@ func TestAdaptersAreLoadedAtStartAndAppliedToNoneByDefault(t *testing.T) {
 		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")
+	}
+}

internal/agent/gpucheck.go

@@ -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, ", ")
+}

internal/agent/gpucheck_test.go

@@ -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)
+	}
+}

internal/agent/provision.go

@@ -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.
@@ -58,6 +79,15 @@ const provisionEvery = 10 * time.Minute
 func (a *Agent) setProvision(p *llmgwv1.Provision) {
 	wants := []provision.Want{}
 	for _, e := range p.GetEngines() {
+		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()})
 	}
 	a.mu.Lock()

internal/agent/provision/install.go

@@ -299,9 +299,55 @@ 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

internal/agent/provision/install_test.go

@@ -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
 	}
@@ -388,3 +396,32 @@ 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-" + cudaForLlamaCpp + "-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])
+	}
+}

internal/agent/provision/recipe.go

@@ -43,6 +43,11 @@ 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
 	// 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.
@@ -124,25 +129,72 @@ const minDriverForCUDA13 = 580
 // Volta, so a V100 cannot run these builds at all.
 const minComputeForCUDA13 = 7.5
 
-// cudaGate says whether a node can run an engine built on CUDA 13.
-func cudaGate(engine string, h discovery.Host) error {
+// The CUDA 12 builds before them run on any driver from 525, by CUDA's minor
+// version compatibility, and on Volta.
+const (
+	minDriverForCUDA12  = 525
+	minComputeForCUDA12 = 7.0
+)
+
+// cuda13Since is the first version of each Python engine built for CUDA 13.
+var cuda13Since = map[string]string{"vLLM": "0.26", "SGLang": "0.5"}
+
+// cudaGate says whether a node can run this version of an engine, and which
+// CUDA it is built for.
+//
+// By version, because the line moved: a V100 cannot run vLLM 0.26 or later,
+// and runs 0.25 — so a cluster with Volta and newer cards gives the Volta
+// nodes the older build (engine_overrides in the server's configuration)
+// rather than going without vLLM, or without the Volta nodes.
+func cudaGate(engine, version string, h discovery.Host) (variant string, err 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)
+	variant, cuda, minDriver, minCC := "cu13", "13", minDriverForCUDA13, minComputeForCUDA13
+	if versionBefore(version, cuda13Since[engine]) {
+		variant, cuda, minDriver, minCC = "cu12", "12", minDriverForCUDA12, minComputeForCUDA12
 	}
-	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)
+	if major := leadingInt(h.NVIDIADriver); major > 0 && major < minDriver {
+		return "", fmt.Errorf("%s %s is built for CUDA %s, which needs NVIDIA driver %d or newer; this node has %s",
+			engine, version, cuda, minDriver, h.NVIDIADriver)
 	}
-	return nil
+	if cc := h.MinComputeCapability; cc > 0 && cc < minCC {
+		hint := ""
+		if cuda == "13" {
+			hint = fmt.Sprintf("; %s before %s is built for CUDA 12 and runs on it", engine, cuda13Since[engine])
+		}
+		return "", fmt.Errorf("%s %s is built for CUDA %s, which supports GPUs of compute capability %.1f and newer; this node has %.1f%s",
+			engine, version, cuda, minCC, cc, hint)
+	}
+	return variant, nil
+}
+
+// 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
+	}
+	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
+		}
+	}
+	return false
 }
 
 // Python engines need a C compiler at run time: Triton builds a small
@@ -151,11 +203,12 @@ func cudaGate(engine string, h discovery.Host) error {
 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 {
+	variant, err := cudaGate("vLLM", version, h)
+	if err != nil {
 		return Build{}, err
 	}
 	return Build{
-		Variant:      "cu13",
+		Variant:      variant,
 		Python:       pythonForEngines,
 		Requirements: []string{"vllm==" + version},
 		Binary:       "vllm",
@@ -165,13 +218,14 @@ func chooseVLLM(version string, h discovery.Host) (Build, error) {
 }
 
 func chooseSGLang(version string, h discovery.Host) (Build, error) {
-	if err := cudaGate("SGLang", h); err != nil {
+	variant, err := cudaGate("SGLang", version, h)
+	if err != nil {
 		return Build{}, err
 	}
 	// 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",
+		Variant:      variant,
 		Python:       pythonForEngines,
 		Requirements: []string{"sglang==" + version},
 		Binary:       "python",
@@ -234,6 +288,7 @@ func chooseLlamaCpp(version string, h discovery.Host) (Build, error) {
 	if h.NVIDIADriver != "" && arch == "x64" {
 		name := fmt.Sprintf("bin-ubuntu-cuda-%s-x64.tar.gz", cudaForLlamaCpp)
 		b.Variant = "cuda-" + cudaForLlamaCpp
+		b.GPUBackend = "libggml-cuda.so"
 		b.Assets = []string{
 			fmt.Sprintf("llama-%s-%s", version, name),
 			fmt.Sprintf("cudart-llama-%s-%s", version, name),

internal/agent/weights/set.go

@@ -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
+}

internal/agent/weights/set_test.go

@@ -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)
+	}
+}

internal/cli/ctl_restart.go

@@ -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
 }
 

internal/cli/doctor.go

@@ -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

internal/cli/enginecheck.go

@@ -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
+}

internal/cli/mirror.go

@@ -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
 		}

internal/cli/node.go

@@ -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()

internal/cli/server.go

@@ -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"
@@ -496,8 +500,10 @@ func startControlPlane(ctx context.Context, cfg config.Server, cat *catalog.Live
 		}
 	}
 	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))
 	// 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

internal/config/config.go

@@ -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.
@@ -548,6 +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")
 	}
@@ -626,6 +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{}

internal/config/engines_test.go

@@ -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)
+		}
+	}
+}

internal/control/adminapi/api.go

@@ -208,10 +208,13 @@ 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"`
@@ -497,7 +500,8 @@ 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) {
@@ -1238,6 +1242,9 @@ 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.
@@ -1261,7 +1268,11 @@ 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
 	}

internal/control/agentsvc/installable_test.go

@@ -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 +26,31 @@ 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"}, reg)); got["vllm"] != "0.25.1" {
+		t.Errorf("the Volta node: %v", got)
+	}
+	if got := want(s.provisionFor(inventory.Node{ID: "hopper"}, 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"}, &llmgwv1.Register{InstallableEngines: []string{"llamacpp"}})); got["vllm"] != "" {
+		t.Errorf("asked a node for an engine it does not install: %v", got)
+	}
+}

internal/control/agentsvc/service.go

@@ -53,8 +53,9 @@ 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
 
 	// logWaits are the log requests awaiting a reply, by request id.
 	logMu    sync.Mutex
@@ -68,7 +69,14 @@ 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 }
 
 // conn is one agent's open stream.
 type conn struct {
@@ -179,15 +187,8 @@ 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{}
-		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]})
-		}
+	if s.enginesFor != nil {
+		p := s.provisionFor(node, reg)
 		if err := stream.Send(&llmgwv1.ControlMessage{
 			Msg: &llmgwv1.ControlMessage_Provision{Provision: p},
 		}); err != nil {
@@ -853,6 +854,21 @@ 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, reg *llmgwv1.Register) *llmgwv1.Provision {
+	p := &llmgwv1.Provision{}
+	can := installable(reg)
+	wants := s.enginesFor(node)
+	for _, e := range slices.Sorted(maps.Keys(wants)) {
+		if !slices.Contains(can, e) {
+			continue
+		}
+		p.Engines = append(p.Engines, &llmgwv1.EngineWant{Engine: e, Version: wants[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

internal/control/catalog/catalog.go

@@ -333,6 +333,13 @@ func (m Model) validate() error {
 	if m.Source == "" {
 		return fmt.Errorf("model %q: source is required", m.Name)
 	}
+	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.
+		return fmt.Errorf("model %q: source %s is a Hugging Face repository page, not its files; "+
+			"mirror it and use the version's source ending in /, or give %q for vLLM to fetch itself",
+			m.Name, m.Source, repo)
+	}
 	if err := validateAdapters(m); err != nil {
 		return err
 	}
@@ -381,6 +388,21 @@ 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, not by tp or pp: "+
+					"write gpus_per_node: %d and leave tp and pp out", m.Name, p.Name, max(p.TP, 1)*max(p.PP, 1))
+			}
+			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) {
@@ -704,3 +726,21 @@ func (m Model) engineAdapters() []engines.Adapter {
 	}
 	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
+}

internal/control/catalog/catalog_test.go

@@ -321,3 +321,48 @@ 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 is a width llama.cpp has no way to use; refused, saying what to
+	// write instead.
+	for _, p := range []Profile{
+		{Name: "tp", Engine: "llamacpp", GPUClass: "t4", TP: 4},
+		{Name: "nodes", Engine: "llamacpp", GPUClass: "t4", Nodes: 2, GPUsPerNode: 1},
+	} {
+		_, err := New([]Model{model("x", p)})
+		if err == nil {
+			t.Errorf("%s: accepted", p.Name)
+		}
+		if p.TP == 4 && (err == nil || !strings.Contains(err.Error(), "gpus_per_node: 4")) {
+			t.Errorf("tp 4 should point at gpus_per_node: %v", err)
+		}
+	}
+}
+
+func TestARepositoryPageIsNotASource(t *testing.T) {
+	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
+		_, err := New([]Model{m})
+		if err == nil || !strings.Contains(err.Error(), `"Qwen/Qwen2.5-32B-Instruct-AWQ"`) {
+			t.Errorf("%s: %v", src, err)
+		}
+	}
+	m := model("file")
+	m.Source = "https://huggingface.co/Qwen/X-GGUF/resolve/abc123/x.gguf"
+	if _, err := New([]Model{m}); err != nil {
+		t.Errorf("a file under /resolve/ is a source: %v", err)
+	}
+}

internal/control/reconciler/reconciler.go

@@ -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 {

internal/control/reconciler/reconciler_test.go

@@ -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++ {

internal/control/reconciler/restart.go

@@ -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 ""
 }

internal/doctor/doctor.go

@@ -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)
 	}
 
@@ -549,15 +588,19 @@ func vllmFindings(f Facts, add func(Level, string, string, string)) {
 		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(),
+			"give this node a vLLM it can run with engine_overrides in the server's configuration, "+
+				"or 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

internal/doctor/doctor_test.go

@@ -151,8 +151,14 @@ 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.
+		{"old driver", Facts{Host: discovery.Host{Arch: "amd64", NVIDIADriver: "550.54.15"}, VLLMVersion: "0.30.0"}, Warn, "driver 580"},
+		{"volta", Facts{Host: discovery.Host{Arch: "amd64", NVIDIADriver: "580.65.06", MinComputeCapability: 7.0}, VLLMVersion: "0.30.0"},
+			Warn, "before 0.26 is built for CUDA 12"},
+		{"volta, a CUDA 12 vLLM", Facts{Host: discovery.Host{Arch: "amd64", NVIDIADriver: "550.54.15", MinComputeCapability: 7.0},
+			VLLMVersion: "0.25.1", 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 +206,38 @@ 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")
+}

internal/hf/hf.go

@@ -258,6 +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

internal/mirror/add.go

@@ -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)

internal/mirror/mirror.go

@@ -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]+)$`)

internal/mirror/mirror_test.go

@@ -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)
+	}
+}

internal/onboard/onboard.go

@@ -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)
 		}

internal/onboard/onboard_test.go

@@ -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)
+	}
+}

internal/store/unique_test.go

@@ -0,0 +1,30 @@
+package store
+
+import (
+	"errors"
+	"fmt"
+	"testing"
+
+	"github.com/jackc/pgx/v5/pgconn"
+)
+
+func TestADuplicateIsRecognisedOnEitherDatabase(t *testing.T) {
+	for _, c := range []struct {
+		name string
+		err  error
+		want bool
+	}{
+		{"sqlite", errors.New("constraint failed: UNIQUE constraint failed: users.email (2067)"), true},
+		// What pgx returns, wrapped as database/sql passes it on: "unique"
+		// in lower case, which a check for "UNIQUE" missed.
+		{"postgres", fmt.Errorf("exec: %w", &pgconn.PgError{Code: "23505",
+			Message: `duplicate key value violates unique constraint "users_email_key"`}), true},
+		{"postgres, as text", errors.New(`ERROR: duplicate key value violates unique constraint "users_email_key" (SQLSTATE 23505)`), true},
+		{"another postgres error", &pgconn.PgError{Code: "23502", Message: "null value in column"}, false},
+		{"nothing", nil, false},
+	} {
+		if got := isUniqueViolation(c.err); got != c.want {
+			t.Errorf("%s: %v, want %v", c.name, got, c.want)
+		}
+	}
+}

internal/store/users.go

@@ -6,6 +6,8 @@ import (
 	"errors"
 	"strings"
 	"time"
+
+	"github.com/jackc/pgx/v5/pgconn"
 )
 
 // UserState is where an account is in its life.
@@ -66,12 +68,29 @@ func (s *DB) CreateUser(ctx context.Context, u User) error {
 		 VALUES (?, ?, ?, ?, ?, ?, ?)`,
 		u.ID, u.Email, u.Name, u.PasswordHash, string(u.State),
 		nullString(u.ProjectID), millis(u.CreatedAt))
-	if err != nil && strings.Contains(err.Error(), "UNIQUE") {
+	if isUniqueViolation(err) {
 		return ErrEmailTaken
 	}
 	return err
 }
 
+// isUniqueViolation reports whether err is a row refused for a value another
+// row already has. Each database says it its own way: SQLite in words
+// ("UNIQUE constraint failed"), PostgreSQL by code, 23505, in a message with
+// "unique" in lower case — which a check for the upper-case word missed, so a
+// second sign-up with an address on PostgreSQL was "something went wrong".
+func isUniqueViolation(err error) bool {
+	if err == nil {
+		return false
+	}
+	var pe *pgconn.PgError
+	if errors.As(err, &pe) {
+		return pe.Code == "23505"
+	}
+	m := err.Error()
+	return strings.Contains(m, "UNIQUE constraint failed") || strings.Contains(m, "SQLSTATE 23505")
+}
+
 // UserByEmail finds an account by its address.
 func (s *DB) UserByEmail(ctx context.Context, email string) (User, error) {
 	return s.scanUser(s.db.QueryRowContext(ctx,

internal/web/account.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"crypto/rand"
 	"encoding/base32"
+	"encoding/hex"
 	"errors"
 	"log/slog"
 	"net/http"
@@ -169,8 +170,12 @@ func (p *AccountPages) signup(w http.ResponseWriter, r *http.Request) {
 	case errors.Is(err, account.ErrWeakPassword), errors.Is(err, account.ErrInvalidEmail):
 		d.Error = err.Error()
 	default:
-		p.log.Error("signup failed", "err", err)
-		d.Error = "something went wrong; try again"
+		// Nothing of the database's words on the page; a reference that finds
+		// them in the log for whoever runs the gateway.
+		ref := errorRef()
+		p.log.Error("signup failed", "ref", ref, "err", err)
+		d.Error = "something went wrong on our side, and your account was not created. " +
+			"Try again later; if it keeps happening, give an operator this reference: " + ref
 	}
 	render(w, "signup", d)
 }
@@ -638,3 +643,11 @@ func (p *AccountPages) renderWithCSRF(w http.ResponseWriter, r *http.Request,
 
 // SetLlamaCppVersion tells the pages which llama.cpp build the nodes run.
 func (p *AccountPages) SetLlamaCppVersion(v string) { p.llamaVersion = v }
+
+// errorRef is a short reference for an error a page does not explain, logged
+// beside the error itself.
+func errorRef() string {
+	var b [4]byte
+	_, _ = rand.Read(b[:])
+	return "err-" + hex.EncodeToString(b[:])
+}

internal/web/admin_mirror.go

@@ -60,6 +60,9 @@ type repoSet struct {
 	// A shard set goes as one action — a fifteen-shard model is not fifteen
 	// decisions, it is one.
 	Sources []repoSource
+	// Companions are the files fetched with a safetensors set that are not
+	// weights: config.json, the tokenizer, the shard index.
+	Companions []string
 	// Model and Version are what the fetch form starts with. Editable, because
 	// the name a catalog entry uses is the operator's to choose and the
 	// registry's idea of it is only a good guess.
@@ -317,6 +320,16 @@ func (p *AccountPages) RegisterAdminMirror(mux *http.ServeMux, d MirrorDeps) {
 				if err != nil {
 					return "", err
 				}
+				// The CPU while GPUs sit free is a choice to make on purpose:
+				// made by default, it was a model running at a tenth of the
+				// speed with nothing saying why.
+				if model.Profiles[0].CPUOnly() && r.FormValue("cpu_ok") == "" {
+					if n := freeGPUs(p.fleet); n > 0 {
+						return "", fmt.Errorf("this fleet has %s free and %s was set to run on the CPU; "+
+							"choose a GPU placement, or tick \"run on the CPU anyway\"",
+							plural(n, "GPU", "GPUs"), model.Name)
+					}
+				}
 				// The same check `llmgw ctl model add` makes. This page builds
 				// the source out of a row the mirror reported, so it is right
 				// far more often than something typed — but a file removed
@@ -332,8 +345,8 @@ func (p *AccountPages) RegisterAdminMirror(mux *http.ServeMux, d MirrorDeps) {
 				// Placed on the next pass, not now: saying it is serving would
 				// be wrong for however long the weights take to reach a node.
 				return fmt.Sprintf(
-					"%s added to the catalog, %s; it is placed on the next pass",
-					model.Name, plural(model.Scaling.Min, "replica", "replicas")), nil
+					"%s added to the catalog, %s: %s; it is placed on the next pass",
+					model.Name, plural(model.Scaling.Min, "replica", "replicas"), describeServe(model)), nil
 			}))
 	}
 
@@ -768,6 +781,17 @@ func (p *AccountPages) openRepo(r *http.Request, d *adminMirrorData, repo string
 			rs.Sources = append(rs.Sources,
 				repoSource{URL: got.DownloadURL("", f.Path), Path: f.Path})
 		}
+		// Safetensors are not a model on their own: vLLM loads a directory,
+		// and reads its config and tokenizer before any weight. Fetched as
+		// part of the set, or the set arrives as shards nothing can load.
+		if strings.HasSuffix(strings.ToLower(set.Name), ".safetensors") {
+			for _, f := range hf.Companions(got.Files) {
+				rs.Sources = append(rs.Sources,
+					repoSource{URL: got.DownloadURL("", f.Path), Path: f.Path})
+				rs.Companions = append(rs.Companions, f.Path)
+				rs.Bytes += f.Bytes
+			}
+		}
 		d.RepoSets = append(d.RepoSets, rs)
 	}
 	p.probeKinds(r.Context(), d.RepoSets)

internal/web/admin_mirror_test.go

@@ -710,3 +710,37 @@ func TestARefusedFetchReturnsToItsFormAndAStartedOneToTheList(t *testing.T) {
 		t.Errorf("a started download should be shown where it is followed: %s", body)
 	}
 }
+
+func TestASafetensorsSetIsFetchedWithWhatTheEngineReadsBesideIt(t *testing.T) {
+	// The field: a four-shard AWQ model mirrored as its shards alone, and
+	// vLLM failing at start for want of config.json.
+	reg := &fakeRegistry{repo: hf.Repo{
+		ID: "Qwen/Qwen2.5-32B-Instruct-AWQ", Revision: "abc123def456",
+		Files: []hf.File{
+			{Path: "model-00001-of-00002.safetensors", Bytes: 4_000_000_000},
+			{Path: "model-00002-of-00002.safetensors", Bytes: 3_000_000_000},
+			{Path: "model.safetensors.index.json", Bytes: 90_000},
+			{Path: "config.json", Bytes: 900},
+			{Path: "tokenizer.json", Bytes: 7_000_000},
+			{Path: "tokenizer_config.json", Bytes: 7_000},
+			{Path: "README.md", Bytes: 9_000},
+			{Path: "modeling_custom.py", Bytes: 40_000},
+			{Path: "onnx/config.json", Bytes: 900},
+		},
+	}}
+	body := mirrorGet(t, mirrorPageWith(t, &fakeMirror{}, reg, true),
+		"/admin/mirror?repo=Qwen/Qwen2.5-32B-Instruct-AWQ").Body.String()
+	for _, want := range []string{"config.json", "tokenizer.json", "tokenizer_config.json", "model.safetensors.index.json"} {
+		if !strings.Contains(body, `name="file" value="`+want+`"`) {
+			t.Errorf("%s is not fetched with the set", want)
+		}
+	}
+	for _, not := range []string{"README.md", "modeling_custom.py", "onnx/config.json"} {
+		if strings.Contains(body, `name="file" value="`+not+`"`) {
+			t.Errorf("%s was swept in with the set", not)
+		}
+	}
+	if !strings.Contains(body, "with 4 files the engine reads beside them") {
+		t.Errorf("the companions are not said; got %s", body)
+	}
+}

internal/web/admin_serve.go

@@ -43,8 +43,20 @@ type serveForm struct {
 	Engine  string
 	// Profile is the profile name, which doubles as how a node_selector would
 	// be written later. Named for the hardware rather than the model, because
-	// a profile is about where a model runs.
+	// a profile is about where a model runs. Derived from the placement, not
+	// typed: a name is only a name, and a field for it read as though calling
+	// a profile "gpu" would put it on one.
 	Profile string
+	// Placements are where it can run: each GPU class the fleet has, and for
+	// llama.cpp the CPU. One is selected.
+	Placements []placeChoice
+	// GPUs is how many cards a replica takes, when placed on a GPU class.
+	GPUs int
+	// Parallel is how many requests one replica serves at once.
+	Parallel int
+	// FreeGPUs is how many cards no replica holds, across the fleet: what
+	// choosing the CPU leaves unused.
+	FreeGPUs int
 	// Files is how many files the version holds, above one for a set.
 	Files int
 	// MinMemoryGB is the host RAM to ask for, estimated from the weights.
@@ -126,6 +138,13 @@ func planServe(st mirror.Status, files []mirror.Status, base string,
 	default:
 		f.Profile = "gpu"
 	}
+	f.GPUs, f.Parallel = 1, defaultParallel
+	f.Placements, f.FreeGPUs = placementsFor(fleet, f.Engine)
+	for _, pc := range f.Placements {
+		if pc.Selected {
+			f.Profile = pc.profile()
+		}
+	}
 	// Whether any node can actually run it. Agents report their installed
 	// engines now, so this is a fact rather than the guess it started as —
 	// which was cards, and would have passed a GPU node with no vLLM on it.
@@ -274,12 +293,38 @@ func buildModel(get func(string) string) (catalog.Model, error) {
 	if engine == "" {
 		return m, errors.New("which engine should run it?")
 	}
-	p := catalog.Profile{
-		Name: strings.TrimSpace(get("profile")), Engine: engine,
-		Weight: 1, Priority: 100,
+	p := catalog.Profile{Engine: engine, Weight: 1, Priority: 100}
+	gpus := 1
+	if v := strings.TrimSpace(get("gpus")); v != "" {
+		if gpus, err = positive(v, "the GPUs per replica"); err != nil {
+			return m, err
+		}
+		gpus = max(gpus, 1)
 	}
-	if p.Name == "" {
-		p.Name = engine
+	place := placeChoice{Value: strings.TrimSpace(get("placement"))}
+	switch class, onGPU := strings.CutPrefix(place.Value, "gpu:"); {
+	case onGPU && class != "":
+		// Placement is decided by gpu_class, which is what the selector
+		// sets; the count is gpus_per_node for llama.cpp, which spreads
+		// layers over a node's cards, and tp for the others.
+		p.GPUClass = class
+		if engine == "llamacpp" {
+			p.GPUsPerNode = gpus
+		} else {
+			p.TP = gpus
+		}
+		place.GPUs = gpus
+	case place.Value == "" || place.Value == "cpu":
+		if engine != "llamacpp" {
+			return m, fmt.Errorf("%s runs on GPUs; choose a GPU class to place it on", engine)
+		}
+		place.Value = "cpu"
+	default:
+		return m, fmt.Errorf("%q is not a placement", place.Value)
+	}
+	p.Name = place.profile()
+	if v := strings.TrimSpace(get("profile")); v != "" {
+		p.Name = v
 	}
 	// The pool, not the per-request limit: llama.cpp divides it across its
 	// parallel slots, so the number a client may use is smaller and is what
@@ -290,6 +335,15 @@ func buildModel(get func(string) string) (catalog.Model, error) {
 		}
 		p.Args = []string{"--ctx-size", v}
 	}
+	if v := strings.TrimSpace(get("parallel")); v != "" && engine == "llamacpp" {
+		n, err := positive(v, "the parallel requests")
+		if err != nil {
+			return m, err
+		}
+		if n > 0 {
+			p.Args = append(p.Args, "--parallel", strconv.Itoa(n))
+		}
+	}
 	if v := strings.TrimSpace(get("min_memory_gb")); v != "" {
 		if p.MinMemoryGB, err = positive(v, "the memory requirement"); err != nil {
 			return m, err
@@ -408,3 +462,125 @@ func modeOf(sc catalog.Scaling) string {
 // defaultCtxPool is the llama.cpp context pool the serve form starts with.
 // --parallel divides it among slots; it does not multiply it.
 const defaultCtxPool = 16384
+
+// defaultParallel is the requests a llama.cpp replica serves at once when
+// the form is not changed: the agent's own default.
+const defaultParallel = 4
+
+// placeChoice is one place a model can run: a GPU class, or the CPU.
+type placeChoice struct {
+	// Value is "gpu:<class>" or "cpu", as the form submits it.
+	Value string
+	Label string
+	// GPUs is how many cards a replica takes, for naming the profile.
+	GPUs     int
+	Selected bool
+}
+
+// profile names the profile for where it runs: the class, with the count when
+// it is more than one card ("t4", "t4-x4"), or "cpu".
+func (c placeChoice) profile() string {
+	class, ok := strings.CutPrefix(c.Value, "gpu:")
+	if !ok || class == "" {
+		return "cpu"
+	}
+	if c.GPUs > 1 {
+		return fmt.Sprintf("%s-x%d", class, c.GPUs)
+	}
+	return class
+}
+
+// placementsFor lists where engine can run on this fleet, with one selected,
+// and how many GPUs are free.
+//
+// A GPU when the fleet has one: a GGUF model was put on the CPU by this form
+// on a cluster of GPUs, ran at a tenth of their speed, and nothing on the way
+// said why — the form had no choice to offer and named the profile "cpu". The
+// class selected is the one with the most cards free.
+func placementsFor(fleet Source, engine string) ([]placeChoice, int) {
+	var out []placeChoice
+	free, best := 0, -1
+	if fleet != nil {
+		for _, c := range fleet.Summary().GPUs {
+			n := c.Free()
+			free += n
+			out = append(out, placeChoice{
+				Value: "gpu:" + c.Class,
+				Label: fmt.Sprintf("GPU: %s, %d of %d free", c.Class, n, c.Total),
+			})
+			if best < 0 || n > fleet.Summary().GPUs[best].Free() {
+				best = len(out) - 1
+			}
+		}
+	}
+	if best >= 0 {
+		out[best].Selected = true
+	}
+	if engine == "llamacpp" {
+		label := "CPU only: no GPU, many times slower"
+		if len(out) == 0 {
+			label = "CPU: this fleet has no GPU"
+		}
+		out = append(out, placeChoice{Value: "cpu", Label: label, Selected: best < 0})
+	}
+	return out, free
+}
+
+// freeGPUs is how many cards no replica holds across the fleet.
+func freeGPUs(fleet Source) int {
+	if fleet == nil {
+		return 0
+	}
+	n := 0
+	for _, c := range fleet.Summary().GPUs {
+		n += c.Free()
+	}
+	return n
+}
+
+// describeServe says, in one line, what a model will run as: the summary
+// shown before it is added and in the message after.
+func describeServe(m catalog.Model) string {
+	if len(m.Profiles) == 0 {
+		return ""
+	}
+	p := m.Profiles[0]
+	engine := map[string]string{"llamacpp": "llama.cpp", "vllm": "vLLM", "sglang": "SGLang"}[p.Engine]
+	if engine == "" {
+		engine = p.Engine
+	}
+	where := "on the CPU"
+	if p.GPUClass != "" {
+		n := max(p.GPUsPerNode, p.TP, 1)
+		where = fmt.Sprintf("on %s %s", plural(n, "GPU", "GPUs"), p.GPUClass)
+		if n > 1 {
+			where = fmt.Sprintf("on %d %s GPUs of one node", n, p.GPUClass)
+		}
+	}
+	parts := []string{engine + " " + where}
+	if p.MinMemoryGB > 0 {
+		parts = append(parts, fmt.Sprintf("%d GB of host memory", p.MinMemoryGB))
+	}
+	if p.GPUClass == "" && p.CPUCores > 0 {
+		parts = append(parts, plural(p.CPUCores, "core", "cores"))
+	}
+	pool, parallel := argInt(p.Args, "--ctx-size"), argInt(p.Args, "--parallel")
+	switch {
+	case pool > 0 && parallel > 0:
+		parts = append(parts, fmt.Sprintf("context pool %d over %d parallel requests (%d each)", pool, parallel, pool/parallel))
+	case pool > 0:
+		parts = append(parts, fmt.Sprintf("context pool %d", pool))
+	}
+	return strings.Join(parts, ", ")
+}
+
+// argInt is the number after flag in args, or 0.
+func argInt(args []string, flag string) int {
+	for i := 0; i+1 < len(args); i++ {
+		if args[i] == flag {
+			n, _ := strconv.Atoi(args[i+1])
+			return n
+		}
+	}
+	return 0
+}

internal/web/admin_serve_test.go

@@ -71,6 +71,8 @@ type fakeFleet struct {
 	gpus     int
 	engines  []string
 	memoryGB int
+	// used is how many of the cards replicas hold.
+	used int
 }
 
 func (f fakeFleet) Nodes() []adminapi.NodeView {
@@ -81,7 +83,12 @@ func (f fakeFleet) Nodes() []adminapi.NodeView {
 func (f fakeFleet) Models() []adminapi.ModelView           { return nil }
 func (f fakeFleet) Replicas(string) []adminapi.ReplicaView { return nil }
 func (f fakeFleet) Explain(string) adminapi.ExplainView    { return adminapi.ExplainView{} }
-func (f fakeFleet) Summary() adminapi.Summary              { return adminapi.Summary{} }
+func (f fakeFleet) Summary() adminapi.Summary {
+	if f.gpus == 0 {
+		return adminapi.Summary{}
+	}
+	return adminapi.Summary{GPUs: []adminapi.ClassUse{{Class: "t4", Total: f.gpus, Used: f.used}}}
+}
 
 func held(model, version, file string) mirror.Status {
 	return mirror.Status{Model: model, Version: version, File: file,
@@ -452,3 +459,90 @@ func TestServingOnDemandLoadsWhenAskedFor(t *testing.T) {
 		t.Errorf("on demand was added as %+v", sc)
 	}
 }
+
+func TestAGGUFOnAFleetWithGPUsIsOfferedTheGPUFirst(t *testing.T) {
+	gguf := []mirror.Status{held("coder", "v1", "coder-q4_k_m.gguf")}
+	body := mirrorGet(t, servePage(t, gguf, &fakeCatalog{}, 4, "llamacpp"),
+		"/admin/mirror?serve=coder:v1").Body.String()
+	// Found in the field: no choice offered, a profile called "cpu", and a
+	// model on the CPU of a GPU cluster at a tenth of the speed.
+	if !strings.Contains(body, `<option value="gpu:t4" selected>GPU: t4, 4 of 4 free</option>`) {
+		t.Errorf("the free GPU class should be offered and chosen; got %s", body)
+	}
+	if !strings.Contains(body, `<option value="cpu">CPU only`) {
+		t.Error("the CPU should still be on offer, unselected")
+	}
+	if !strings.Contains(body, `name="cpu_ok"`) {
+		t.Error("choosing the CPU with GPUs free needs saying so")
+	}
+
+	cpuOnly := mirrorGet(t, servePage(t, gguf, &fakeCatalog{}, 0, "llamacpp"),
+		"/admin/mirror?serve=coder:v1").Body.String()
+	if !strings.Contains(cpuOnly, `<option value="cpu" selected>CPU: this fleet has no GPU</option>`) ||
+		strings.Contains(cpuOnly, `name="cpu_ok"`) {
+		t.Errorf("a fleet without GPUs has only the CPU, and nothing to confirm; got %s", cpuOnly)
+	}
+}
+
+func TestTheChosenPlacementDecidesWhereItRuns(t *testing.T) {
+	form := func(engine, place, gpus string) url.Values {
+		return url.Values{
+			"name": {"m"}, "source": {"http://x/m"}, "engine": {engine}, "replicas": {"1"},
+			"context_length": {"4096"}, "ctx_size": {"16384"}, "parallel": {"4"},
+			"placement": {place}, "gpus": {gpus}, "min_memory_gb": {"8"},
+		}
+	}
+	m, err := buildModel(form("llamacpp", "gpu:t4", "4").Get)
+	if err != nil {
+		t.Fatal(err)
+	}
+	p := m.Profiles[0]
+	if p.GPUClass != "t4" || p.GPUsPerNode != 4 || p.TP > 1 || p.Name != "t4-x4" {
+		t.Errorf("llama.cpp on four t4s: %+v", p)
+	}
+	if got := describeServe(m); got != "llama.cpp on 4 t4 GPUs of one node, 8 GB of host memory, "+
+		"context pool 16384 over 4 parallel requests (4096 each)" {
+		t.Errorf("summary %q", got)
+	}
+	if m, _ := buildModel(form("vllm", "gpu:t4", "2").Get); m.Profiles[0].TP != 2 || m.Profiles[0].GPUsPerNode != 0 {
+		t.Errorf("vLLM's GPU count is tp: %+v", m.Profiles[0])
+	}
+	if m, _ := buildModel(form("llamacpp", "cpu", "1").Get); m.Profiles[0].GPUClass != "" || m.Profiles[0].Name != "cpu" {
+		t.Errorf("the CPU: %+v", m.Profiles[0])
+	}
+	if _, err := buildModel(form("vllm", "cpu", "1").Get); err == nil {
+		t.Error("vLLM on the CPU was accepted")
+	}
+}
+
+func TestTheCPUWithGPUsFreeIsAChoiceMadeOnPurpose(t *testing.T) {
+	values := func(extra ...string) url.Values {
+		v := url.Values{
+			"csrf": {"tok"}, "name": {"coder"}, "source": {reachableSource(t)}, "engine": {"llamacpp"},
+			"replicas": {"1"}, "context_length": {"4096"}, "placement": {"cpu"}, "min_memory_gb": {"8"},
+		}
+		for i := 0; i+1 < len(extra); i += 2 {
+			v.Set(extra[i], extra[i+1])
+		}
+		return v
+	}
+	cat := &fakeCatalog{}
+	page := mirrorPageDeps(t, MirrorDeps{Mirror: &fakeMirror{}, Catalog: cat,
+		Fleet: fakeFleet{gpus: 4, used: 1, engines: []string{"llamacpp"}}}, true)
+	body := mirrorPost(t, page, "/admin/mirror/serve", values()).Body.String()
+	if len(cat.added) != 0 || !strings.Contains(body, "3 GPUs free") {
+		t.Fatalf("the CPU with three GPUs free went through unasked: %s", body)
+	}
+	mirrorPost(t, page, "/admin/mirror/serve", values("cpu_ok", "yes"))
+	if len(cat.added) != 1 {
+		t.Fatal("ticking the box should add it")
+	}
+	// Every card held: nothing to ask.
+	cat = &fakeCatalog{}
+	full := mirrorPageDeps(t, MirrorDeps{Mirror: &fakeMirror{}, Catalog: cat,
+		Fleet: fakeFleet{gpus: 4, used: 4, engines: []string{"llamacpp"}}}, true)
+	mirrorPost(t, full, "/admin/mirror/serve", values())
+	if len(cat.added) != 1 {
+		t.Error("with no GPU free the CPU is the only choice and should not be questioned")
+	}
+}

internal/web/signup_test.go

@@ -0,0 +1,54 @@
+package web
+
+import (
+	"context"
+	"errors"
+	"log/slog"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"strings"
+	"testing"
+)
+
+// signupAccounts is an account service whose sign-up answers with err.
+type signupAccounts struct {
+	fakeAccounts
+	err error
+}
+
+func (s signupAccounts) SignUp(context.Context, string, string, string) error { return s.err }
+func (signupAccounts) SignupOpen() bool                                       { return true }
+
+func signupBody(t *testing.T, err error) string {
+	t.Helper()
+	mux := http.NewServeMux()
+	NewAccountPages(signupAccounts{err: err}, nil, nil, "v0", slog.New(slog.DiscardHandler)).Register(mux)
+	form := url.Values{"email": {"a@b.c"}, "name": {"A"}, "password": {"long enough password"}}
+	req := httptest.NewRequest(http.MethodPost, "/signup", strings.NewReader(form.Encode()))
+	req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+	rec := httptest.NewRecorder()
+	mux.ServeHTTP(rec, req)
+	return rec.Body.String()
+}
+
+func TestASignupThatFailsSaysSoWithAReferenceAndNoDatabaseWords(t *testing.T) {
+	body := signupBody(t, errors.New(`duplicate key value violates unique constraint "users_email_key"`))
+	if strings.Contains(body, "users_email_key") || strings.Contains(body, "constraint") {
+		t.Errorf("the database's own words reached the page: %s", body)
+	}
+	if !strings.Contains(body, "was not created") || !strings.Contains(body, "reference: err-") {
+		t.Errorf("no plain failure with a reference to give an operator: %s", body)
+	}
+}
+
+func TestASignupSaysWhatToDoIfTheAddressHadAnAccount(t *testing.T) {
+	// The service answers a taken address as it does a new one, so nobody
+	// can find out from this page who has an account; what the page can do
+	// is say what to do in either case.
+	body := signupBody(t, nil)
+	if !strings.Contains(body, "already had an account") || !strings.Contains(body, `href="/login"`) ||
+		!strings.Contains(body, "reset it") {
+		t.Errorf("the done page leaves someone who already had an account stuck: %s", body)
+	}
+}

internal/web/templates/account.html

@@ -11,6 +11,11 @@
 <p class="note">You will not get an email from this gateway — it does not send any.
 Tell whoever you asked for access that you have signed up.</p>
 <p><a href="/login">Sign in</a> once you have been approved.</p>
+{{/* The same page whether or not the address already had an account, so
+     it says what to do in either case without saying which this was. */}}
+<p class="note">If this address already had an account, nothing new was made:
+<a href="/login">sign in</a> with it, or, if you have forgotten its password, ask
+an operator to reset it.</p>
 {{else}}
 {{if .Error}}<p class="error">{{.Error}}</p>{{end}}
 <form method="post" autocomplete="on">

internal/web/templates/admin_mirror.html

@@ -131,7 +131,7 @@ reachable from this machine.</p>
 <p class="error">Cannot be served as it stands: {{.Refusal}}.</p>
 {{else}}
 {{if .Warning}}<p class="note error-row">{{.Warning}}.</p>{{end}}
-<form method="post" action="/admin/mirror/serve">
+<form method="post" action="/admin/mirror/serve" class="serve-form" data-engine="{{.Engine}}" data-free-gpus="{{.FreeGPUs}}">
   <input type="hidden" name="csrf" value="{{$.CSRF}}">
   <input type="hidden" name="source" value="{{.Source}}">
   <input type="hidden" name="engine" value="{{.Engine}}">
@@ -163,8 +163,23 @@ reachable from this machine.</p>
     <label>Engine context pool <span class="opt">divided across slots</span>
       <input type="text" name="ctx_size" value="{{.CtxPool}}"></label>
   </div>
+  {{/* Where it runs is a choice with consequences, so it is a choice: the
+       GPU classes the fleet has, and the CPU for llama.cpp. A profile name
+       field stood here once, and reading as though naming a profile "gpu"
+       would put it on one, it put GGUF models on the CPU of GPU clusters. */}}
+  <div class="field-row">
+    <label>Runs on
+      <select name="placement">
+        {{range .Placements}}<option value="{{.Value}}"{{if .Selected}} selected{{end}}>{{.Label}}</option>{{end}}
+      </select></label>
+    <label>GPUs per replica <span class="opt">on one node</span>
+      <input type="text" name="gpus" value="{{.GPUs}}" size="3"></label>
+    {{if eq .Engine "llamacpp"}}<label>Parallel requests <span class="opt">per replica</span>
+      <input type="text" name="parallel" value="{{.Parallel}}" size="3"></label>{{end}}
+  </div>
+  {{if .FreeGPUs}}<label class="cpu-ok"><input type="checkbox" name="cpu_ok" value="yes">
+    Run on the CPU anyway: {{.FreeGPUs}} GPUs are free and would sit unused</label>{{end}}
   <div class="field-row">
-    <label>Profile name <input type="text" name="profile" value="{{.Profile}}"></label>
     <label>Needs RAM (GB) <span class="opt">for the context pool beside it</span>
       <input type="text" name="min_memory_gb" value="{{.MinMemoryGB}}" required></label>
     <label>CPU cores <span class="opt">reserved; the engine's threads</span>
@@ -173,8 +188,44 @@ reachable from this machine.</p>
       <input type="text" name="selector_key" placeholder="role" size="8">
       <input type="text" name="selector_value" placeholder="cpu-large" size="10"></label>
   </div>
+  <p class="sub serve-summary" hidden></p>
   <button type="submit">Add to the catalog</button>
 </form>
+<script>
+// What it will run as, in one line, kept up to date as the form changes:
+// engine, where, how many cards, memory, context and how it is shared.
+(function () {
+  const f = document.currentScript.previousElementSibling;
+  if (!f || !f.classList.contains("serve-form")) return;
+  const out = f.querySelector(".serve-summary"), ok = f.querySelector(".cpu-ok");
+  const v = (n) => (f.elements[n] ? f.elements[n].value.trim() : "");
+  const engine = { llamacpp: "llama.cpp", vllm: "vLLM", sglang: "SGLang" }[f.dataset.engine] || f.dataset.engine;
+  function draw() {
+    const place = v("placement"), gpus = parseInt(v("gpus"), 10) || 1;
+    const onCPU = place === "cpu" || place === "";
+    let where = "on the CPU";
+    if (!onCPU) {
+      const cls = place.replace(/^gpu:/, "");
+      where = gpus > 1 ? "on " + gpus + " " + cls + " GPUs of one node" : "on 1 GPU " + cls;
+    }
+    const parts = [engine + " " + where];
+    if (v("min_memory_gb")) parts.push(v("min_memory_gb") + " GB of host memory");
+    if (onCPU && v("cpu_cores")) parts.push(v("cpu_cores") + " cores");
+    const pool = parseInt(v("ctx_size"), 10), par = parseInt(v("parallel"), 10);
+    if (pool && par) parts.push("context pool " + pool + " over " + par + " parallel requests (" + Math.floor(pool / par) + " each)");
+    else if (pool) parts.push("context pool " + pool);
+    out.textContent = "Will run as: " + parts.join(", ") + ".";
+    const warn = onCPU && +f.dataset.freeGpus > 0;
+    out.classList.toggle("error", warn);
+    if (warn) out.textContent += " " + f.dataset.freeGpus + " GPUs are free; on the CPU this runs many times slower.";
+    if (ok) ok.hidden = !onCPU;
+    out.hidden = false;
+  }
+  f.addEventListener("input", draw);
+  f.addEventListener("change", draw);
+  draw();
+})();
+</script>
 <p class="note">The context pool is what the engine is given; llama.cpp divides
 it across its parallel slots, so the number one request may use is the smaller
 one above.</p>
@@ -416,6 +467,7 @@ version and the old weights stay where they are.</p>
 {{range .RepoSets}}<tr>
   <td><code>{{.Name}}</code>
     {{if .Shards}}<span class="opt">{{.Shards}} shards, fetched together</span>{{end}}
+    {{with .Companions}}<span class="opt" title="{{join .}}">with {{len .}} files the engine reads beside them</span>{{end}}
     {{if eq .Kind "adapter"}}<span class="tag new" title="a LoRA adapter: it runs on a {{.Arch}} model it was made for, not on its own">LoRA adapter · {{.Arch}}</span>
     {{else if eq .Kind "model"}}<span class="opt">model{{with .Arch}} · {{.}}{{end}}</span>{{end}}</td>
   <td class="num">{{bytes .Bytes}}</td>
@@ -441,7 +493,8 @@ version and the old weights stay where they are.</p>
 </tr>{{end}}
 </table>
 <p class="note">Only the weights are listed; a repository is mostly a README and
-a tokenizer. Shards of one file are one row and one button — a fifteen-shard
+a tokenizer. A safetensors set is fetched with the files vLLM reads beside it —
+config.json, the tokenizer, the shard index — and never with code. Shards of one file are one row and one button — a fifteen-shard
 model is one decision, not fifteen. Two separate files are two rows because they
 are usually alternatives: one quantisation or the other, not both.</p>
 <p class="note">The source URLs are pinned to

internal/web/templates/node.html

@@ -82,9 +82,9 @@ the adapter came up: a number that grows is the one to look at.</p>
   </tr>{{end}}
 </table>
 <table class="rows compact gpu-list">
-  <tr><th>GPU</th><th>Name</th><th class="num">NUMA</th><th>Nearest adapter</th><th>Multi-node NVLink</th></tr>
+  <tr><th>GPU</th><th>Name</th><th class="num">Compute</th><th class="num">NUMA</th><th>Nearest adapter</th><th>Multi-node NVLink</th></tr>
   {{range .Node.GPUs}}<tr>
-    <td>{{.Index}}</td><td>{{.Name}}</td><td class="num">{{.NUMA}}</td>
+    <td>{{.Index}}</td><td>{{.Name}}</td><td class="num">{{.ComputeCapability}}</td><td class="num">{{.NUMA}}</td>
     <td>{{if .NearestNIC}}<code>{{.NearestNIC}}</code> <span class="note">{{.NICPath}}</span>{{else}}—{{end}}</td>
     <td>{{if .Domain}}<span class="state-ready">joined</span>{{if and .Bandwidth (ne .Bandwidth "Full")}} <span class="bad">{{.Bandwidth}} bandwidth</span>{{end}}{{else if .FabricState}}<span class="bad">{{.FabricState}}</span>{{else}}<span class="note">—</span>{{end}}</td>
   </tr>{{end}}