v0.16.3: changelog and README

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

.gitea/workflows/ci.yml

@@ -34,11 +34,23 @@ jobs:
         run: go vet ./...
 
       - name: Test with the race detector
-        run: go test -race ./...
+        # Under the job's limit, so a package that hangs panics with the
+        # test it was in and where every goroutine waited, rather than the
+        # job ending at its deadline with nothing said: from 2026-10-15 every
+        # run did, printing nothing after internal/release.
+        run: go test -race -timeout 12m ./...
 
       - name: Build
         run: go build ./...
 
+      # llmgw-kb runs on people's own machines, Windows included, which no
+      # test here runs on: building it for each catches a Unix-only call.
+      - name: Build llmgw-kb for every system
+        run: |
+          for target in windows/amd64 windows/arm64 darwin/amd64 darwin/arm64 linux/amd64 linux/arm64; do
+            CGO_ENABLED=0 GOOS="${target%/*}" GOARCH="${target#*/}" go build -o /dev/null ./cmd/llmgw-kb
+          done
+
       # The Terraform provider is a module of its own, which ./... above does
       # not reach. Its tests check its client against docs/admin-api.openapi.json,
       # so a change to the admin API that breaks the provider fails here. The

.gitea/workflows/engine-releases.yml

@@ -0,0 +1,20 @@
+# Fails when an engine llmgw follows release by release (FreeToken) has
+# published a version newer than the one DefaultEngines installs: the failed
+# run is the notice to check the new version and move to it. See
+# scripts/engine_releases.py.
+name: Engine releases
+
+on:
+  schedule:
+    # Mondays 06:00 UTC.
+    - cron: "0 6 * * 1"
+  workflow_dispatch:
+
+jobs:
+  check:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Compare with PyPI
+        run: python3 scripts/engine_releases.py

.gitea/workflows/release.yml

@@ -55,6 +55,21 @@ jobs:
             (cd dist && sha256sum "${name}" >> checksums.txt)
           done
 
+          # llmgw-kb, which people run on their own machines to put their
+          # documents into a knowledge base: Windows too, which llmgw itself
+          # does not build for.
+          for target in windows/amd64 windows/arm64 darwin/amd64 darwin/arm64 linux/amd64 linux/arm64; do
+            os="${target%/*}"; arch="${target#*/}"
+            ext=""; [ "$os" = windows ] && ext=".exe"
+            name="llmgw-kb-${version}-${os}-${arch}${ext}"
+            echo "building $name"
+            CGO_ENABLED=0 GOOS="$os" GOARCH="$arch" go build \
+              -trimpath \
+              -ldflags "-s -w -X gitea.mynodes.xyz/ilker/llmgw/internal/version.Version=${version}" \
+              -o "dist/${name}" ./cmd/llmgw-kb
+            (cd dist && sha256sum "${name}" >> checksums.txt)
+          done
+
           # The reverse proxy configurations, so a release page carries
           # everything install.md asks for and the repository is not needed.
           for f in deploy/proxy/Caddyfile deploy/proxy/nginx.conf; do
@@ -91,6 +106,7 @@ jobs:
           # test the cross-compiled ones, which is why the native build is
           # checked rather than assumed.
           ./dist/llmgw-${version}-linux-amd64 --version
+          ./dist/llmgw-kb-${version}-linux-amd64 --version
           cat dist/checksums.txt
 
       # Signed, so a machine that trusts the key built into it installs

.gitignore

@@ -9,3 +9,5 @@ coverage.out
 /terraform-provider-llmgw/terraform-provider-llmgw
 /terraform-provider-llmgw/examples/.terraform*
 /terraform-provider-llmgw/examples/terraform.tfstate*
+# What a development server run from the checkout writes beside it.
+/secrets/

Makefile

@@ -6,6 +6,7 @@ LDFLAGS := -X '$(MODULE)/internal/version.Version=$(VERSION)'
 
 build:
 	go build -ldflags "$(LDFLAGS)" -o bin/llmgw ./cmd/llmgw
+	go build -ldflags "$(LDFLAGS)" -o bin/llmgw-kb ./cmd/llmgw-kb
 
 test:
 	go test ./...

README.md

@@ -3,8 +3,9 @@
 A single-binary LLM gateway for clusters with multiple InfiniBand fabrics,
 NVLink domains and heterogeneous GPUs.
 
-It serves LLMs running on your own GPU servers (vLLM / SGLang) behind one
-OpenAI-compatible endpoint. Design document: [llmgw-plan.md](llmgw-plan.md).
+It serves LLMs running on your own GPU and CPU servers (vLLM, SGLang,
+llama.cpp, TensorRT-LLM) behind one OpenAI- and Anthropic-compatible
+endpoint, and answers from your own documents. Design document: [llmgw-plan.md](llmgw-plan.md).
 
 ## Getting started
 
@@ -29,7 +30,7 @@ Releases track the roadmap phases; see [CHANGELOG.md](CHANGELOG.md).
 | v0.1.0 | 0 | Skeleton and sim-engine | released |
 | v0.2.0 | 1 | Static router (filter/score/pick, streaming proxy) | released |
 | v0.3.0 | 1.5 | Admission control (priority bands) | released |
-| v0.4.0 | 2 | Agent: GPU/IB discovery, engine supervision | released |
+| v0.4.0 | 2 | Agent: GPU/IB discovery, engine supervision | released as part of v0.5.0 |
 | v0.5.0 | 3 | Control plane: catalog, placement, reconciler, ctl | released |
 | v0.6.0 | 4 | Multi-node placement and fabric awareness | released |
 | v0.7.0 | — | Conversations, agent memory, disk management | released |
@@ -37,45 +38,119 @@ Releases track the roadmap phases; see [CHANGELOG.md](CHANGELOG.md).
 | v0.9.0 | — | Managing people and the tenancy tree | released |
 | v0.10.0 | — | InfiniBand fabrics and NVLink, doctor, node add, engines installed by the nodes | released |
 | v0.11.0 | — | High availability from setup, the setup and console screens, usage sums and per-tenant alerts | released |
-| v1.0.0 | 5 | Tenancy, prefix affinity, benchmarking | planned |
+| v0.12.0 | — | The cluster's settings in the store, in force when changed; the firewall switch; wipe | released |
+| v0.13.0 | — | The whole system stopped and started from one command; Qwen3Guard judges, on a GPU and CPUs together | released |
+| v0.14.0 | — | Knowledge bases: documents from people's own machines, kept apart per tenant, answered from | released (v0.14.1: adapters, the tree over the API, the firewall on every machine; v0.14.2: pictures from ComfyUI; v0.14.3: 3D models as jobs; v0.14.4: any ComfyUI workflow as a job, presets, licences; v0.14.5: picture and 3D models that add themselves; v0.14.6: setup that explains every part and lets it be chosen, the gateway on all three HA machines; v0.14.7: the judge sized for its machine; v0.14.8: sign-in with Active Directory, several keys for one outside provider; v0.14.9: sign-in set up from the console, a node's console that shows the machine) |
+| v0.15.0 | — | The company's MCP tools and VS Code through the gateway; guardrails that ask outside services and put redacted values back; answers to similar questions | released (v0.15.1: Strata, a 125B model of experts on one card; a second mirror kept in step; pickled weights refused; a served model's context changed; v0.15.2: a served model moved to more cards; v0.15.3: a model stopped and kept, or removed with its files) |
+| v0.16.0 | — | Weights pinned to a commit and checked against what Hugging Face published; documents embedded by the gateway; Strata offered where models are chosen; cluster upgrades that say how far they are | released (v0.16.1: the mirror's Strata tab, Strata's ready-made packs; v0.16.2: standbys that answer clients, FreeToken, a model fetched once per node; v0.16.3: replicas per profile, a served model's settings on the mirror page, the CPU's speed before serving) |
+| v1.0.0 | 5 | What phase 5 has left: precise KV routing proven on vLLM, the block cost picker, defragmentation across NVLink domains, prefill/decode disaggregation | planned |
 
 As of v0.6.0 a cluster runs unattended: agents report their hardware, the
 control plane places replicas — including ones larger than a single machine —
 models load when someone asks for them and stop when nobody does, and keys,
 budgets and rate limits decide who may use any of it.
 
-Since v0.11.0, not yet released (see [CHANGELOG.md](CHANGELOG.md),
-"Unreleased"):
+Since v0.11.0 (each release's detail is in [CHANGELOG.md](CHANGELOG.md)):
 
-- **Security hardening** — signed releases only, a token on the admin API,
-  engines not run as root, TLS for the agent port, etcd and Patroni, a
-  firewall for llmgw's internal ports, encrypted backups, bounded
-  connections, sign-in throttling, a Content-Security-Policy.
-- **Guardrails** — policies that check requests and answers (PII, secrets,
-  prompt injection, deny patterns, a judge model such as Llama Guard), in
-  observe or enforce mode, attached to any part of the tenancy tree:
+- **Setup that explains, and high availability that is whole** (v0.14.6).
+  `sudo llmgw` lists the parts of an install, each with what it is for and
+  what is missing without it, and every one can be turned off or set: the
+  starter model and its context, the mirror's directory, the judge's
+  categories and what it does when it does not answer, limits, the
+  portal's language, pictures and 3D. With three machines each runs the
+  control plane, with a floating address on the leader, and any one can
+  be lost: on three VMs, cutting the store primary's power lost no
+  request. One ComfyUI serves an environment's models, pictures, 3D and
+  workflows are MCP tools, and `/v1/messages` gives a model's reasoning as
+  thinking blocks.
+- **Picture and 3D models that add themselves** (v0.14.5). Where a node's
+  card fits flux-schnell or triposr and the catalog makes nothing of that
+  kind, the leader adds it on demand, so a new cluster's portal has its
+  Images and 3D pages; the cluster page and the console show each preset
+  and add it by hand, and `presets.auto` turns it off. Stable Fast 3D's
+  colours can be brought to the picture's (`albedo_match`), and every mesh
+  carries its whole licence.
+- **Any ComfyUI workflow, and models by name** (v0.14.4). `POST
+  /v1/workflows/runs` runs a ComfyUI model's workflow as a job, its inputs
+  (text, numbers, choices, images, masks, audio, video, files) declared in
+  the catalog and listed in `/v1/models`. `llmgw ctl model add triposr` or
+  `flux-schnell` adds a tested model with no YAML; a profile may name the
+  card it needs instead of a class, and ComfyUI now installs on V100s. A
+  licence carries what a machine can match, and a person can leave
+  licences out on the profile page so no key of theirs reaches such a
+  model. A 3D job waiting for the one card says what holds it and for
+  about how long.
+- **3D models as jobs** (v0.14.3). `POST /v1/3d/generations` answers with a
+  job, asked after until its mesh is ready to download. ComfyUI makes it on
+  the cluster's own cards, TripoSR being the tested model. The job says how
+  far along it is, the model's licence is written into the GLB, and
+  `per_mesh` prices it. People make them at `/3d` on the portal; `llmgw ctl
+  3d` shows every owner's jobs. A model profile may also load weights of
+  its own, and an admission band may wait longer than the rest, for bulk
+  work.
+- **Pictures from ComfyUI** (v0.14.2). `POST /v1/images/generations` is made
+  by ComfyUI, behind a gate that demands the replica's key, as an engine the
+  agents install from its source checked against a digest. `per_image`
+  prices it, and people make pictures at `/images` on the portal.
+- **Adapters, and the tenancy tree over the API** (v0.14.1). `llmgw ctl
+  kb dataset` and `llmgw ctl train` make a LoRA adapter from a collection
+  on a node's spare GPU memory, kept on the mirror and seen only by its
+  owner. Organizations, teams, projects, keys and memberships are in the
+  admin API and the Terraform provider. The agents' settings are edited
+  with the cluster's, on the portal too, and the firewall setting reaches
+  every machine.
+- **Knowledge bases** (v0.14.0). People put a company's documents in from
+  their own machines with `llmgw-kb`, a separate download for Windows,
+  macOS and Linux that reads PDF, Office, OpenDocument, mail and many more
+  formats and sends text and vectors, never the files. Collections are kept
+  apart per tenant, down to row-level security in PostgreSQL. A chat
+  request, or a part of the tree by default, is answered from them with
+  its sources: [docs/knowledge.md](docs/knowledge.md).
+- **One command for the whole system** (v0.13.0): `llmgw stop`, `start` and
+  `restart`. Qwen3Guard is the guardrails' judge, on a GPU and on CPUs
+  together.
+- **Settings that apply when changed** (v0.12.0). The cluster's
+  configuration lives in the store, the same on every control plane and in
+  force within seconds (`llmgw ctl config`, the console's Settings). The
+  firewall can be switched on and off for the whole cluster.
+- **Security hardening** (v0.11.2): signed releases only, a token on the
+  admin API, engines not run as root, TLS for the agent port, etcd and
+  Patroni, a firewall for llmgw's internal ports, encrypted backups,
+  bounded connections, sign-in throttling, a Content-Security-Policy.
+- **Guardrails** (v0.11.2). Policies check requests and answers for PII,
+  secrets, prompt injection, deny patterns, and with a judge model; they
+  observe or enforce, and attach to any part of the tenancy tree:
   [docs/guardrails.md](docs/guardrails.md).
-- **Administration** — an audit log of every change (`ctl audit`), staff
-  roles (viewer, operator, billing, admin), single sign-on through OpenID
-  Connect and provisioning through SCIM 2.0, key lifetime limits and
-  rotation with overlap (`keys ttl`, `keys rotate`), usage export as CSV
-  (`ctl usage export`).
-- **Notifications** — webhooks, signed or Slack-shaped, for a model parked,
-  a node lost, a budget spent and new weights published (`notify.webhooks`,
-  `ctl notify test`), and Alertmanager receiver templates.
-- **Traffic and models** — request tags, canaries of new weights with
-  automatic rollback, quality suites gating updates, a response cache,
-  kept and replayed requests, batches, Anthropic's Messages API, more of
-  OpenAI's (structured output, images, rerank, audio, moderations),
-  scaling by the clock, SLO-aware admission, TensorRT-LLM, MIG, models
-  imported from Ollama, routing by what each vLLM replica's cache holds,
-  LoRA adapters loaded on demand, other llmgw clusters as providers, and
-  placement that knows racks, PDUs and power caps.
-- **Operations** — `server restore` and off-machine backup copies, secrets
-  in Vault, forgetting a person and keeping a project's requests in the
-  cluster, offline installs from a bundle, a Helm chart, tracing from the
-  control plane to the engines, the admin API in OpenAPI with a Terraform
-  provider, and the portal in English and Turkish with its docs at /docs.
+- **Administration** (v0.11.2):
+  - an audit log of every change (`ctl audit`);
+  - staff roles (viewer, operator, billing, admin);
+  - single sign-on through OpenID Connect and provisioning through SCIM
+    2.0;
+  - key lifetime limits, and rotation with overlap (`keys ttl`, `keys
+    rotate`);
+  - usage export as CSV (`ctl usage export`).
+- **Notifications** (v0.11.2): webhooks, signed or Slack-shaped, for a
+  model parked, a node lost, a budget spent and new weights published, and
+  Alertmanager receiver templates.
+- **Traffic and models** (v0.11.2):
+  - request tags, canaries of new weights with automatic rollback, and
+    quality suites gating updates;
+  - a response cache, kept and replayed requests, and batches;
+  - Anthropic's Messages API, and more of OpenAI's (structured output,
+    images, rerank, audio, moderations);
+  - scaling by the clock, and SLO-aware admission;
+  - TensorRT-LLM, MIG, and models imported from Ollama;
+  - routing by what each vLLM replica's cache holds, and LoRA adapters
+    loaded on demand;
+  - other llmgw clusters as providers, and placement that knows racks,
+    PDUs and power caps.
+- **Operations** (v0.11.2):
+  - `server restore`, off-machine backup copies, and secrets in Vault;
+  - forgetting a person, and keeping a project's requests in the cluster;
+  - offline installs from a bundle, and a Helm chart;
+  - tracing from the control plane to the engines;
+  - the admin API in OpenAPI, with a Terraform provider;
+  - the portal in English and Turkish, with its docs at /docs.
 
 Every one of these is in the console as well as on the command line:
 [docs/console.md](docs/console.md).
@@ -110,6 +185,8 @@ Every page is also in the binary: `llmgw docs`, or **Manual** in the console.
   cluster that exists
 - [docs/kubernetes.md](docs/kubernetes.md) — the control plane on
   Kubernetes with the Helm chart, and why GPU nodes stay outside it
+- [docs/kong.md](docs/kong.md) — llmgw behind Kong Gateway, or Kong behind
+  llmgw: the configuration, what Kong counts, and which to choose
 - [docs/fabrics.md](docs/fabrics.md) — InfiniBand fabrics and NVLink: what
   llmgw needs of them, checking a node with `llmgw doctor`, and what to look at
   when a replica spanning nodes will not run
@@ -124,6 +201,9 @@ Every page is also in the binary: `llmgw docs`, or **Manual** in the console.
   official OpenAI and Anthropic SDKs, and llmgw's extensions
 - [docs/guardrails.md](docs/guardrails.md) — checking requests and answers:
   policies, detectors, a judge model, and where each applies
+- [docs/knowledge.md](docs/knowledge.md) — knowledge bases: collections,
+  who may read them, `llmgw-kb` on people's machines, and answers from
+  documents
 - [docs/choosing-models.md](docs/choosing-models.md) — which models to run,
   how many of each, and where the weights should live
 
@@ -152,7 +232,7 @@ Every page is also in the binary: `llmgw docs`, or **Manual** in the console.
 ## Building
 
 ```sh
-make build      # bin/llmgw
+make build      # bin/llmgw and bin/llmgw-kb
 make test       # unit tests
 make race       # with the race detector
 ```
@@ -169,6 +249,7 @@ llmgw ctl cordon gpu-a-01            # stop new placement, leave what runs
 sudo llmgw doctor                    # on a node: what its GPUs, NVLink and fabric need
 llmgw ctl audit --since 24h          # who changed what
 llmgw ctl usage export --from 2026-09-01 --to 2026-10-01 --by project,model
+llmgw ctl kb list                    # knowledge-base collections, grants, defaults
 ```
 
 Bare `llmgw` opens the console, which does all of this from one screen.
@@ -180,6 +261,33 @@ Add `--json` to any command for machine-readable output.
 says nothing about which of six constraints bit, so it prints every candidate
 considered and the reason it was rejected.
 
+## What llmgw adds to a request
+
+Measured with `llmgw ctl bench overhead`. It starts an engine that answers
+at once and a gateway in front of it, in one process on loopback. It sends
+the same load straight to the engine and through the gateway, and reports
+what the gateway added. Run on an Apple M5 (10 CPUs, load about 2) on
+2026-10-16, 10 s a step, 72,000 requests with none failed:
+
+| Through the gateway | added at p50 | added at p99 |
+|---|---|---|
+| bare | 0.06–0.24 ms | 0.10–0.33 ms |
+| with authentication and usage accounting | 0.05–0.32 ms | 0.09–0.37 ms |
+| with the guardrails' pattern checks as well | 0.10–0.42 ms | 0.15–0.49 ms |
+
+The range is across 100 to 2,000 requests a second. The overhead was
+*lowest* at 2,000, most likely because a busier processor keeps its cores
+awake. The guardrails' judge model is not in these figures: a judgment
+takes seconds on CPUs (see [install.md](docs/install.md)) and is a model's
+time, not the gateway's.
+
+The command also prints the gateway's CPU per thousand requests. Gateway,
+engine and load share one process there, so that figure is rough (0.04 to
+0.65 s per thousand in the run above); the latency is the one to quote.
+Figures from a VM on a busy host move with the host: run it where you will
+deploy, and keep the load it prints beside the result. CI runs a one-second
+version on every change (`TestTheGatewaysOverheadIsMeasured`).
+
 ## Running the router
 
 ```sh
@@ -207,8 +315,9 @@ within 1.24x of idle while 500 concurrent batch requests run, against 58x with
 admission off.
 
 Routing profiles select how endpoints are scored: `chat` favours session
-affinity, `batch` favours the shortest queue, `long-context` favours free KV
-cache. Pick one per model in the config, or per request with the
+affinity and the replica that has been answering soonest (its time to first
+token, smoothed), `batch` favours the shortest queue, `long-context` favours
+free KV cache. Pick one per model in the config, or per request with the
 `x-llmgw-profile` header.
 
 ## sim-engine

api/proto/llmgw/v1/agent.proto

@@ -28,6 +28,8 @@ message AgentMessage {
     FabricTestReply fabric_test = 6;
     UpgradeReply upgrade = 7;
     KVEvents kv_events = 8;
+    TrainStatus train_status = 9;
+    TrainArtifact train_artifact = 10;
   }
 }
 
@@ -73,6 +75,9 @@ message ControlMessage {
     ControlPlanes control_planes = 11;
     Firewall firewall = 13;
     RestartAgent restart_agent = 14;
+    AgentSettings agent_settings = 15;
+    TrainStart train_start = 16;
+    TrainCancel train_cancel = 17;
   }
   // trace is the W3C trace context (traceparent, and tracestate when there
   // is one) of the control plane's span that sent this command, so what the
@@ -103,6 +108,17 @@ message Firewall {
   int32 engine_ports_to = 4;
 }
 
+// AgentSettings is what the cluster's settings say of this node's agent
+// (agents.default, with agents.nodes.<id> over it), sent when it registers
+// and whenever they change. The agent keeps it beside agent.yaml, which
+// holds only what it needs to start; over that file, it is what the agent
+// runs on. A change from what it runs on has the agent start again on it,
+// its engines serving throughout.
+message AgentSettings {
+  // yaml is agent.yaml's keys, as YAML; empty for none.
+  string yaml = 1;
+}
+
 // ControlPlanes is every control plane an agent may dial, sent by the leader
 // when an agent registers and whenever the list changes. The agent keeps it
 // beside its configuration, so a second control plane added after the node
@@ -222,6 +238,13 @@ message Host {
   // nvidia_driver is the driver version, empty when there is none. A card
   // without a driver is not a GPU anything can use.
   string nvidia_driver = 6;
+  // mem_bandwidth_gbs is how fast the memory reads, in GB/s, measured when
+  // the agent started: what decides a model's speed on the CPU. 0 unknown.
+  double mem_bandwidth_gbs = 7;
+  // mem_bandwidth_core_gbs is what one core reads alone, in GB/s: a replica
+  // given a few threads of a large machine reads at a few times this, not at
+  // the whole machine's rate. 0 unknown.
+  double mem_bandwidth_core_gbs = 8;
 }
 
 // Storage is one filesystem a node keeps weights on.
@@ -318,6 +341,13 @@ message GPU {
   // memory the instance's share; utilization_percent stays zero, because no
   // driver measures it per instance.
   MIGDevice mig = 22;
+  // How the card is linked to its host: the fastest generation card and
+  // slot allow together, the card's own fastest, and the lanes in use and
+  // the card's most. Zero when the driver did not say.
+  int32 pcie_gen = 23;
+  int32 pcie_gpu_gen = 24;
+  int32 pcie_width = 25;
+  int32 pcie_width_max = 26;
 }
 
 // MIGDevice is which card a MIG instance is carved from, and how.
@@ -422,6 +452,21 @@ message GPUSensor {
   int32 last_xid = 26;
   string last_xid_text = 27;
   int64 last_xid_unix = 28;
+  // How much of the chip the work on it uses, which utilization_percent
+  // does not say: that is the share of time any kernel ran, and a kernel on
+  // one SM of a hundred counts in full. has_profile says these are present;
+  // -1 is not reported. Read from NVML's GPM on Hopper and later, or from a
+  // dcgm-exporter on the node; profile_source says which ("gpm", "dcgm").
+  bool has_profile = 29;
+  // sm_active_percent is the share of time the SMs had work, averaged over
+  // all of them; sm_occupancy_percent how full those SMs were with warps.
+  double sm_active_percent = 30;
+  double sm_occupancy_percent = 31;
+  // tensor_active_percent is the tensor cores; dram_active_percent the
+  // memory interface, which decode is bound by.
+  double tensor_active_percent = 32;
+  double dram_active_percent = 33;
+  string profile_source = 34;
 }
 
 // ShardLoad is one engine process group's use.
@@ -572,6 +617,13 @@ message Heartbeat {
   // is not handed out. Empty from an agent that does not measure it, or from
   // one that could not this time.
   repeated int32 gpu_foreign_memory_mb = 19;
+
+  // mem_bandwidth_gbs is how fast the node's memory reads, in GB/s, once
+  // the agent has measured it — in the background after it starts, so a
+  // registration usually has none yet. 0 means not measured.
+  double mem_bandwidth_gbs = 20;
+  // mem_bandwidth_core_gbs is one core's read rate, as in Host.
+  double mem_bandwidth_core_gbs = 21;
 }
 
 // Provision is the software a node should have: which engines, at which
@@ -674,6 +726,9 @@ message FetchProgress {
   // is a real answer: a percentage would then be invented, and an invented
   // percentage that stops moving is worse than a byte count that keeps rising.
   int64 bytes_total = 3;
+  // phase is what the bytes are: empty while they arrive, "checking" while the
+  // node checks a set's files against the mirror's checksums.
+  string phase = 4;
 }
 
 // StartShard launches one shard. It is idempotent by shard_id: the control
@@ -733,6 +788,13 @@ message EngineProfile {
   // kv_events has a vLLM engine publish its KV cache events to the agent, for
   // the router's precise index.
   bool kv_events = 13;
+  // allow_pickle lets the node fetch weights in a pickle format, which run
+  // code when loaded: the catalog's allow_pickle, set on purpose.
+  bool allow_pickle = 14;
+  // model_sha256 is the sha256 the registry published for the file
+  // model_path names, from the catalog's origin: the node refuses a copy
+  // that is not it. Empty when nothing is known to check against.
+  string model_sha256 = 15;
 }
 
 // RemoveWeights deletes cached model files a node no longer needs.
@@ -872,3 +934,55 @@ message LoraAdapter {
   string name = 1;
   string source = 2;
 }
+
+// TrainStart has a node train a LoRA adapter (BACKLOG F4): on a base model's
+// weights from its registry, on question/answer pairs, with a Python trainer
+// the agent carries, in an environment it makes over an installed engine's.
+// One at a time per node.
+message TrainStart {
+  string job_id = 1;
+  // base_repo and base_revision name the base model's weights on its
+  // registry (Hugging Face): safetensors, not the GGUF a catalog may serve.
+  string base_repo = 2;
+  string base_revision = 3;
+  // data is the training pairs, NDJSON of {"q", "a"}.
+  bytes data = 4;
+  // params is the trainer's settings as JSON: rank, alpha, epochs, lr,
+  // batch, max_len.
+  string params = 5;
+  // gpu_memory_mib is what the job needs of a card; the agent starts it
+  // only when one has that much free, which lets it run beside an engine.
+  int32 gpu_memory_mib = 6;
+  // gguf asks for the adapter in llama.cpp's format too.
+  bool gguf = 7;
+}
+
+// TrainCancel stops a training job.
+message TrainCancel {
+  string job_id = 1;
+}
+
+// TrainStatus is how a training job is going, sent as it changes.
+message TrainStatus {
+  string job_id = 1;
+  // state is preparing, running, done or failed.
+  string state = 2;
+  int32 step = 3;
+  int32 steps = 4;
+  double loss = 5;
+  // message is what is being done, or why it failed.
+  string message = 6;
+  // peak_mib is the most the job held of the card.
+  int32 peak_mib = 7;
+}
+
+// TrainArtifact is a piece of what a training job made, sent before its
+// final status: a gzipped tar of the PEFT adapter (kind peft), or the GGUF
+// adapter (kind gguf).
+message TrainArtifact {
+  string job_id = 1;
+  string kind = 2;
+  int32 seq = 3;
+  bytes data = 4;
+  bool last = 5;
+}

cmd/llmgw-kb/main.go

@@ -0,0 +1,469 @@
+// Command llmgw-kb puts a person's documents into an llmgw knowledge base
+// from their own machine (BACKLOG F2): it reads the files there, cuts them
+// into chunks, has the gateway embed them with the person's key, and sends
+// the text and vectors — only for what changed since the last run, which a
+// ledger on the machine remembers.
+//
+// Its own binary, for Windows, macOS and Linux: it carries none of the
+// server, agent or setup code that llmgw does.
+package main
+
+import (
+	"bufio"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"os/signal"
+	"path/filepath"
+	"strings"
+	"text/tabwriter"
+	"time"
+
+	"github.com/spf13/cobra"
+	"golang.org/x/term"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/kb/extract"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/kb/kbsync"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/version"
+)
+
+func main() {
+	if err := root().Execute(); err != nil {
+		fmt.Fprintln(os.Stderr, "error:", err)
+		os.Exit(1)
+	}
+}
+
+// config is what `login` keeps: the gateway, the key, and for a personal key
+// the project it acts in.
+type config struct {
+	Gateway string `json:"gateway"`
+	Key     string `json:"key"`
+	Project string `json:"project,omitempty"`
+	CA      string `json:"ca,omitempty"`
+}
+
+func configPath() (string, error) {
+	dir, err := os.UserConfigDir()
+	if err != nil {
+		return "", err
+	}
+	return filepath.Join(dir, "llmgw-kb", "config.json"), nil
+}
+
+// loadConfig reads the saved configuration, with the environment over it:
+// LLMGW_URL, LLMGW_API_KEY, LLMGW_PROJECT.
+func loadConfig() (config, error) {
+	var c config
+	if p, err := configPath(); err == nil {
+		if b, err := os.ReadFile(p); err == nil {
+			if err := json.Unmarshal(b, &c); err != nil {
+				return c, fmt.Errorf("%s: %w", p, err)
+			}
+		}
+	}
+	for env, dst := range map[string]*string{"LLMGW_URL": &c.Gateway, "LLMGW_API_KEY": &c.Key, "LLMGW_PROJECT": &c.Project} {
+		if v := strings.TrimSpace(os.Getenv(env)); v != "" {
+			*dst = v
+		}
+	}
+	return c, nil
+}
+
+func client(flags *globalFlags) (*kbsync.Client, error) {
+	c, err := loadConfig()
+	if err != nil {
+		return nil, err
+	}
+	if flags.gateway != "" {
+		c.Gateway = flags.gateway
+	}
+	if flags.project != "" {
+		c.Project = flags.project
+	}
+	if flags.ca != "" {
+		c.CA = flags.ca
+	}
+	if c.Gateway == "" || c.Key == "" {
+		return nil, errors.New("no gateway or key: run `llmgw-kb login`, or set LLMGW_URL and LLMGW_API_KEY")
+	}
+	h, err := kbsync.NewHTTP(c.CA)
+	if err != nil {
+		return nil, err
+	}
+	return &kbsync.Client{Base: c.Gateway, Key: c.Key, Project: c.Project, HTTP: h}, nil
+}
+
+type globalFlags struct {
+	gateway, project, ca, ledger string
+}
+
+func (g *globalFlags) openLedger(ctx context.Context) (*kbsync.Ledger, error) {
+	p := g.ledger
+	if p == "" {
+		var err error
+		if p, err = kbsync.DefaultLedgerPath(); err != nil {
+			return nil, err
+		}
+	}
+	return kbsync.OpenLedger(ctx, p)
+}
+
+func root() *cobra.Command {
+	g := &globalFlags{}
+	cmd := &cobra.Command{
+		Use:   "llmgw-kb",
+		Short: "Put your documents into an llmgw knowledge base, from this machine",
+		Long: `llmgw-kb reads documents on this machine, cuts them into chunks, has the
+gateway embed them with your key, and sends the text and vectors to a
+knowledge base. Files are never uploaded. Run it as often as you like: a
+ledger on this machine remembers what was sent, and only what changed is
+read and sent again.`,
+		Version:       version.Version,
+		SilenceUsage:  true,
+		SilenceErrors: true,
+	}
+	cmd.PersistentFlags().StringVar(&g.gateway, "gateway", "", "the gateway's address (default: from login, or LLMGW_URL)")
+	cmd.PersistentFlags().StringVar(&g.project, "project", "", "the project a personal key acts in (default: from login, or LLMGW_PROJECT)")
+	cmd.PersistentFlags().StringVar(&g.ca, "ca", "", "a PEM file of certificates to trust besides the system's")
+	cmd.PersistentFlags().StringVar(&g.ledger, "ledger", "", "the ledger file (default: in your configuration directory)")
+	cmd.AddCommand(loginCmd(g), collectionsCmd(g), syncCmd(g), statusCmd(g), historyCmd(g), formatsCmd())
+	return cmd
+}
+
+func loginCmd(g *globalFlags) *cobra.Command {
+	return &cobra.Command{
+		Use:   "login",
+		Short: "Save the gateway's address and your key",
+		Long: `Asks for the gateway's address and your API key, checks them, and saves them
+in your configuration directory, readable by you only. The key is typed
+without being shown, so it does not end up in your shell's history.`,
+		Args: cobra.NoArgs,
+		RunE: func(cmd *cobra.Command, _ []string) error {
+			c, _ := loadConfig()
+			in := bufio.NewReader(os.Stdin)
+			ask := func(q, def string) string {
+				if def != "" {
+					fmt.Fprintf(cmd.OutOrStdout(), "%s [%s]: ", q, def)
+				} else {
+					fmt.Fprintf(cmd.OutOrStdout(), "%s: ", q)
+				}
+				s, _ := in.ReadString('\n')
+				if s = strings.TrimSpace(s); s == "" {
+					return def
+				}
+				return s
+			}
+			c.Gateway = strings.TrimRight(ask("Gateway address (https://…)", or(g.gateway, c.Gateway)), "/")
+			fmt.Fprint(cmd.OutOrStdout(), "API key: ")
+			var key string
+			if term.IsTerminal(int(os.Stdin.Fd())) {
+				b, err := term.ReadPassword(int(os.Stdin.Fd()))
+				fmt.Fprintln(cmd.OutOrStdout())
+				if err != nil {
+					return err
+				}
+				key = strings.TrimSpace(string(b))
+			} else {
+				key, _ = in.ReadString('\n')
+				key = strings.TrimSpace(key)
+			}
+			if key != "" {
+				c.Key = key
+			}
+			c.Project = ask("Project, for a personal key (empty for a project's key)", or(g.project, c.Project))
+			if g.ca != "" {
+				c.CA = g.ca
+			}
+			h, err := kbsync.NewHTTP(c.CA)
+			if err != nil {
+				return err
+			}
+			cl := &kbsync.Client{Base: c.Gateway, Key: c.Key, Project: c.Project, HTTP: h, Wait: time.Second}
+			ctx, cancel := context.WithTimeout(cmd.Context(), 30*time.Second)
+			defer cancel()
+			cs, err := cl.Collections(ctx)
+			if err != nil {
+				return fmt.Errorf("the gateway refused: %w", err)
+			}
+			p, err := configPath()
+			if err != nil {
+				return err
+			}
+			if err := os.MkdirAll(filepath.Dir(p), 0o700); err != nil {
+				return err
+			}
+			b, _ := json.MarshalIndent(c, "", "  ")
+			if err := os.WriteFile(p, b, 0o600); err != nil {
+				return err
+			}
+			fmt.Fprintf(cmd.OutOrStdout(), "Saved in %s. This key reaches %d collection(s); `llmgw-kb collections` lists them.\n", p, len(cs))
+			return nil
+		},
+	}
+}
+
+func or(a, b string) string {
+	if a != "" {
+		return a
+	}
+	return b
+}
+
+func collectionsCmd(g *globalFlags) *cobra.Command {
+	return &cobra.Command{
+		Use:   "collections",
+		Short: "List the collections your key may read, and which it may add to",
+		Args:  cobra.NoArgs,
+		RunE: func(cmd *cobra.Command, _ []string) error {
+			c, err := client(g)
+			if err != nil {
+				return err
+			}
+			cs, err := c.Collections(cmd.Context())
+			if err != nil {
+				return err
+			}
+			if len(cs) == 0 {
+				fmt.Fprintln(cmd.OutOrStdout(), "No collections. An administrator creates them and grants access (llmgw ctl kb).")
+				return nil
+			}
+			tw := tabwriter.NewWriter(cmd.OutOrStdout(), 2, 4, 2, ' ', 0)
+			fmt.Fprintln(tw, "NAME\tACCESS\tOWNER\tEMBEDDED BY")
+			for _, x := range cs {
+				access := "read"
+				if x.Writable {
+					access = "read, write"
+				}
+				fmt.Fprintf(tw, "%s\t%s\t%s\t%s (%d)\n", x.Name, access, or(x.OwnerName, x.Owner), x.EmbedModel, x.Dims)
+			}
+			return tw.Flush()
+		},
+	}
+}
+
+func syncCmd(g *globalFlags) *cobra.Command {
+	var o kbsync.Options
+	var verbose bool
+	var ocr string
+	var maxMiB int64
+	cmd := &cobra.Command{
+		Use:   "sync COLLECTION PATH...",
+		Short: "Send a collection what changed in these folders and files",
+		Long: `Reads the files under each PATH that the collection does not already have as
+they are, and sends their text and vectors. Run it again whenever you like;
+unchanged files are not even read.
+
+Documents are named by their path under PATH, with forward slashes; with
+several PATHs, each folder's own name comes first.
+
+  --dry-run   reads and cuts everything, sends nothing, and says what would
+              leave this machine
+  --delete    removes documents this machine sent whose files are gone
+              (only under these PATHs; nothing anybody else sent)
+  --retry     reads again files skipped before (after installing LibreOffice
+              or Tesseract, say)
+
+Without --include, every file of a known format is read (llmgw-kb formats);
+hidden files, version control folders and Office lock files are not.`,
+		Args: cobra.MinimumNArgs(2),
+		RunE: func(cmd *cobra.Command, args []string) error {
+			c, err := client(g)
+			if err != nil {
+				return err
+			}
+			ctx, stop := signal.NotifyContext(cmd.Context(), os.Interrupt)
+			defer stop()
+			l, err := g.openLedger(ctx)
+			if err != nil {
+				return err
+			}
+			defer l.Close()
+			o.Collection, o.Roots = args[0], args[1:]
+			o.Extract = extract.Options{Helpers: extract.FindHelpers(), OCRLanguages: ocr, MaxBytes: maxMiB << 20}
+			out := cmd.OutOrStdout()
+			c.Waiting = func(why string, d time.Duration) {
+				fmt.Fprintf(out, "  ..       waiting %s: %s\n", d, why)
+			}
+			start := time.Now()
+			rep, err := kbsync.Sync(ctx, c, l, o, func(e kbsync.Event) {
+				switch e.Kind {
+				case "unchanged":
+					if !verbose {
+						return
+					}
+					fmt.Fprintf(out, "  =        %s\n", e.Name)
+				case "sent", "would-send":
+					fmt.Fprintf(out, "  %-8s %s (%d chunks, %s)\n", label(e.Kind), e.Name, e.Chunks, size(e.Bytes))
+				default:
+					line := fmt.Sprintf("  %-8s %s", label(e.Kind), e.Name)
+					if e.Detail != "" {
+						line += ": " + e.Detail
+					}
+					fmt.Fprintln(out, line)
+				}
+			})
+			if rep.Collection.Name != "" {
+				verb := "sent"
+				if o.DryRun {
+					verb = "would send"
+				}
+				fmt.Fprintf(out, "\n%s: %d files; %s %d (%d chunks, %s), %d unchanged, %d skipped, %d failed, %d deleted, in %s\n",
+					rep.Collection.Name, rep.Files, verb, rep.Sent, rep.Chunks, size(rep.Bytes), rep.Unchanged,
+					rep.Skipped, rep.Failed, rep.Deleted, time.Since(start).Round(time.Second))
+			}
+			if err != nil {
+				return err
+			}
+			if rep.Failed > 0 {
+				return fmt.Errorf("%d document(s) failed; run again to retry them", rep.Failed)
+			}
+			return nil
+		},
+	}
+	f := cmd.Flags()
+	f.StringSliceVar(&o.Include, "include", nil, "read only files matching these globs (\"*.pdf,docs/**/*.md\")")
+	f.StringSliceVar(&o.Exclude, "exclude", nil, "never read files matching these globs")
+	f.BoolVar(&o.Hidden, "hidden", false, "read hidden files and folders too")
+	f.BoolVar(&o.DryRun, "dry-run", false, "send nothing; say what would be sent")
+	f.BoolVar(&o.Delete, "delete", false, "remove documents this machine sent whose files are gone")
+	f.BoolVar(&o.Retry, "retry", false, "read again files skipped before")
+	f.IntVar(&o.Workers, "workers", 4, "files read and sent at once")
+	f.BoolVar(&o.EmbedHere, "embed-here", false,
+		"embed on this machine and send the vectors, even to a gateway that embeds text it is sent")
+	f.StringVar(&ocr, "ocr-languages", "tur+eng", "Tesseract's languages for images and scanned pages")
+	f.Int64Var(&maxMiB, "max-size", 256, "skip files larger than this many MiB")
+	f.BoolVarP(&verbose, "verbose", "v", false, "list unchanged files too")
+	return cmd
+}
+
+func label(kind string) string {
+	switch kind {
+	case "would-send":
+		return "send"
+	case "would-delete":
+		return "delete"
+	}
+	return kind
+}
+
+func size(b int64) string {
+	switch {
+	case b >= 1<<20:
+		return fmt.Sprintf("%.1f MiB", float64(b)/(1<<20))
+	case b >= 1<<10:
+		return fmt.Sprintf("%.0f KiB", float64(b)/(1<<10))
+	}
+	return fmt.Sprintf("%d B", b)
+}
+
+func statusCmd(g *globalFlags) *cobra.Command {
+	var all bool
+	return &cobra.Command{
+		Use:   "status COLLECTION",
+		Short: "Show what this machine has sent to a collection, and what it skipped",
+		Args:  cobra.ExactArgs(1),
+		RunE: func(cmd *cobra.Command, args []string) error {
+			c, err := client(g)
+			if err != nil {
+				return err
+			}
+			m, err := c.Manifest(cmd.Context(), args[0])
+			if err != nil {
+				return err
+			}
+			l, err := g.openLedger(cmd.Context())
+			if err != nil {
+				return err
+			}
+			defer l.Close()
+			es, err := l.Entries(cmd.Context(), strings.TrimRight(c.Base, "/"), m.Collection.ID)
+			if err != nil {
+				return err
+			}
+			counts := map[kbsync.State]int{}
+			tw := tabwriter.NewWriter(cmd.OutOrStdout(), 2, 4, 2, ' ', 0)
+			fmt.Fprintln(tw, "STATE\tDOCUMENT\tCHUNKS\tWHEN\tWHY")
+			for _, e := range es {
+				counts[e.State]++
+				if e.State == kbsync.Sent && !all {
+					continue
+				}
+				fmt.Fprintf(tw, "%s\t%s\t%d\t%s\t%s\n", e.State, e.Name, e.Chunks, e.At.Format("2006-01-02 15:04"), e.Reason)
+			}
+			_ = tw.Flush()
+			fmt.Fprintf(cmd.OutOrStdout(), "\n%s holds %d documents; from this machine %d sent, %d skipped, %d failed.",
+				m.Collection.Name, len(m.Documents), counts[kbsync.Sent], counts[kbsync.Skipped], counts[kbsync.Failed])
+			if !all {
+				fmt.Fprint(cmd.OutOrStdout(), " (--all lists the sent ones too)")
+			}
+			fmt.Fprintln(cmd.OutOrStdout())
+			return nil
+		},
+	}
+}
+
+func historyCmd(g *globalFlags) *cobra.Command {
+	return &cobra.Command{
+		Use:   "history",
+		Short: "List the latest syncs from this machine",
+		Args:  cobra.NoArgs,
+		RunE: func(cmd *cobra.Command, _ []string) error {
+			l, err := g.openLedger(cmd.Context())
+			if err != nil {
+				return err
+			}
+			defer l.Close()
+			runs, err := l.Runs(cmd.Context(), 20)
+			if err != nil {
+				return err
+			}
+			tw := tabwriter.NewWriter(cmd.OutOrStdout(), 2, 4, 2, ' ', 0)
+			fmt.Fprintln(tw, "STARTED\tTOOK\tCOLLECTION\tSENT\tUNCHANGED\tSKIPPED\tFAILED\tDELETED")
+			for _, r := range runs {
+				took := "unfinished"
+				if !r.Finished.IsZero() {
+					took = r.Finished.Sub(r.Started).Round(time.Second).String()
+				}
+				fmt.Fprintf(tw, "%s\t%s\t%s\t%d\t%d\t%d\t%d\t%d\n", r.Started.Format("2006-01-02 15:04"), took,
+					r.Collection, r.Sent, r.Unchanged, r.Skipped, r.Failed, r.Delete)
+			}
+			return tw.Flush()
+		},
+	}
+}
+
+func formatsCmd() *cobra.Command {
+	return &cobra.Command{
+		Use:   "formats",
+		Short: "List the formats read, and the helpers found on this machine",
+		Args:  cobra.NoArgs,
+		RunE: func(cmd *cobra.Command, _ []string) error {
+			builtIn, helper := extract.Supported()
+			out := cmd.OutOrStdout()
+			fmt.Fprintf(out, "Read here:\n%s\n\n", wrap(builtIn))
+			fmt.Fprintf(out, "Read with a helper program:\n%s\n\n", wrap(helper))
+			fmt.Fprintln(out, "Helpers:")
+			for _, l := range extract.FindHelpers().Describe() {
+				fmt.Fprintln(out, "  "+l)
+			}
+			return nil
+		},
+	}
+}
+
+func wrap(words []string) string {
+	var b strings.Builder
+	line := "  "
+	for _, w := range words {
+		if len(line)+len(w) > 78 {
+			b.WriteString(strings.TrimRight(line, " ") + "\n")
+			line = "  "
+		}
+		line += w + " "
+	}
+	b.WriteString(strings.TrimRight(line, " "))
+	return b.String()
+}

cmd/llmgw-kb/main_test.go

@@ -0,0 +1,24 @@
+package main
+
+import (
+	"os/exec"
+	"strings"
+	"testing"
+)
+
+// The client goes onto employees' laptops: none of the server, agent, store
+// or setup code may come with it, and it must build where llmgw does not.
+func TestTheClientCarriesNoServerCode(t *testing.T) {
+	out, err := exec.Command("go", "list", "-deps", ".").Output()
+	if err != nil {
+		t.Skipf("go list: %v", err)
+	}
+	for _, dep := range strings.Fields(string(out)) {
+		for _, banned := range []string{"/internal/store", "/internal/router", "/internal/agent", "/internal/control",
+			"/internal/cli", "/internal/guard", "github.com/jackc/pgx"} {
+			if strings.Contains(dep, banned) {
+				t.Errorf("llmgw-kb depends on %s", dep)
+			}
+		}
+	}
+}

deploy/ha/README.md

@@ -1,4 +1,4 @@
-# Two control planes
+# More than one control plane
 
 llmgw handles the hard half of this itself: only one control plane places work,
 and agents find whichever one is leading. What it does not do is move client
@@ -37,7 +37,7 @@ nothing in front of the two had moved.
 So something has to. The options, in the order most people should consider
 them:
 
-**A load balancer you already run.** Health-check `/ready` on each control
+**A load balancer you already run.** Health-check `/leader` on each control
 plane and send traffic to whichever answers. This is the right answer if you
 have one, and it needs nothing else from this directory.
 
@@ -50,25 +50,50 @@ in whatever your resolvers actually honour rather than what you set, which on
 a bad day is minutes. Fine for a cluster people use interactively, wrong for
 one an agent is driving.
 
-## Check `/ready`, not `/health`
+## Check `/leader`, not `/health`
 
 This is the one thing to get right, and the obvious choice is the wrong one.
 
-`/health` says the process is up. It says nothing about whether it can serve,
+`/health` says the process is up. It says nothing about whether it leads,
 and a standby is exactly that case: alive, answering, and holding no endpoints
 at all — its reconciler placed nothing and it has no agent streams to adopt
-from. A load balancer checking `/health` would send it half the traffic, and
-half the traffic would be a 503.
+from. A standby passes client requests on to the leader, so traffic sent to
+one still works — until the leader is the machine that was lost, and there
+is nobody to pass them to.
+
+`/leader` is 200 on the control plane that leads and 503 on a standby. It is
+what keepalived follows for the floating address, and what a load balancer
+should send traffic by: to the one machine that answers 200.
+
+`/ready` is narrower: 200 only when this process has somewhere to send a
+request, so 503 on a standby and also on a leader with no model running yet.
+Checking it takes the portal and the admin pages away with the models — on a
+new cluster with no node yet, nobody can sign in to add one. Use it to stop
+sending inference only, if your load balancer can tell the paths apart.
 
-`/ready` says this process has somewhere to send a request. It is 503 on a
-standby, 503 on a leader whose replicas have all failed, and 200 when there is
-something to serve — which is the question a load balancer is actually asking.
 Both are open, because a load balancer has no key.
 
-During a failover neither passes for about the length of the lease handover:
+During a failover nothing passes for about the length of the lease handover:
 three seconds when the leader is stopped cleanly, up to the lease TTL when it
-is not. That gap is real and there is no way to have it both ways — traffic
-cannot go to a control plane that has nothing to route to.
+is not. A proxy that waits that long for a healthy machine rather than
+failing at once (Caddy's `lb_try_duration`) hides it from clients.
+
+**Caddy**, in front of three control planes — the end of
+[deploy/proxy/Caddyfile](../proxy/Caddyfile) has it with the rest of the
+settings:
+
+```
+reverse_proxy 10.0.0.157:8080 10.0.0.158:8080 10.0.0.159:8080 {
+	lb_policy first
+	health_uri /leader
+	health_interval 2s
+	lb_try_duration 15s
+	flush_interval -1
+}
+```
+
+With a floating address the address already follows the leader: give Caddy
+that one address instead.
 
 Not the admin port, for a different reason: it serves pprof, and pprof can
 dump the process heap, where API keys sit in plaintext.

deploy/ha/keepalived.conf.example

@@ -41,7 +41,9 @@ vrrp_instance llmgw {
     # MASTER on the machine you would rather serve; BACKUP on the other.
     state MASTER
     interface eth0
-    # Any number both machines agree on and nothing else on the segment uses.
+    # Any number both machines agree on and nothing else on the segment uses:
+    # another cluster's keepalived with the same one reads as an invalid
+    # password. `llmgw server add --vip` derives one from the address.
     virtual_router_id 51
     # The higher one wins while both are healthy. The weight above is enough
     # to invert this when the preferred machine's gateway is not answering.

deploy/monitoring/grafana/llmgw-overview.json

@@ -504,7 +504,7 @@
     {
       "type": "timeseries",
       "title": "GPU busy, by card",
-      "description": "Share of each sample period a kernel was running, sampled every heartbeat.",
+      "description": "Share of each sample period in which any kernel ran (nvidia-smi's utilization.gpu), sampled every heartbeat. Not how much of the chip: one SM busy reads 100%. The SMs panel and power against its limit say how much.",
       "datasource": {
         "type": "prometheus",
         "uid": "llmgw-prom"
@@ -908,6 +908,51 @@
           "mode": "multi"
         }
       }
+    },
+    {
+      "type": "timeseries",
+      "title": "GPU SMs and tensor cores active, by card",
+      "description": "From the cards' profiling counters: NVML's GPM on Hopper and later, or a dcgm-exporter on the node. SMs active is the share of time the SMs had work, averaged over all of them; the gap to busy is chip a running kernel leaves idle. Tensor cores stay low in decode at batch one; prefill and batches are where they rise. Empty for cards nothing can profile.",
+      "datasource": {
+        "type": "prometheus",
+        "uid": "llmgw-prom"
+      },
+      "gridPos": {
+        "x": 0,
+        "y": 73,
+        "w": 24,
+        "h": 7
+      },
+      "id": 25,
+      "targets": [
+        {
+          "expr": "llmgw_gpu_sm_active_percent",
+          "legendFormat": "{{node}} gpu {{gpu}} SMs",
+          "refId": "A"
+        },
+        {
+          "expr": "llmgw_gpu_tensor_active_percent",
+          "legendFormat": "{{node}} gpu {{gpu}} tensor",
+          "refId": "B"
+        }
+      ],
+      "fieldConfig": {
+        "defaults": {
+          "unit": "percent",
+          "min": 0,
+          "max": 100
+        },
+        "overrides": []
+      },
+      "options": {
+        "legend": {
+          "displayMode": "list",
+          "placement": "bottom"
+        },
+        "tooltip": {
+          "mode": "multi"
+        }
+      }
     }
   ]
 }

deploy/monitoring/prometheus.yml

@@ -21,11 +21,18 @@ scrape_configs:
   # The control plane's dedicated metrics port. Not its admin listener: that one
   # is loopback because it also serves pprof, and pprof can dump the heap where
   # API keys sit in plaintext.
+  #
+  # Every control plane, one line each: the leader moves on restarts and the
+  # router's metrics are the leading one's, so scraping one alone leaves gaps
+  # — and alerts on them — whenever another leads.
   - job_name: llmgw
     static_configs:
       - targets: ['10.0.0.140:9091']
         labels:
-          instance: control-plane
+          instance: control-plane-140
+      # - targets: ['10.0.0.150:9091']
+      #   labels:
+      #     instance: control-plane-150
 
   # Every machine of the cluster, for its node exporter: network and disk
   # traffic, filesystems, CPU time. The control plane lists them — nodes,
@@ -33,10 +40,15 @@ scrape_configs:
   # nothing edited here, and each carries its llmgw id as the node label.
   # `llmgw node add` installs the exporter; on a machine added before that,
   # apt-get install prometheus-node-exporter.
+  # Asked of every control plane too: a standby lists only the control
+  # planes and the store, and they label every machine alike, so the same
+  # target from two of them is scraped once.
   - job_name: nodes
     http_sd_configs:
       - url: http://10.0.0.140:9091/prometheus/targets
         refresh_interval: 1m
+      # - url: http://10.0.0.150:9091/prometheus/targets
+      #   refresh_interval: 1m
 
   - job_name: prometheus
     static_configs:

deploy/proxy/Caddyfile

@@ -14,6 +14,7 @@ llm.example.com {
 		max_size 32MB
 	}
 
+	# One control plane. With more than one, see the end of this file.
 	reverse_proxy 127.0.0.1:8080 {
 		# Tokens are written the moment they are generated. Without this a
 		# streamed reply can arrive in bursts, and a client measuring time to
@@ -39,3 +40,36 @@ llm.example.com {
 	# Access log to stderr, which under systemd is the journal.
 	log
 }
+
+# More than one control plane (server init --standbys, server add): one of
+# them leads, the others stand by. Name them all and send to the one that
+# leads — /leader is 200 there and 503 on a standby. Not /ready: that is 503
+# on a leader with no model running yet, and would take the portal away with
+# it. A standby would pass a request on to the leader anyway, but not while
+# there is none, and the hop is for nothing.
+#
+# With a floating address (--vip) the address already follows the leader:
+# give reverse_proxy that one address instead, and keep the rest as above.
+#
+# llm.example.com {
+# 	request_body {
+# 		max_size 32MB
+# 	}
+# 	reverse_proxy 10.0.0.157:8080 10.0.0.158:8080 10.0.0.159:8080 {
+# 		# The first that is healthy: the leader, whichever it is.
+# 		lb_policy first
+# 		health_uri /leader
+# 		health_interval 2s
+# 		health_timeout 2s
+# 		# While leadership moves — a few seconds when a leader is lost —
+# 		# none is healthy; a request waits that long for one rather than
+# 		# failing at once.
+# 		lb_try_duration 15s
+# 		lb_try_interval 250ms
+# 		flush_interval -1
+# 		transport http {
+# 			response_header_timeout 15m
+# 		}
+# 	}
+# 	log
+# }

deploy/systemd/llmgw-agent.service

@@ -2,6 +2,9 @@
 Description=llmgw node agent
 After=network-online.target
 Wants=network-online.target
+# Never given up on: a new release that fails its first starts, or an upgrade
+# that restarts it twice in a minute, would otherwise leave the node off.
+StartLimitIntervalSec=0
 
 [Service]
 ExecStart=/usr/local/bin/llmgw agent --config /etc/llmgw/agent.yaml

docs.go

@@ -39,7 +39,7 @@ func (d Doc) Source() string {
 // than with either reader, because the console's manual and the gateway's
 // docs portal both list them, and two lists drift.
 var docOrder = []string{"console", "quick-install", "install", "node-onboarding", "onboarding",
-	"clients", "guardrails", "choosing-models", "fabrics", "facility-aware-placement", "data",
+	"clients", "guardrails", "knowledge", "choosing-models", "fabrics", "facility-aware-placement", "data",
 	"reference", "development"}
 
 // Manual is every document, in reading order: those docOrder names, then any
@@ -96,6 +96,12 @@ func titleOf(src, path string) string {
 //go:embed docs/admin-api.openapi.json
 var AdminOpenAPI []byte
 
+// ClientOpenAPI is the client API described in OpenAPI 3.1, generated from
+// internal/clientapi's route table and served at /v1/openapi.json.
+//
+//go:embed docs/client-api.openapi.json
+var ClientOpenAPI []byte
+
 // EvalSuites are the example quality suites under deploy/eval, which the
 // control plane runs by name when its eval directory has none of the name.
 //

docs/admin-api.md

@@ -8,8 +8,9 @@ description is.
 anything they do can be done with `curl`, a generated client or the
 [Terraform provider](../terraform-provider-llmgw/README.md). It adds and
 removes models, scales them, cordons and upgrades nodes, sets capacity
-quotas, attaches guardrail policies, and reads the fleet, usage and the audit
-log.
+quotas, attaches guardrail policies, adds organizations, teams and projects,
+issues and revokes API keys, puts people in the tree, and reads the fleet,
+usage and the audit log.
 
 ## Reaching it
 
@@ -99,6 +100,14 @@ curl -s "${H[@]}" -X POST $API/models/qwen-small/scale -d '{"min": 0, "max": 2,
 # A team's capacity quota.
 curl -s "${H[@]}" -X PUT $API/quotas -d '{"scope": "team:Acme/ml", "limit": {"gpus": 8}}'
 
+# An organization, a team in it and a project in that; then a key for the
+# project. The answer holds the key once: only its hash is kept.
+curl -s "${H[@]}" -X POST $API/orgs -d '{"name": "Acme"}'
+curl -s "${H[@]}" -X POST $API/teams -d '{"name": "ml", "in": "Acme"}'
+curl -s "${H[@]}" -X POST $API/projects -d '{"name": "search", "in": "Acme/ml"}'
+curl -s "${H[@]}" -X POST $API/keys -d '{"name": "ci", "project": "Acme/ml/search", "expires_in": "720h"}'
+curl -s "${H[@]}" -X POST $API/users/ayse@example.com/memberships -d '{"target": "team:Acme/ml"}'
+
 # Remove the model; a DELETE carries the content type too.
 curl -s "${H[@]}" -X DELETE $API/models/qwen-small
 ```
@@ -149,11 +158,11 @@ go generate ./internal/docgen
 
 ## What it does not cover
 
-- **Organizations, teams, projects, people and API keys.** These live in the
-  store and are managed with `llmgw org`, `llmgw users` and `llmgw keys` on the
-  control plane, or in the web console's `/admin` pages on the client port
-  behind a session. The admin API reads the tree only where a quota or a
-  policy names a part of it.
+- **Making people's accounts.** People sign up, are invited, or come from
+  the identity provider (OIDC, SCIM); the API lists them and puts them in
+  the tree, and `llmgw users` approves, disables and deletes them. Key
+  lifetime limits and spending limits are `llmgw keys ttl` and `llmgw keys
+  limits`.
 - **The client-facing API** (`/v1/chat/completions` and the rest), which is
   OpenAI's and Anthropic's, on `listen`, with API keys.
 - **The rest of the admin listener**: `/metrics`, `/health`, the dashboard

docs/clients.md

@@ -38,6 +38,14 @@ export ANTHROPIC_BASE_URL=https://llm.example.com
 export ANTHROPIC_API_KEY=$LLMGW_API_KEY
 ```
 
+### The API described
+
+Every endpoint is described in OpenAPI 3.1 at `/v1/openapi.json`. It needs
+no key, so a client generator can be pointed at it. `/docs/api` shows it in
+Swagger UI, where *Authorize* takes a key and *Try it out* sends a request to
+the gateway. People with a part in running the gateway also find the admin
+API's document there.
+
 ## A first request
 
 ### curl
@@ -526,6 +534,73 @@ curl https://llm.example.com/mcp \
   -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/list"}'
 ```
 
+**Pictures, 3D models, vision and workflows.** Where the gateway has the
+models for them, the same server offers more tools — only those it has a
+model for:
+
+| Tool | Does |
+|---|---|
+| `generate_image` | A picture from a prompt. `style` adds what a picture for a mesh wants: `concept` (one object, three-quarter view, grey background), `side` (an exact side view), `texture` (a seamless tile) or `none`. Answers the PNG's `file_id`, `url` and `seed`, with a 512 px preview to look at. |
+| `generate_mesh` | A GLB from a picture — `image` is a `file-…` id (a `generate_image` result, say), a data URL or base64 — or from a `prompt`, for which a concept picture is made first. `seed`, `resolution`, `texture_resolution`, and any other input the model's workflow declares in `inputs`. Answers the GLB's `file_id`, `url` and licence. |
+| `review_images` | Asks a vision model about one to four pictures (each shrunk to 1280 px on its long side); `json_schema` makes it answer that JSON. |
+| `run_workflow` | Runs a ComfyUI workflow from the inputs its catalog entry declares, and answers its result file. |
+| `job_status` | Where a mesh or a workflow run stands, and its file when done; `wait_seconds` waits for it. |
+| `gateway_status` | The picture, 3D, vision and workflow models, whether each is loaded on a GPU now, a workflow's inputs, and your last jobs. |
+
+Each tool makes its request of the gateway under your key, so your
+allowlist, limits, left-out licences and budget apply, and it is charged
+and listed like any other request. What it makes is kept as your file for
+as long as job results are: fetch it from `url` with the same key (`curl -H
+"Authorization: Bearer $LLMGW_API_KEY" -o rock.glb <url>`). A tool server
+over HTTP cannot write to your disk, so an agent that needs the file
+locally — to import it in Blender — fetches it, or goes through a small
+local bridge that does.
+
+A picture can be uploaded once and named by its id after: `POST /v1/files`
+with `purpose=vision` (or `user_data`) keeps it, for as long as job results
+are kept, as a file of yours that `generate_mesh`, `review_images` and a
+workflow's file inputs take as `file-…`. A call is at most 64 MB.
+
+A model that is not loaded is waited for: a picture or a look at one whose
+model is starting, or waiting for another to give up the GPU, waits as a
+mesh does, saying so, rather than failing with "model loading".
+
+A mesh takes from half a minute to a few minutes. A client that takes the
+answer as a stream (`Accept: text/event-stream`, as Claude Code does) is
+answered that way: `notifications/progress` for the call's
+`progressToken` while it runs — queued, waiting for the GPU, the node and
+its percent — then the result, and the call waits up to ten minutes. One
+answered in one piece waits 50 seconds; a job not done by then is answered
+with its `job_id`, to ask after with `job_status`. `wait_seconds` (up to
+900) sets either.
+
+```
+claude mcp add --transport http llmgw https://llm.example.com/mcp \
+  --header "Authorization: Bearer $LLMGW_API_KEY"
+```
+
+**Your company's tools.** Where an administrator has registered the
+company's own MCP servers (Jira, a file server, a database) and your key is
+granted their tools, the same `/mcp` lists them too, each as `server.tool`
+(`jira.search`). Give your agent only llmgw's `/mcp` and your key: the
+servers' own credentials stay with llmgw. A tool your key is not granted is
+not listed, and calling it by name is refused. Each call is in the audit
+log under your key: which tool, the size of what was sent, how it ended and
+how long it took, never what it said.
+
+### Licences you leave out
+
+A model's licence is in `/v1/models` (`license`: its `id`, `name`, whether
+it allows commercial use and the regions it grants nothing in). On your
+profile page, under **Licences**, you can leave licences out: one by one,
+every licence that forbids commercial use, or every licence that grants
+nothing in a region you name (`EU` for the union). The rules cover models
+added later too. From then on those models are not listed to your keys and
+a request for one is refused with 403 and the code `licence_excluded`,
+naming the choice that left it out — on every key you own and on the
+portal's pages, whatever a key's own allowlist says. A change reaches keys
+already in use within a minute.
+
 ### Headers llmgw sends back
 
 | Header | Says |
@@ -545,9 +620,10 @@ its usual exception with the gateway's message in it. The ones worth knowing:
 |---|---|
 | 400 | Besides a malformed request: a prompt too long for the model, or a guardrail policy refused it (the code `content_policy_violation`). |
 | 401 | No key, a mistyped one, or one that was revoked or has expired. |
-| 403 | The key may not use that model. |
+| 403 | The key may not use that model; with the code `licence_excluded`, its licence is one you left out on your profile page. |
 | 429 | A rate limit or budget of the key, its project or its team was reached. `Retry-After` says when to try again, and is absent when waiting will not help — a spent monthly budget. |
-| 503 | No model of that name is served (the message lists the ones that are), the model is still loading, or every replica is busy and the request waited as long as it may. The SDKs retry by themselves; a name that is not served will not start to be. |
+| 503 | No model of that name is served (the message lists the ones that are), the model is still loading, or every replica is busy and the queue is full. The SDKs retry by themselves; a name that is not served will not start to be. |
+| 504 | Every replica was busy for as long as the key's band may wait. `Retry-After` says when to send it again. |
 
 ## Priority
 

docs/console.md

@@ -2,8 +2,8 @@
 
 `sudo llmgw`, with nothing after it, on a machine llmgw runs on. On a machine
 with nothing installed it opens the setup screens; on a control plane, the
-console described here; on a node, the node's own console, which edits
-`agent.yaml`. Everything on these screens can also be done by command — each
+console described here; on a node, the node's own console, which shows
+what runs on that machine and edits `agent.yaml`. Everything on these screens can also be done by command — each
 screen says which — so a script never needs them and a person never needs
 to remember the commands.
 
@@ -164,6 +164,32 @@ which): requests, tokens, errors and spend.
 team or project is guaranteed and may at most hold, and what it holds now.
 **A** adds one; on a row, change or remove it.
 
+**Knowledge** lists the knowledge-base collections: each one's owner, the
+model that embeds it, how many documents and chunks it holds and who else it
+is granted to. **A** makes one. On a row: its **Documents**, **Grant…** a
+project, team, organization, person or key read or write access, **Revoke**
+a grant, **Answer from it by default…** for a part of the tree, **Try a
+search…** as an administrator, or **Delete…** it. People
+fill collections from their own machines with `llmgw-kb`
+([knowledge.md](knowledge.md)). Off on a SQLite store, and says why.
+
+**MCP servers** lists the company's own MCP servers that agents reach through
+the gateway's `/mcp`, their tools named `server.tool`: whether each is on,
+who has been granted its tools, its rate per caller, the names of the
+credentials it is sent (never their values) and its address. **A** adds one:
+its name, which is its tools' prefix, its address, and a bearer token or
+another header, kept sealed in the store. On a row: **Test, and list its
+tools**, with its credential, as a call would; **Grant its tools to…** a
+team, project, organization, person or key — all of them, `jira.*`, or some
+by name — beside whatever else that level is granted, and saying when a
+level above keeps them out of reach; **Revoke** a grant;
+**Change…** its address, timeout or rate; **Replace its credential…**;
+**Turn off** (its tools are listed nowhere and every call refused) or on;
+**Remove…**. A server's tools are seen nowhere until granted. The screen
+also says how a person points an agent at the gateway, with a key of their
+own from `/profile/keys`. `llmgw ctl mcp` does the same from a shell, and
+the web page **MCP servers** (`/admin/mcp`, administrators only) in a browser.
+
 **Mirror** is what the model mirror holds, one row for each version of each
 model — a file, or the set of files a model of several is: shards, a vision
 projector, a config and tokenizer — with how far its download has got and
@@ -237,6 +263,28 @@ long a key made there may live, and is the expiry of one made without any
 Limits are requests and tokens a minute, a budget per day or month, and the
 models allowed. Each level only narrows what is above it.
 
+**Sign-in** — how people get in besides a password kept here: single
+sign-on, Active Directory and provisioning, each on or off, with the
+directory's last sync (how many people, accounts kept in step, suspended,
+reinstated).
+- **E** sets up Entra ID. Give the tenant ID, the client ID and secret,
+  and the groups for administrators and operators. A provider other than
+  Entra takes its issuer URL in place of the tenant. The form shows the
+  redirect address to register.
+- **D** sets up Active Directory: the servers, where people are, the
+  service account and its password, the domain's CA certificate as a PEM
+  file on this machine, the group whose members may sign in, and groups
+  for roles and teams.
+- Before it is kept, the directory is tried from the control plane: each
+  server reached and bound, how many people may sign in, how many are in
+  each group. If any step fails, nothing is kept and the steps are shown.
+  **T** tries the settings in force.
+- On the provisioning row, Enter makes a SCIM token, shown once.
+- The password, the client secret and the token are kept in the store,
+  sealed by the master key when one is set. The settings name them as
+  `store:auth/…`, so each control plane has them with no file put on it.
+  The CA is written into the settings themselves.
+
 **Guardrails…** on a row lists the guardrail policies server.yaml defines.
 Each policy is marked in one of three ways:
 
@@ -326,5 +374,18 @@ to restart it.
 
 ## The node's console
 
-On a node, `sudo llmgw` opens `agent.yaml` in the same settings view, and the
-manual. The cluster is watched and changed from a control plane.
+On a node, `sudo llmgw` opens on **Overview**: what this machine is doing, read
+on the machine itself, so it works without an admin token and while the control
+plane is out of reach. Refreshed every three seconds, it shows
+
+- llmgw-agent's state and since when, and the version of llmgw;
+- each control plane the agent dials, from `agent.yaml` and the list the
+  leader last sent, and whether it can be reached;
+- each GPU: how busy, its memory, its heat, and the models on it;
+- the engines running here, from the agent's own records: replica, model,
+  engine, GPUs, port, whether the process is still there, and for how long;
+- the free space where weights and engines are kept;
+- the agent's warnings and errors of the last hour.
+
+Below it, `agent.yaml` in the same settings view, and the manual. The cluster
+is watched and changed from a control plane.

docs/data.md

@@ -72,7 +72,9 @@ each answer that may be reused, as the client got it, and a hash of the
 request it answers, in the gateway process's memory — never in the store,
 never on disk — for the project's TTL or until pushed out by newer ones
 (`router.cache.max_bytes`), and gone with a restart. Never under a `no_store`
-policy. Forgetting a person does not reach into it; the answers age out by
+policy. A project that reuses answers to similar questions (`--similar`)
+also has each question's last user message embedded, and the vector kept
+beside its answer, in the same memory and for as long. Forgetting a person does not reach into it; the answers age out by
 the TTL. The store keeps only the setting, and a usage record's mark that it
 was answered from the cache.
 

docs/guardrails.md

@@ -91,6 +91,21 @@ guard:
   value is the same placeholder throughout a request).
 - **`block`** stops it.
 
+**Putting them back.** With `restore: true` in a policy's `input`, the
+values a request had redacted are put back in its answer: the model reads
+"write to [PERSON_1] at [EMAIL_1]", the client reads the name and address
+it sent. Streamed answers too — a placeholder split across two events is
+held until it is whole. Only that request's own values, only into its
+answer; the map is kept in memory for the request and nowhere else. The
+output rules then read the answer as the client will, so a policy that
+redacts personal data in answers still does.
+
+```yaml
+input:
+  pii: {email: redact, phone: redact}
+  restore: true
+```
+
 **Shapes.**
 
 - `system_prompt` is put first.
@@ -258,6 +273,47 @@ themselves.
     context rather than overflowing it and going unjudged.
 - **Cache:** the same text is not asked about twice within five minutes.
 
+## An outside service
+
+A site required to use a particular guardrail — Azure Content Safety,
+Lakera Guard, NeMo Guardrails, a classifier of its own — names it in a
+policy's `services`. It is asked over HTTP beside the patterns and before
+the judge; its "unsafe" blocks or flags like theirs, and a service that does
+not answer is what the policy's `fail` says.
+
+```yaml
+policies:
+  strict:
+    fail: closed
+    services:
+      - name: lakera
+        url: https://api.lakera.ai/v2/guard
+        header_files: {Authorization: /etc/llmgw/secrets/lakera}   # "Bearer …", read each call
+        body: '{"messages": [{"role": "user", "content": {{json .Text}}}]}'
+        blocked: flagged                 # where the answer says unsafe
+        reason: breakdown.0.detector_type
+        on: both                         # input, output or both
+        action: block                    # or flag
+        timeout: 2s
+```
+
+`body` is a template given `.Text` (`{{json .Text}}` writes it as a JSON
+string; empty sends `{"input": …}`). `blocked` and `reason` are dotted paths
+into the JSON answer, numbers for list items; true, a number above zero, a
+non-empty list or a string other than "false" is unsafe. On input the
+service reads the latest user turn as the model will, with what `redact`
+took out still out; on output the whole answer, once, at its end, as the
+client reads it — with `restore`, the values put back. Being at the end, a
+service can stop a streamed answer only where it is held back
+(`stream: buffer_all`): released a sentence at a time, the client has read
+it by then, and what the service finds ends the stream with
+`content_filter`, flags it and is recorded. The rule recorded is
+`service:<name>:<reason>`. The service is told the text only, never who
+sent it. Under `fail: open`, the default, a service that cannot be asked
+lets the text through — and a caller can make it fail, sending what it
+refuses to read; give `fail: closed` where the service must decide. Of an
+answer asked for with `n` above 1, the service reads the first.
+
 ## Where a policy applies
 
 `guard.default` applies everywhere. Others are attached to a part of the

docs/knowledge.md

@@ -0,0 +1,428 @@
+# Knowledge bases
+
+A knowledge base is a collection of a company's documents that the cluster
+can answer from. People put their documents into it from their own machines
+with `llmgw-kb`. The tool reads the files on that machine, cuts them into
+passages ("chunks") and sends their text; the gateway embeds each passage
+with the person's own key and keeps the text and the vectors. The files
+themselves never leave the machine.
+
+## Why retrieval, and not training
+
+Retrieval is the tool for *knowledge*: policies, procedures, contracts. The
+answer can name its source, a changed document is sent again in minutes, and
+a passage only ever reaches people allowed to read its collection.
+
+Training a model on the documents (LoRA) changes how it *writes*, not what it
+reliably *knows*. It has no access control: whatever goes into the weights
+is available to everyone who may use them.
+
+## What it needs
+
+- **The PostgreSQL store with pgvector.** On a SQLite store knowledge bases
+  are off, and `llmgw ctl kb list` and the console say why.
+  - `llmgw server init` and the high-availability database setup install
+    pgvector and enable it.
+  - A cluster set up before this build needs it by hand, once:
+    1. Install the package on every database machine:
+       `sudo apt-get install postgresql-16-pgvector` (use the server's
+       version number).
+    2. On the leader, enable it in llmgw's database:
+       `sudo -u postgres psql -d llmgw -c 'CREATE EXTENSION vector'`.
+    3. Restart the control planes: `llmgw restart`.
+- **An embedding model in the catalog**, with `capabilities.embeddings`.
+  Two good choices handle Turkish and run on CPUs: `BAAI/bge-m3` (1024
+  dimensions) and `Qwen/Qwen3-Embedding-0.6B`.
+
+  Add it as any model is added (**Serve a model** in the console, or the
+  catalog), with its entry saying that it embeds:
+
+  ```yaml
+  capabilities:
+    embeddings: true
+  ```
+
+  A collection stays with the model it was made with: vectors from two
+  models cannot be compared.
+
+  Some embedding models are trained with an instruction before the
+  question and none before the documents. Give it as the entry's
+  `query_prefix`; llmgw puts it before questions only. For
+  Qwen3-Embedding:
+
+  ```yaml
+  query_prefix: "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: "
+  ```
+
+  With pgvector 0.7 or later a collection's index is kept at half
+  precision (`halfvec`), which halves its memory and takes models of up to
+  4000 dimensions; with an older pgvector the limit is 2000. A collection
+  keeps the kind it was made with.
+
+## For administrators
+
+### Make a collection
+
+A collection belongs to one part of the tenancy tree: an organization, a
+team, a project or a person. Whoever is in the owner, at any level below it,
+may read it and add to it.
+
+```sh
+llmgw ctl kb create team:Acme/hr policies bge-m3
+llmgw ctl kb create user:USER-ID notes bge-m3 --org Acme
+```
+
+The model's dimension is asked of the model itself, so the model has to be
+able to serve when the collection is made. `--dims 1024` gives it instead.
+
+In the console: **Knowledge**, then **A**.
+
+### Who may read and add
+
+The owner, and everything below it, may read and add. A **grant** gives the
+same to another part of the same organization, to a person, or to one key:
+
+```sh
+llmgw ctl kb grant Acme/policies project:Acme/sales/crm           # read
+llmgw ctl kb grant Acme/policies team:Acme/legal --write          # read and add
+llmgw ctl kb revoke Acme/policies project:Acme/sales/crm
+```
+
+How access works:
+
+- **Grants add up down the tree.** A grant to a team reaches its projects
+  and their keys.
+- **Nothing crosses organizations.** A grant to another organization's
+  team, project, key or person is refused.
+- **A person's own key** (a key with no project) reads nothing by itself.
+  The person names the project they act in (`llmgw-kb login` asks for it,
+  or `--project`) and must be a member of it. Joining a second team
+  therefore does not quietly widen what every key of theirs can read.
+- **No access looks like no collection.** A collection a key may not read
+  is answered exactly as one that does not exist, so even its name does not
+  leak.
+
+### Answer from collections by default
+
+Clients that cannot add a header still get answers from the documents when
+their part of the tree has a default:
+
+```sh
+llmgw ctl kb default team:Acme/hr Acme/policies Acme/forms
+llmgw ctl kb default key:KEY-ID Acme/policies
+llmgw ctl kb default team:Acme/hr --clear
+```
+
+- **Which default applies.** The narrowest one: a key's over its project's,
+  a project's over its team's, a team's over its organization's.
+- **A request can override it.** A request that names collections uses
+  those instead, and `X-Llmgw-Knowledge: none` turns the default off for
+  that request.
+- **A default never refuses a request.** A collection the caller may not
+  read is left out for them. If the search fails (the embedding model is
+  down, say), the request is answered without passages and the failure is
+  logged.
+- **Changes take effect within 30 seconds.**
+
+In the console: **Knowledge**, Enter on a collection, then **Answer from it
+by default…**.
+
+### Look after it
+
+```sh
+llmgw ctl kb list                          # collections, owners, documents, grants
+llmgw ctl kb docs Acme/policies            # its documents, who added each and when
+llmgw ctl kb rm Acme/policies old/form.pdf # one document out
+llmgw ctl kb search Acme/policies how many days of annual leave
+llmgw ctl kb delete Acme/policies          # the collection, its documents and grants
+```
+
+Rules that apply:
+
+- A project, team or organization that owns a collection is not deleted
+  until its collections are. This is the same rule as for one with members.
+- A person's own collections go with their account (`users delete`,
+  `users forget`).
+- Every document added or removed is in the audit log (`llmgw ctl audit`),
+  by name. The text is never recorded.
+
+### A large import
+
+A few thousand documents ask the embedding model for far more than it can
+answer at once. Give the key that sends them a band of its own, below the
+people's, that may wait for room for minutes rather than the 30 seconds the
+others wait:
+
+```sh
+llmgw ctl config set router.admission.bands '[{name: gold, priority: 300, saturation_ceiling: 0.85},
+  {name: bronze, priority: 100, saturation_ceiling: 0.3}, {name: bulk, priority: 50, saturation_ceiling: 0.3, queue_ttl: 5m}]'
+llmgw keys create import --project Acme/docs --band bulk
+```
+
+The people's questions then go first, and the import waits behind them. The
+gateway embeds a document's passages 32 at a time, four batches at once,
+enough to fill an embedding replica's slots; a batch told to wait (`Retry-After`)
+waits up to five minutes before the document is answered the same, and
+`llmgw-kb` sends it again rather than counting it as failed. It gives up
+after 15 minutes.
+
+## For everyone: `llmgw-kb`
+
+### Get it
+
+Download `llmgw-kb` for your system from the release page, next to `llmgw`:
+
+| System | File |
+|---|---|
+| Windows (Intel/AMD) | `llmgw-kb-VERSION-windows-amd64.exe` |
+| Windows (ARM) | `llmgw-kb-VERSION-windows-arm64.exe` |
+| macOS (Apple silicon) | `llmgw-kb-VERSION-darwin-arm64` |
+| macOS (Intel) | `llmgw-kb-VERSION-darwin-amd64` |
+| Linux | `llmgw-kb-VERSION-linux-amd64` / `-arm64` |
+
+Rename it to `llmgw-kb` (`llmgw-kb.exe` on Windows) and put it somewhere on
+your PATH. You do not need to install anything else.
+
+On macOS, a downloaded program that is not signed is stopped the first time.
+Either right-click it and choose Open, or run
+`xattr -d com.apple.quarantine llmgw-kb`.
+
+### Log in once
+
+```sh
+llmgw-kb login
+```
+
+It asks for three things, checks them, and saves them where only you can
+read them:
+
+- **The gateway's address.**
+- **Your API key.** It is typed without being shown.
+- **A project.** Needed only for a personal key.
+
+Instead of logging in you can set `LLMGW_URL`, `LLMGW_API_KEY` and
+`LLMGW_PROJECT`.
+
+```sh
+llmgw-kb collections        # what your key may read, and which you may add to
+```
+
+### Send your documents
+
+```sh
+llmgw-kb sync policies ~/Documents/HR
+llmgw-kb sync policies "C:\Users\ayse\Documents\İK" --include "*.pdf,*.docx"
+llmgw-kb sync policies ./handbook ./forms --exclude "drafts/**"
+```
+
+Run it as often as you like. The first run sends everything. Later runs
+send only what changed, and do not even read files whose size and time have
+not changed.
+
+| Option | What it does |
+|---|---|
+| `--dry-run` | Reads and cuts everything, sends nothing, and lists what would leave the machine and how much. |
+| `--delete` | Also removes documents whose files are gone. Only those this machine sent, from these folders; never anything someone else sent. |
+| `--retry` | Reads again files skipped before, for example after installing LibreOffice. |
+| `--include`, `--exclude` | Glob patterns. `*.pdf` matches a file name; `reports/**/*.xlsx` matches a path. |
+| `--hidden` | Reads hidden files and folders too (left out by default). |
+| `-v` | Lists unchanged files too. |
+| `--embed-here` | Embeds the passages from this machine and sends the vectors. With a gateway that does not embed them itself it does so anyway. |
+
+Documents are named by their path under the folder, with forward slashes,
+on every system. With several folders, each folder's own name comes first.
+
+```sh
+llmgw-kb status policies    # what this machine sent and skipped, and why
+llmgw-kb history            # the latest runs
+llmgw-kb formats            # what can be read here
+```
+
+### What it reads
+
+| Read here | Formats |
+|---|---|
+| Documents | PDF (text), Word (`.docx`), Excel (`.xlsx`), PowerPoint (`.pptx`, with speaker notes), OpenDocument (`.odt`, `.ods`, `.odp`), RTF, EPUB |
+| Web and mail | HTML, MHTML, e-mail (`.eml`, with its attachments) |
+| Text | Markdown, plain text, CSV, JSON, YAML, XML, subtitles, notebooks (`.ipynb`), source code |
+| Archives | ZIP, with every file inside it read in turn |
+
+| Read with a helper on your machine | Helper |
+|---|---|
+| Word, Excel and PowerPoint 97–2003 (`.doc`, `.xls`, `.ppt`), Pages, Numbers, Keynote and others | [LibreOffice](https://www.libreoffice.org) |
+| Images and scanned PDFs | [Tesseract](https://tesseract-ocr.github.io) (with the Turkish language data), plus Poppler's `pdftoppm` for PDFs |
+| PDFs the built-in reader cannot parse | Poppler's `pdftotext` |
+
+`llmgw-kb formats` says which helpers it found. Text in UTF-8, UTF-16, or
+the Windows Turkish and Western code pages is read correctly.
+
+A file that cannot be read is skipped with the reason, and the run goes on.
+Reasons include a locked Office file, a format with no helper, and a
+scanned page with no OCR.
+
+### What leaves your machine
+
+For each new or changed document, one request leaves the machine, over the
+gateway's TLS connection: its name, its file's SHA-256 digest and size, and
+its passages' text. The gateway embeds the passages through `/v1/embeddings`
+as you, so your key's limits, budget and guardrails apply as if you had
+asked. That is about a kilobyte a passage; embedding here (`--embed-here`)
+brings each vector down and sends it up again, some thirty times more.
+
+The file itself is never sent. `--dry-run` shows exactly what would be.
+
+### The ledger
+
+`llmgw-kb` keeps a small SQLite database of what it did with every
+document: the file it came from, its size, time and digest, and whether it
+was sent, skipped or failed, with the reason.
+
+| System | Where it is |
+|---|---|
+| Windows | `%AppData%\llmgw-kb\ledger.db` |
+| macOS | `~/Library/Application Support/llmgw-kb/ledger.db` |
+| Linux | `~/.config/llmgw-kb/ledger.db` |
+
+Deleting the ledger is safe: the next run hashes every file once, finds what
+the collection already has, and sends nothing that is unchanged. `--ledger`
+uses another file.
+
+## Asking questions of a collection
+
+A chat request names the collections to answer from in a header. Any client
+that can add a header can do it: the OpenAI and Anthropic SDKs, Continue, or
+curl.
+
+```sh
+curl https://llm.example.com/v1/chat/completions \
+  -H "Authorization: Bearer $LLMGW_API_KEY" \
+  -H "X-Llmgw-Knowledge: politikalar" \
+  -d '{"model": "qwen2.5-coder-14b-instruct",
+       "messages": [{"role": "user", "content": "Kaç gün yıllık iznim var?"}]}'
+```
+
+```python
+client.chat.completions.create(
+    model="qwen2.5-coder-14b-instruct",
+    messages=[{"role": "user", "content": "Kaç gün yıllık iznim var?"}],
+    extra_headers={"X-Llmgw-Knowledge": "politikalar"},
+)
+```
+
+What the gateway does with such a request:
+
+1. **It finds the collections.** It looks up the ones named, by name or ID,
+   among those the key may read. A collection it may not read is answered
+   404, the same as one that does not exist.
+2. **It embeds the question.** The last user message is embedded by the
+   collection's model, on the caller's own account: their limits, budget and
+   guardrails apply.
+3. **It searches.** The collections are searched by meaning and by words
+   (below).
+4. **It adds the passages.** The best passages are numbered and added to the
+   system message, with an instruction to answer from them, cite them as
+   [1], [2], and say so when they do not hold the answer. The system message
+   is created if the request has none. Passages that would not fit the
+   model's context are left out, the weakest first.
+5. **It judges and sends the request.** The request, passages included, then
+   goes through the guardrails and to the model as usual. A document
+   carrying instructions aimed at its reader is judged like anything else
+   the model reads.
+
+On the portal's chat page, **Documents** picks a collection the person may
+read in the chosen project, the default, or none. Under each answer, the
+sources it was given are listed by number.
+
+How it is shaped:
+
+| | |
+|---|---|
+| `X-Llmgw-Knowledge` | Collections, comma separated; `none` for no passages, whatever the default. |
+| `X-Llmgw-Knowledge-K` | How many passages, 1 to 20; 5 by default. |
+| `X-Llmgw-Project` | For a personal key: the project it acts in. |
+| `x-llmgw-sources` (answer) | The passages given, as `document#chunk`, in their numbered order; empty when nothing matched. |
+
+Behaviour to know about:
+
+- **No passages matched.** The model is told that nothing in the documents
+  answers the question, and asked to say so rather than answer from general
+  knowledge.
+- **Other APIs.** Requests through `/v1/responses` and `/v1/messages`
+  (Codex, Claude Code) work the same way, since they are answered as chat
+  completions.
+- **Separate prefix caches.** On vLLM replicas, a request given passages
+  carries a `cache_salt` derived from the collections they came from.
+  Requests that could be given the same passages still share cached blocks,
+  and no one else's do: not another organization's, nor another team's of
+  the same one. How fast an answer starts therefore says nothing about what
+  someone else asked.
+- **No answer cache.** Answers from knowledge bases are never served from
+  the answer cache, because two keys of one project can read different
+  collections.
+
+### How a passage is chosen
+
+Each passage scores the following, and the highest scores win:
+
+- its similarity in meaning to the question (cosine, 0 to 1), plus
+- up to 0.3 for the share of the question's words it contains.
+
+Words almost every passage has ("ne", "için", "the", "of") are not counted.
+The meaning does most of the work: "how much leave do fathers get" finds the
+paternity page, not the annual-leave page that shares more of its words.
+The word match helps with what meaning cannot carry, such as a form number,
+a product code or a name.
+
+Metrics: `llmgw_kb_retrievals_total{result}` (found, nothing, refused) and
+`llmgw_kb_retrieval_seconds`.
+
+## Training an adapter on a collection
+
+An adapter (LoRA) changes how a model *writes*: the organisation's words,
+format and tone. It does not reliably teach facts. On the PoC, a small
+model trained on nine policy documents answered 4 of 9 rephrased questions
+right and mixed facts up on the rest, while the same questions through the
+knowledge base got 9 of 9. Use both: the knowledge base for what answers
+say, an adapter for how they sound.
+
+```sh
+# 1. Questions and answers from the collection, written by a model of the
+#    cluster. Read the file through and correct it.
+llmgw ctl kb dataset Acme/policies --model qwen2.5-coder-14b-instruct --out policies.jsonl
+
+# 2. Train on a node. --gpu-memory is what the job needs of a card; it
+#    runs beside an engine when the engine leaves that much free.
+llmgw ctl train start policies-tone --model qwen2.5-0.5b-instruct \
+  --base-repo Qwen/Qwen2.5-0.5B-Instruct --node cuda01 --gpu-memory 2300 --data policies.jsonl
+llmgw ctl train show JOB --follow
+
+# 3. Make it the model's adapter, kept on the mirror.
+llmgw ctl train publish JOB --restart
+```
+
+The node trains with its installed vLLM's torch and a small overlay (peft
+and accelerate), as the engines' user. For a model llama.cpp serves, it also
+converts the adapter to GGUF with the installed llama.cpp's own code. What
+it makes is kept by the control plane (`llmgw ctl train download`).
+
+An adapter is private like its collection:
+
+- **It has an owner.** An adapter trained on a collection carries some of
+  what the collection says in its weights, so it belongs to a part of the
+  tree.
+  - By default the owner is the collection's owner.
+  - `--owner` names another, which must be able to read every collection
+    the pairs came from.
+  - Only callers in the owner's part of the tree see the adapter. To anyone
+    else it is a model that does not exist, absent from `/v1/models` too.
+- **It is public only by choice.** `--public` makes an adapter every
+  caller's. It is refused for pairs from a knowledge base, so it is only for
+  pairs written by hand.
+
+## How it is kept apart
+
+Passages live in PostgreSQL, partitioned by organization, with an index per
+collection. Every search and every write names the organization and the
+collections the caller may use. On top of that, row-level security on the
+table checks the same thing again, so a query that forgot its filter
+returns nothing rather than another tenant's passages.

docs/kong.md

@@ -0,0 +1,95 @@
+# llmgw and Kong
+
+For a site that already runs Kong. The two meet over plain HTTP: Kong in
+front of llmgw, as one more API it governs, or Kong behind llmgw, as an
+OpenAI-compatible provider. Nothing of either is built into the other.
+
+Tried with Kong Gateway 3.9 (`kong:3.9`, the last open-source release; 3.10
+and later need a Kong licence) in DB-less mode, in front of llmgw.
+
+## Kong in front of llmgw
+
+Two ways, both tried.
+
+**A plain route.** Kong forwards the request as it came. Tried with chat
+completions, streamed and not; llmgw's other APIs — Anthropic's
+`/v1/messages`, `/v1/responses`, embeddings, pictures — go the same way,
+since Kong does not read them. Kong adds its own authentication, rate limits
+and logging; llmgw still checks the key, the budget and the guardrails.
+
+```yaml
+_format_version: "3.0"
+services:
+- name: llmgw
+  url: https://llm.example.com          # llmgw's client port
+  routes:
+  - name: llmgw
+    paths: [/llmgw]                      # /llmgw/v1/chat/completions → /v1/chat/completions
+```
+
+**Kong's `ai-proxy` plugin**, for Kong's own token statistics on a chat
+route. llmgw is an OpenAI provider whose `upstream_url` is llmgw:
+
+```yaml
+services:
+- name: llmgw-chat
+  url: https://llm.example.com
+  routes:
+  - name: chat
+    paths: [/chat]
+  plugins:
+  - name: ai-proxy
+    config:
+      route_type: llm/v1/chat
+      logging: {log_statistics: true, log_payloads: false}
+      model:
+        provider: openai
+        name: qwen-coder                 # a model llmgw serves
+        options:
+          upstream_url: https://llm.example.com/v1/chat/completions
+```
+
+**Keys.** Without an `auth` block, `ai-proxy` passes the client's
+`Authorization` header on: each team keeps its own llmgw key, and llmgw's
+budgets, limits, guardrails and usage stay per team. With
+`auth: {header_name: Authorization, header_value: "Bearer sk-llmgw-…"}` Kong
+sends one key of its own instead, and llmgw sees every caller as that key —
+then budgets per team are Kong's to keep.
+
+**What Kong counts.** For a whole answer Kong reads llmgw's `usage` and
+records prompt and completion tokens as llmgw did. For a streamed answer
+Kong 3.9 records an estimate of the prompt tokens and no completion tokens,
+even with `stream_options.include_usage` set — the usage chunk reaches it
+and is not read. llmgw's own accounting counts both; use its usage export
+(`llmgw ctl usage export`) or `/admin/usage` for figures to bill by.
+
+**Timeouts.** A long answer from a model still loading outlasts Kong's
+60-second defaults: set the service's `read_timeout` to the time a model may
+take to start, as for nginx in [install.md](install.md).
+
+## Kong behind llmgw
+
+Not tried here. Kong's AI routes speak OpenAI's API, so llmgw reaches models Kong fronts as
+it reaches any provider: a model whose source is `external`.
+
+```yaml
+- name: kong-gpt
+  source: external
+  external:
+    url: https://kong.example.com/openai     # Kong's route, without /v1
+    model: gpt-4o-mini
+    api_key_file: /etc/llmgw/secrets/kong-consumer.key
+```
+
+The key is the one Kong's own authentication asks of a consumer. As a
+fallback (`fallbacks: [kong-gpt]`) it is used only when the cluster has no
+room.
+
+## Which to choose
+
+Kong in front when the site's API policy says every API goes through Kong:
+llmgw is then one more upstream, and its per-team accounting still holds
+when each team keeps its own key. Kong behind when the site already reaches
+outside providers through Kong and llmgw should too. llmgw does not need
+Kong for anything it does itself — keys, budgets, guardrails, routing to
+its own replicas.

docs/large-models.md

@@ -0,0 +1,145 @@
+# Large models: several cards, several nodes
+
+A model too big for one card runs on several. llmgw chooses how from the
+model's own figures and the cluster as it is; this page is the rules it
+follows, the commands, and what to look at when a first one is brought up
+on new hardware.
+
+## How a model is split
+
+In order, the first that fits:
+
+1. **One card.** Nothing to split.
+2. **Several cards of one node, tensor parallel.** Every layer is cut
+   across the cards, so they talk for every token: they must be on one
+   NVLink domain. The agent reads `nvidia-smi topo -m` and gives a replica
+   the cards with the most NVLink between them
+   ([install.md](install.md#nvlink-if-your-gpus-have-it)).
+3. **Several nodes, pipeline parallel between them.** Each node holds a
+   run of layers on all its cards (tensor parallel inside it), and only a
+   layer's output crosses to the next node. The nodes must share one
+   InfiniBand or RoCE fabric: a replica never spans two, and never a node
+   with none. Any number of nodes, not only two or four.
+
+The plan takes the fewest cards, then the fewest nodes, then the better
+precision up to eight bits.
+
+**Tensor width and the KV heads.** The cards split the model's KV heads
+among them, or past their number each card holds a copy of one. So the
+width divides the KV heads or is a multiple of them: a model with 8 KV
+heads runs 2, 4, 8 or 16 wide, not 12. Qwen3-235B has 4 and runs on 8.
+
+**Mixture of experts.** A model of experts (Qwen3-235B-A22B, DeepSeek-V3,
+Llama 4, Mixtral) on more than one card is served with vLLM's
+`--enable-expert-parallel`: each card holds whole experts rather than a
+slice of every one, which is less traffic per token. It is read from
+`config.json` (`num_experts`, `n_routed_experts`, `num_local_experts`).
+
+**When it does not fit.** A smaller precision first: an FP8 or AWQ
+repository of the same model, or a GGUF quantisation on llama.cpp if one
+node's cards hold it. Then more nodes. `llmgw ctl model plan REPO
+--precision bf16,fp8` shows each precision's fit from the model's config.
+
+## What a way needs
+
+Each way the plan offers says what a replica needs of each card, part by
+part, and of all its cards together:
+
+```
+$ llmgw ctl model plan meta-llama/Llama-3.3-70B-Instruct --format safetensors --context 32768
+FILE      SIZE    BEST WAY        A CARD   OF     ALL CARDS  NOW
+original  131 GB  4 × h100-80g    37.4 GB  72 GB  150 GB     free
+
+Best: original on 4 × h100-80g: 37.4 GB a card (weights 32.9 + context 2.5 + engine 2.0) of 72 usable; 150 GB on 4 cards.
+The context costs 0.31 GB per 1000 tokens, the whole model (--context to plan another).
+```
+
+- **weights:** the files, divided over the cards.
+- **context:** the KV cache for the context planned, divided likewise.
+- **engine:** the engine's own: 2 GB and a fiftieth of the weights for vLLM.
+- **usable:** 90% of a card; the rest is the driver's and fragmentation's.
+
+The console's **Serve a model** shows the same for every way a file runs.
+A context longer than planned is checked against it before anything is
+fetched.
+
+## Serving one
+
+Without flags the plan's best way is taken:
+
+```
+llmgw ctl model serve Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 --format safetensors --context-length 32768
+```
+
+Or a way given by hand, for instance two nodes of eight:
+
+```
+llmgw ctl model serve meta-llama/Llama-3.3-70B-Instruct --format safetensors \
+    --gpu-class h100-80g --gpus 8 --nodes 2 --env NCCL_DEBUG=INFO
+```
+
+`--gpus` is the cards on each node and `--nodes` the nodes. The pipeline
+is the nodes and the tensor width the cards of each.
+
+`--env KEY=VALUE` goes into the engine's environment on every node:
+`NCCL_DEBUG=INFO` while bringing a cluster up, `NCCL_IB_*` to steer it.
+Only the variables engines are tuned with are taken. llmgw already sets
+these:
+- `NCCL_IB_HCA` to the fabric's adapters;
+- `NCCL_IB_GID_INDEX` on RoCE;
+- `NCCL_SOCKET_IFNAME` to the management network;
+- for an NVLink domain spanning nodes, `NCCL_MNNVL_ENABLE`.
+
+The job's first step says what it needs of each card. The weights then
+come to the mirror once and from it to each node. The startup timeout (20
+minutes unless the profile's `startup_timeout` says otherwise) runs from
+when a node's download last moved. So a model of hundreds of gigabytes is
+not failed while it is still arriving, and one whose download stops is
+failed that long after.
+
+## Watching it come up
+
+| What | Where |
+|---|---|
+| The replica and its shards, one per node, each with its download | `llmgw ctl replicas` |
+| Why it is not placed | `llmgw ctl explain-placement MODEL` |
+| A node's engine output (NCCL's lines with `NCCL_DEBUG=INFO`) | `llmgw ctl logs NODE --source SHARD` |
+| The bandwidth between two nodes, rail by rail | `llmgw ctl fabric-test NODE NODE` |
+| Each node's GPUs, NVLink and adapters | `llmgw doctor` on the node; its page under /admin |
+| Every step of a start, across nodes | the `replica start` trace ([install.md](install.md)) |
+
+With `NCCL_DEBUG=INFO`, NCCL says which transport it took between nodes:
+`NET/IB` is the fabric. `NET/Socket` means it fell back to TCP, and the
+model will run, slowly.
+
+## Bringing it up on new hardware
+
+A first run on a fabric, in order. Each step is worth its output if
+something goes wrong.
+
+1. **The nodes.** Add each with its fabric:
+   `llmgw node add host --fabric ib-a:mlx5_0,mlx5_1`. Then
+   `llmgw ctl nodes` shows every node ready, on `ib-a`.
+2. **The fabric.** `llmgw ctl fabric-test node1 node2`: each rail at its
+   rated speed (a 400 Gb/s rail near 45 GB/s).
+3. **One node, several cards.** A dense 70B at bf16 on four or eight
+   cards of one node, by the plan:
+   `llmgw ctl model serve meta-llama/Llama-3.3-70B-Instruct --format safetensors`.
+   Ready, then a question answered.
+4. **Two nodes.** The same model forced over two:
+   `--gpu-class CLASS --gpus 8 --nodes 2 --env NCCL_DEBUG=INFO`, under
+   another `--name`. Ready, answering, and `NET/IB` in both shards' logs.
+5. **Experts.** A mixture-of-experts model on one node's eight cards,
+   for instance `Qwen/Qwen3-235B-A22B-Instruct-2507-FP8`: its profile
+   carries `--enable-expert-parallel` (`llmgw ctl models`, or its
+   definition from the admin API).
+6. **A node lost.** With the two-node replica serving, stop one node's
+   agent (`systemctl stop llmgw-agent`): the replica fails and is placed
+   again once the node is back or another has room. Nothing is left
+   holding cards.
+
+For each, worth keeping:
+- the time from serve to ready;
+- tokens a second from `llmgw ctl bench`;
+- the plan's per-card figure against what `nvidia-smi` shows once loaded;
+- anything it said that was wrong.

docs/onboarding.md

@@ -54,10 +54,20 @@ on the page that creates it and never again.
 
 ### Connecting an editor
 
-The account page's **Connect your editor** makes a key for Continue, Codex or
-Cline and shows what the tool needs with that key, this gateway's address and a
-model already in it:
+The account page's **Connect your editor** makes a key for VS Code, Continue,
+Codex or Cline and shows what the tool needs with that key, this gateway's
+address and a model already in it:
 
+- **VS Code** — llmgw's own extension, downloaded from the gateway at
+  `/setup/vscode/llmgw.vsix` with this gateway's address already in it. It
+  offers every chat model the key may use to VS Code itself: the chat's model
+  picker, agent mode, and any extension that asks VS Code for a language model
+  (Cline's "VS Code LM API" provider among them), with no setup of their own.
+  **llmgw: Sign In** in VS Code opens this page with VS Code chosen; the page's
+  **Open in VS Code** link hands the key back, and the extension keeps it in
+  VS Code's secret storage. A link the extension did not ask for is confirmed
+  first, with the gateway's address shown, since it decides where code is sent.
+  **llmgw: Enter API Key** takes a pasted one instead.
 - **Continue** — an assistant file of its own, `~/.continue/assistants/llmgw.yaml`,
   so their `config.yaml` is not touched.
 - **Codex** — a provider and a profile in `~/.codex/config.toml`, run with

docs/quick-install.md

@@ -8,7 +8,7 @@ explains every choice.
 
 | | |
 |---|---|
-| **A control plane** | One machine, or three for high availability (the store replicated, either one able to take over). Linux, x86-64 or ARM64: Debian 12, Ubuntu 22.04+ or RHEL 10. No GPU needed; 2 GB RAM, 20 GB disk each — 8 GB RAM when it is also a node running the starter model and the two judges. |
+| **A control plane** | One machine, or three for high availability (the store replicated, either one able to take over). Linux, x86-64 or ARM64: Debian 12, Ubuntu 22.04+ or RHEL 10. No GPU needed; 2 GB RAM, 20 GB disk each — 12 GB RAM when it is also a node running the starter model and the two judges, which setup checks before it installs anything; four cores or more, or the judges on machines of their own, keep a judgment within its wait. |
 | **GPU nodes** (optional to start) | The same Linux, with the NVIDIA driver installed. vLLM needs driver 525 or newer and a Turing GPU or newer; a V100 gets an older vLLM. |
 | **root** | on the control plane. On the other two of a three-machine setup, `sudo` (its password is asked for if it needs one); on nodes, `sudo` without a password. |
 | **SSH** | from the control plane to the other machines. The setup screens take a key, the ssh agent or a password; `node add` uses ssh keys. A machine nothing may log into joins with a code instead (`llmgw join`). |
@@ -146,7 +146,35 @@ agent is stopped.
 ## 4. Add a model
 
 Open `/admin/mirror` and choose **Explore models**. It lists what is popular on
-Hugging Face, showing only what your nodes can run.
+Hugging Face, showing only what your nodes can run: judged by the cluster as it
+is now — the largest card and its compute capability, the most memory on one
+node — so a model too big for them, in a form those cards lack (FP8 needs 8.9),
+or in neither GGUF nor safetensors is left out, and the page says how many.
+Without GPUs it lists GGUF only. Untick **only what these nodes can run** to
+see everything. On a model's page a file too big for the cluster is shown
+without a **Fetch** button, with what it would need.
+
+The **CPU tok/s** column is about how fast one reply would come on the CPU of
+the fastest node whose memory holds the model at Q4_K_M; a model's page gives
+it for each GGUF file by its own size. Every node's agent measures how fast its
+memory reads when it starts (it is in the agent's log as `memory_gbs`), and a
+token on a CPU reads every weight once: the estimate is that speed over the
+model's size — for a model of experts named like `30B-A3B`, its active part.
+On the test cluster it came within 15% of what llama.cpp measured. Several
+replies at once share the reads and each is slower.
+
+**GPU tok/s** is the same on the cards, for the layout this cluster would
+give the model — the fewest cards that hold it, of each class, the fastest of
+those; its tooltip says which ("2× h100-80g, NVLink, vLLM", "2 nodes × 1×
+a100-80g over ib0 (InfiniBand 200 Gb/s), vLLM"). A card's memory bandwidth is
+known from its model name (an H100 SXM and an H100 PCIe differ by half); a card
+not in the table gets no figure. Tensor parallel across a node's cards divides
+what each reads, and adds two all-reduces a layer, short over NVLink and
+longer over PCIe; llama.cpp across cards splits by layer, so the cards take
+turns and a reply is no faster than on one; a pipeline across nodes takes turns
+too, and adds a hop over the fabric for each node. One card's figures were
+measured on the test cluster's RTX 6000; tensor parallel and the fabric are
+from what NCCL is known to take, not measured here.
 
 1. Click a model to read its card.
 2. Pick one of its files, usually a `q4_k_m` for a first try, and click

go.mod

@@ -12,12 +12,14 @@ require (
 	github.com/charmbracelet/x/ansi v0.11.6
 	github.com/coreos/go-oidc/v3 v3.21.0
 	github.com/go-jose/go-jose/v4 v4.1.4
+	github.com/go-ldap/ldap/v3 v3.4.15
 	github.com/go-zeromq/zmq4 v0.17.0
 	github.com/jackc/pgx/v5 v5.11.0
 	github.com/knadh/koanf/parsers/yaml v1.1.1
 	github.com/knadh/koanf/providers/env/v2 v2.0.1
 	github.com/knadh/koanf/providers/file v1.2.1
 	github.com/knadh/koanf/v2 v2.3.6
+	github.com/ledongthuc/pdf v0.0.0-20260907135840-6c8c28e0e8a0
 	github.com/mattn/go-isatty v0.0.24
 	github.com/prometheus/client_golang v1.24.1
 	github.com/prometheus/client_model v0.6.3
@@ -34,6 +36,8 @@ require (
 	golang.org/x/net v0.59.0
 	golang.org/x/oauth2 v0.37.0
 	golang.org/x/sync v0.23.0
+	golang.org/x/sys v0.48.0
+	golang.org/x/term v0.46.0
 	golang.org/x/text v0.42.0
 	google.golang.org/protobuf v1.36.12
 	modernc.org/sqlite v1.59.0
@@ -42,6 +46,7 @@ require (
 require (
 	filippo.io/edwards25519 v1.2.0 // indirect
 	filippo.io/hpke v0.4.0 // indirect
+	github.com/Azure/go-ntlmssp v0.1.1 // indirect
 	github.com/aymanbagabas/go-osc52/v2 v2.0.1 // indirect
 	github.com/beorn7/perks v1.0.1 // indirect
 	github.com/cenkalti/backoff/v5 v5.0.3 // indirect
@@ -54,6 +59,7 @@ require (
 	github.com/dustin/go-humanize v1.0.1 // indirect
 	github.com/erikgeiser/coninput v0.0.0-20211004153227-1c3628e74d0f // indirect
 	github.com/fsnotify/fsnotify v1.9.0 // indirect
+	github.com/go-asn1-ber/asn1-ber v1.5.8 // indirect
 	github.com/go-logr/logr v1.4.4 // indirect
 	github.com/go-logr/stdr v1.2.2 // indirect
 	github.com/go-viper/mapstructure/v2 v2.5.0 // indirect
@@ -86,7 +92,6 @@ require (
 	go.opentelemetry.io/otel/exporters/otlp/otlptrace v1.46.0 // indirect
 	go.opentelemetry.io/otel/metric v1.46.0 // indirect
 	go.opentelemetry.io/proto/otlp v1.11.0 // indirect
-	golang.org/x/sys v0.48.0 // indirect
 	google.golang.org/genproto/googleapis/api v0.0.0-20260819154853-08b0e4226688 // indirect
 	google.golang.org/genproto/googleapis/rpc v0.0.0-20260819154853-08b0e4226688 // indirect
 	google.golang.org/grpc v1.83.1 // indirect

go.sum

@@ -8,8 +8,12 @@ filippo.io/edwards25519 v1.2.0 h1:crnVqOiS4jqYleHd9vaKZ+HKtHfllngJIiOpNpoJsjo=
 filippo.io/edwards25519 v1.2.0/go.mod h1:xzAOLCNug/yB62zG1bQ8uziwrIqIuxhctzJT18Q77mc=
 filippo.io/hpke v0.4.0 h1:p575VVQ6ted4pL+it6M00V/f2qTZITO0zgmdKCkd5+A=
 filippo.io/hpke v0.4.0/go.mod h1:EmAN849/P3qdeK+PCMkDpDm83vRHM5cDipBJ8xbQLVY=
+github.com/Azure/go-ntlmssp v0.1.1 h1:l+FM/EEMb0U9QZE7mKNEDw5Mu3mFiaa2GKOoTSsNDPw=
+github.com/Azure/go-ntlmssp v0.1.1/go.mod h1:NYqdhxd/8aAct/s4qSYZEerdPuH1liG2/X9DiVTbhpk=
 github.com/NVIDIA/go-nvml v0.13.4-0 h1:o3jp9u2x1R9ShFE3v+Aesp55XOSIQFMJz/VGNUcJaNE=
 github.com/NVIDIA/go-nvml v0.13.4-0/go.mod h1:id63qwpoDWpFXwnwM6psDCSqW4BmNu6mWpr4YeQtPGo=
+github.com/alexbrainman/sspi v0.0.0-20250919150558-7d374ff0d59e h1:4dAU9FXIyQktpoUAgOJK3OTFc/xug0PCXYCqU0FgDKI=
+github.com/alexbrainman/sspi v0.0.0-20250919150558-7d374ff0d59e/go.mod h1:cEWa1LVoE5KvSD9ONXsZrj0z6KqySlCCNKHlLzbqAt4=
 github.com/aymanbagabas/go-osc52/v2 v2.0.1 h1:HwpRHbFMcZLEVr42D4p7XBqjyuxQH5SMiErDT4WkJ2k=
 github.com/aymanbagabas/go-osc52/v2 v2.0.1/go.mod h1:uYgXzlJ7ZpABp8OJ+exZzJJhRNQ2ASbcXHWsFqH8hp8=
 github.com/beorn7/perks v1.0.1 h1:VlbKKnNfV8bJzeqoa4cOKqO6bYr3WgKZxO8Z16+hsOM=
@@ -46,8 +50,12 @@ github.com/erikgeiser/coninput v0.0.0-20211004153227-1c3628e74d0f h1:Y/CXytFA4m6
 github.com/erikgeiser/coninput v0.0.0-20211004153227-1c3628e74d0f/go.mod h1:vw97MGsxSvLiUE2X8qFplwetxpGLQrlU1Q9AUEIzCaM=
 github.com/fsnotify/fsnotify v1.9.0 h1:2Ml+OJNzbYCTzsxtv8vKSFD9PbJjmhYF14k/jKC7S9k=
 github.com/fsnotify/fsnotify v1.9.0/go.mod h1:8jBTzvmWwFyi3Pb8djgCCO5IBqzKJ/Jwo8TRcHyHii0=
+github.com/go-asn1-ber/asn1-ber v1.5.8 h1:H9AZkK22UOmfX8J84ubyaZxKJZ3FMHVwn8swoMML7iQ=
+github.com/go-asn1-ber/asn1-ber v1.5.8/go.mod h1:hEBeB/ic+5LoWskz+yKT7vGhhPYkProFKoKdwZRWMe0=
 github.com/go-jose/go-jose/v4 v4.1.4 h1:moDMcTHmvE6Groj34emNPLs/qtYXRVcd6S7NHbHz3kA=
 github.com/go-jose/go-jose/v4 v4.1.4/go.mod h1:x4oUasVrzR7071A4TnHLGSPpNOm2a21K9Kf04k1rs08=
+github.com/go-ldap/ldap/v3 v3.4.15 h1:eRB4f4Ecn1POL1mO86+uUDlecmyWsdiJwh60bAUS4UY=
+github.com/go-ldap/ldap/v3 v3.4.15/go.mod h1:jeD0c98EdZQCKKKKe4kTWs5aCf6y+j8KuYsYLznZSQk=
 github.com/go-logr/logr v1.2.2/go.mod h1:jdQByPbusPIv2/zmleS9BjJVeZ6kBagPoEUsqbVz/1A=
 github.com/go-logr/logr v1.4.4 h1:tG4xh9yMsRCAiodLVTxyrkzSZ9+o0L1Kg/+cPVcbP/8=
 github.com/go-logr/logr v1.4.4/go.mod h1:9T104GzyrTigFIr8wt5mBrctHMim0Nb2HLGrmQ40KvY=
@@ -69,6 +77,8 @@ github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
 github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
 github.com/grpc-ecosystem/grpc-gateway/v2 v2.30.0 h1:/Tnpcb2E0Pz/tN9s3bfEY2Q8ePCEX9iuS+cneUwncnw=
 github.com/grpc-ecosystem/grpc-gateway/v2 v2.30.0/go.mod h1:zOBXOsUaBSjKgmH4OGzV1esUpR3oUSCPYVd2cUBjKYY=
+github.com/hashicorp/go-uuid v1.0.3 h1:2gKiV6YVmrJ1i2CKKa9obLvRieoRGviZFL26PcT/Co8=
+github.com/hashicorp/go-uuid v1.0.3/go.mod h1:6SBZvOh/SIDV7/2o3Jml5SYk/TvGqwFJ/bN7x4byOro=
 github.com/hashicorp/golang-lru/v2 v2.0.7 h1:a+bsQ5rvGLjzHuww6tVxozPZFVghXaHOwFs4luLUK2k=
 github.com/hashicorp/golang-lru/v2 v2.0.7/go.mod h1:QeFd9opnmA6QUJc5vARoKUSoFhyfM2/ZepoAG6RGpeM=
 github.com/inconshreveable/mousetrap v1.1.0 h1:wN+x4NVGpMsO7ErUn/mUI3vEoE6Jt13X2s0bqwp9tc8=
@@ -81,6 +91,18 @@ github.com/jackc/pgx/v5 v5.11.0 h1:IzBBtyK9AHqf98cctWFifYSci2hgQR/cd56wB4p+ogg=
 github.com/jackc/pgx/v5 v5.11.0/go.mod h1:mal1tBGAFfLHvZzaYh77YS/eC6IX9OWbRV1QIIM0Jn4=
 github.com/jackc/puddle/v2 v2.2.2 h1:PR8nw+E/1w0GLuRFSmiioY6UooMp6KJv0/61nB7icHo=
 github.com/jackc/puddle/v2 v2.2.2/go.mod h1:vriiEXHvEE654aYKXXjOvZM39qJ0q+azkZFrfEOc3H4=
+github.com/jcmturner/aescts/v2 v2.0.0 h1:9YKLH6ey7H4eDBXW8khjYslgyqG2xZikXP0EQFKrle8=
+github.com/jcmturner/aescts/v2 v2.0.0/go.mod h1:AiaICIRyfYg35RUkr8yESTqvSy7csK90qZ5xfvvsoNs=
+github.com/jcmturner/dnsutils/v2 v2.0.0 h1:lltnkeZGL0wILNvrNiVCR6Ro5PGU/SeBvVO/8c/iPbo=
+github.com/jcmturner/dnsutils/v2 v2.0.0/go.mod h1:b0TnjGOvI/n42bZa+hmXL+kFJZsFT7G4t3HTlQ184QM=
+github.com/jcmturner/gofork v1.7.6 h1:QH0l3hzAU1tfT3rZCnW5zXl+orbkNMMRGJfdJjHVETg=
+github.com/jcmturner/gofork v1.7.6/go.mod h1:1622LH6i/EZqLloHfE7IeZ0uEJwMSUyQ/nDd82IeqRo=
+github.com/jcmturner/goidentity/v6 v6.0.1 h1:VKnZd2oEIMorCTsFBnJWbExfNN7yZr3EhJAxwOkZg6o=
+github.com/jcmturner/goidentity/v6 v6.0.1/go.mod h1:X1YW3bgtvwAXju7V3LCIMpY0Gbxyjn/mY9zx4tFonSg=
+github.com/jcmturner/gokrb5/v8 v8.4.4 h1:x1Sv4HaTpepFkXbt2IkL29DXRf8sOfZXo8eRKh687T8=
+github.com/jcmturner/gokrb5/v8 v8.4.4/go.mod h1:1btQEpgT6k+unzCwX1KdWMEwPPkkgBtP+F6aCACiMrs=
+github.com/jcmturner/rpc/v2 v2.0.3 h1:7FXXj8Ti1IaVFpSAziCZWNzbNuZmnvw/i6CqLNdWfZY=
+github.com/jcmturner/rpc/v2 v2.0.3/go.mod h1:VUJYCIDm3PVOEHw8sgt091/20OJjskO/YJki3ELg/Hc=
 github.com/klauspost/compress v1.19.1 h1:VsB4HPswih7mmZ8WleSFQ75c/Ui1M4trX5oAsJnhSlk=
 github.com/klauspost/compress v1.19.1/go.mod h1:cwPg85FWrGar70rWktvGQj8/hthj3wpl0PGDogxkrSQ=
 github.com/knadh/koanf/maps v0.1.2 h1:RBfmAW5CnZT+PJ1CVc1QSJKf4Xu9kxfQgYVQSu8hpbo=
@@ -95,6 +117,8 @@ github.com/knadh/koanf/v2 v2.3.6 h1:JoQPSJmvS4aP0xNc8xMDr5tcrkSEInL23/Il7pITAKo=
 github.com/knadh/koanf/v2 v2.3.6/go.mod h1:gRb40VRAbd4iJMYYD5IxZ6hfuopFcXBpc9bbQpZwo28=
 github.com/kylelemons/godebug v1.1.0 h1:RPNrshWIDI6G2gRW9EHilWtl7Z6Sb1BR0xunSBf0SNc=
 github.com/kylelemons/godebug v1.1.0/go.mod h1:9/0rRGxNHcop5bhtWyNeEfOS8JIWk580+fNqagV/RAw=
+github.com/ledongthuc/pdf v0.0.0-20260907135840-6c8c28e0e8a0 h1:7Q+xNAZFmnfYOMweHN3c/PDFUKKfY1pVJ26K++QvVfU=
+github.com/ledongthuc/pdf v0.0.0-20260907135840-6c8c28e0e8a0/go.mod h1:1fEHWurg7pvf5SG6XNE5Q8UZmOwex51Mkx3SLhrW5B4=
 github.com/lucasb-eyer/go-colorful v1.3.0 h1:2/yBRLdWBZKrf7gB40FoiKfAWYQ0lqNcbuQwVHXptag=
 github.com/lucasb-eyer/go-colorful v1.3.0/go.mod h1:R4dSotOR9KMtayYi1e77YzuveK+i7ruzyGqttikkLy0=
 github.com/mattn/go-isatty v0.0.24 h1:tGZZoVgT/KiqK1c8ocVLeDS8BSWMRd47J3Lbz7vsReI=

internal/account/directory_test.go

@@ -0,0 +1,63 @@
+package account
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/store"
+)
+
+type fakeDirectory struct {
+	asked []string
+	u     store.User
+	err   error
+}
+
+func (f *fakeDirectory) SignIn(_ context.Context, login, _ string) (store.User, error) {
+	f.asked = append(f.asked, login)
+	return f.u, f.err
+}
+
+// With a directory, a name the store has no password for is the
+// directory's to check; an account with a password here — the
+// administrator init made — is still checked here, so a directory that is
+// down locks nobody out.
+func TestPasswordsKeptElsewhereAreTheDirectorysToCheck(t *testing.T) {
+	svc, st := newService(t)
+	ctx := t.Context()
+	hash, _ := HashPassword("local-secret")
+	admin := store.User{ID: "usr-admin", Email: "admin@example.com", Name: "a", PasswordHash: hash,
+		State: store.UserActive, CreatedAt: time.Now()}
+	ada := store.User{ID: "usr-ada", Email: "ada@corp.example", Name: "Ada", PasswordHash: "!ldap",
+		State: store.UserActive, CreatedAt: time.Now()}
+	for _, u := range []store.User{admin, ada} {
+		if err := st.CreateUser(ctx, u); err != nil {
+			t.Fatal(err)
+		}
+	}
+	dir := &fakeDirectory{u: ada}
+	svc.SetDirectory(dir)
+
+	if _, _, err := svc.SignIn(ctx, "admin@example.com", "local-secret"); err != nil || len(dir.asked) != 0 {
+		t.Fatalf("the local administrator: %v, directory asked %v", err, dir.asked)
+	}
+	// A name, not an address, and an account kept by the directory.
+	for _, login := range []string{`CORP\ada`, "ada@corp.example"} {
+		if tok, _, err := svc.SignIn(ctx, login, "ad-secret"); err != nil || tok == "" {
+			t.Errorf("%s: %v", login, err)
+		}
+	}
+	if len(dir.asked) != 2 || dir.asked[0] != `CORP\ada` {
+		t.Errorf("the directory was asked %v", dir.asked)
+	}
+	dir.err = ErrBadCredentials
+	if _, _, err := svc.SignIn(ctx, "ada", "wrong"); !errors.Is(err, ErrBadCredentials) {
+		t.Errorf("a password the directory refused: %v", err)
+	}
+	// A reset link would give a directory's account a password here.
+	if _, _, err := svc.RequestReset(ctx, st, "ada@corp.example"); !errors.Is(err, ErrNoReset) {
+		t.Errorf("a directory's account was sent a reset link: %v", err)
+	}
+}

internal/account/service.go

@@ -80,8 +80,26 @@ type Service struct {
 
 	mu       sync.Mutex
 	attempts map[string]*attemptWindow
+
+	// dir checks the passwords of accounts this store keeps none for.
+	dir Directory
 }
 
+// Directory checks a password against an outside directory — Active
+// Directory over LDAP — and answers the account to sign in: made or kept in
+// step with the directory. Its errors are this package's ErrBadCredentials
+// and ErrSuspended for the person's own, and anything else for the
+// directory's.
+type Directory interface {
+	SignIn(ctx context.Context, login, password string) (store.User, error)
+}
+
+// SetDirectory has passwords for accounts with none here, and names that are
+// not addresses, checked by the directory. An account with a password here
+// — the administrator init made — is still checked here, so a directory that
+// is down locks nobody out.
+func (s *Service) SetDirectory(d Directory) { s.dir = d }
+
 type attemptWindow struct {
 	count int
 	since time.Time
@@ -137,6 +155,7 @@ func (s *Service) SignUp(ctx context.Context, email, name, password string) erro
 // SignIn checks a password and opens a session, returning the session token and
 // the CSRF token that must accompany the forms in it.
 func (s *Service) SignIn(ctx context.Context, email, password string) (token, csrf string, err error) {
+	login := strings.TrimSpace(email)
 	email = store.NormalizeEmail(email)
 	// Counted per address and the client trying it, so somebody guessing
 	// from elsewhere cannot lock the owner out of their own account; and per
@@ -151,6 +170,15 @@ func (s *Service) SignIn(ctx context.Context, email, password string) (token, cs
 	}
 
 	u, err := s.store.UserByEmail(ctx, email)
+	// No password kept here — the directory's, or nobody's — and a
+	// directory to ask: it decides.
+	if s.dir != nil && (errors.Is(err, store.ErrNotFound) || (err == nil && strings.HasPrefix(u.PasswordHash, "!"))) {
+		if u, err = s.dir.SignIn(ctx, login, password); err != nil {
+			return "", "", err
+		}
+		s.clearAttempts(byClient)
+		return s.open(ctx, u)
+	}
 	if errors.Is(err, store.ErrNotFound) {
 		// Verify anyway, against a hash that cannot match. Returning early here
 		// would let a stopwatch enumerate every account on the system.
@@ -180,7 +208,11 @@ func (s *Service) SignIn(ctx context.Context, email, password string) (token, cs
 	case store.UserSuspended:
 		return "", "", ErrSuspended
 	}
+	return s.open(ctx, u)
+}
 
+// open makes a session for a person whose password was right.
+func (s *Service) open(ctx context.Context, u store.User) (token, csrf string, err error) {
 	if token, err = store.NewSessionToken(); err != nil {
 		return "", "", err
 	}
@@ -402,7 +434,10 @@ func (s *Service) RequestReset(ctx context.Context, rs ResetStore, email string)
 		return "", store.User{}, ErrTooManyTries
 	}
 	u, err := rs.UserByEmail(ctx, email)
-	if err != nil || u.PasswordHash == "" || u.State != store.UserActive {
+	// An account whose password is kept elsewhere — "!oidc", "!scim",
+	// "!ldap" — gets no link: setting one here would give it a second way
+	// in that its directory knows nothing of, and outlives its leaving.
+	if err != nil || u.PasswordHash == "" || strings.HasPrefix(u.PasswordHash, "!") || u.State != store.UserActive {
 		return "", store.User{}, ErrNoReset
 	}
 	token, err := rs.CreatePasswordReset(ctx, u.ID, ResetLinkTTL)

internal/accounting/accounting.go

@@ -78,10 +78,41 @@ func (a *Accountant) RestoreBudgets(ctx context.Context, s store.Store, period s
 	if start.IsZero() {
 		return nil
 	}
-	rows, err := s.SpendSinceByScope(ctx, start)
+	spend, keys, err := storedSpend(ctx, s, start)
 	if err != nil {
 		return fmt.Errorf("restoring budgets: %w", err)
 	}
+	a.budgets.Restore(period, spend)
+	a.log.Info("restored budgets from the store",
+		"keys", keys, "counters", len(spend), "since", start)
+	return nil
+}
+
+// CatchUpBudgets raises each counter to what the store has recorded, where
+// that is more. Run every so often by a router that is one of several taking
+// requests: the others' spend reaches it only through the store. Never
+// lowers a counter, which leads the store by whatever this process has not
+// written yet.
+func (a *Accountant) CatchUpBudgets(ctx context.Context, s store.Store, period store.Period) error {
+	start := PeriodStart(period, time.Now())
+	if start.IsZero() {
+		return nil
+	}
+	spend, _, err := storedSpend(ctx, s, start)
+	if err != nil {
+		return fmt.Errorf("reading budgets: %w", err)
+	}
+	a.budgets.Raise(period, spend)
+	return nil
+}
+
+// storedSpend is the spend since start the store has recorded, by counter,
+// and how many keys it came from.
+func storedSpend(ctx context.Context, s store.Store, start time.Time) (map[string]int64, int, error) {
+	rows, err := s.SpendSinceByScope(ctx, start)
+	if err != nil {
+		return nil, 0, err
+	}
 	// Every scope a key sits under, not just the key. A budget owned by a team
 	// is one counter fed by every key beneath it, and restoring only per-key
 	// totals would bring the control plane back up believing the team had spent
@@ -113,10 +144,7 @@ func (a *Accountant) RestoreBudgets(ctx context.Context, s store.Store, period s
 			}
 		}
 	}
-	a.budgets.Restore(period, spend)
-	a.log.Info("restored budgets from the store",
-		"keys", len(rows), "counters", len(spend), "since", start)
-	return nil
+	return spend, len(rows), nil
 }
 
 // Reservation is what Check returns: the estimate it charged, so Record can
@@ -323,7 +351,7 @@ func (a *Accountant) Record(res Reservation, u store.Usage) {
 	if u.ServedModel != "" && u.ServedModel != u.Model && a.pricing != nil {
 		price = a.pricing.For(u.ServedModel)
 	}
-	u.CostMicros = price.Cost(u.PromptTokens, u.CompletionTokens)
+	u.CostMicros = price.Cost(u.PromptTokens, u.CachedTokens, u.CompletionTokens, u.Images, u.Meshes)
 	for _, b := range res.Budgets {
 		a.budgets.Add(b.Counter, b.Period, u.CostMicros)
 	}

internal/accounting/accounting_test.go

@@ -39,17 +39,44 @@ func TestCostRoundsUpSoSmallRequestsAreNotFree(t *testing.T) {
 	// Rounding down would make a stream of tiny requests cost nothing, which is
 	// exactly the traffic shape a quota exists to bound.
 	p := Price{PromptPer1k: 1000, CompletionPer1k: 3000}
-	if got := p.Cost(1, 0); got != 1 {
+	if got := p.Cost(1, 0, 0, 0, 0); got != 1 {
 		t.Errorf("one prompt token cost %d", got)
 	}
-	if got := p.Cost(1000, 1000); got != 4000 {
+	if got := p.Cost(1000, 0, 1000, 0, 0); got != 4000 {
 		t.Errorf("a thousand of each cost %d, want 4000", got)
 	}
-	if got := p.Cost(0, 0); got != 0 {
+	if got := p.Cost(0, 0, 0, 0, 0); got != 0 {
 		t.Errorf("an empty request cost %d", got)
 	}
 }
 
+func TestACachedPromptCostsTheCachedRate(t *testing.T) {
+	p := Price{PromptPer1k: 1000, CachedPer1k: 100, CompletionPer1k: 3000}
+	// 10,000 prompt tokens of which 9,000 were read from the cache: 1,000 at
+	// the prompt rate and 9,000 at the cached one.
+	if got, want := p.Cost(10000, 9000, 0, 0, 0), int64(1000+900); got != want {
+		t.Errorf("a 90%% cached prompt cost %d, want %d", got, want)
+	}
+	// More cached than prompt is an engine's miscount, not a refund.
+	if got := p.Cost(100, 500, 0, 0, 0); got != 10 {
+		t.Errorf("cached beyond the prompt cost %d, want 10", got)
+	}
+	// No cached rate: cached tokens cost what any prompt token does.
+	if got := (Price{PromptPer1k: 1000}).Cost(10000, 9000, 0, 0, 0); got != 10000 {
+		t.Errorf("without a cached rate a cached prompt cost %d, want 10000", got)
+	}
+}
+
+func TestAPictureCostsItsPriceOnTopOfItsPrompt(t *testing.T) {
+	p := Price{PromptPer1k: 1000, PerImage: 50000}
+	if got := p.Cost(20, 0, 0, 3, 0); got != 20+150000 {
+		t.Errorf("three pictures from a 20-token prompt cost %d", got)
+	}
+	if (Price{PerImage: 1}).Zero() {
+		t.Error("a model priced per picture reads as free, and would pass a spent budget")
+	}
+}
+
 func TestZeroPricedModelsStayAvailableWhenTheBudgetIsSpent(t *testing.T) {
 	// This is what makes budget fallback work: a key with nothing left can
 	// still be served something that costs nothing.
@@ -189,6 +216,22 @@ func TestBudgetSurvivesARestart(t *testing.T) {
 	}
 }
 
+func TestCatchingUpRaisesButNeverLowers(t *testing.T) {
+	// Another router's spend arrives through the store; this one's own, not
+	// written yet, must not be wiped by a store that is behind.
+	b := NewBudgets()
+	b.Add("mine", store.PeriodMonth, 900)
+	b.Add("shared", store.PeriodMonth, 100)
+
+	b.Raise(store.PeriodMonth, map[string]int64{"mine": 400, "shared": 700, "theirs": 50})
+
+	for key, want := range map[string]int64{"mine": 900, "shared": 700, "theirs": 50} {
+		if got := b.Spent(key, store.PeriodMonth); got != want {
+			t.Errorf("%s: spent %d, want %d", key, got, want)
+		}
+	}
+}
+
 func TestBudgetsResetAtThePeriodBoundary(t *testing.T) {
 	b := NewBudgets()
 	now := time.Date(2026, 1, 31, 23, 59, 0, 0, time.UTC)

internal/accounting/budget.go

@@ -61,6 +61,19 @@ func (b *Budgets) Restore(period store.Period, spend map[string]int64) {
 	}
 }
 
+// Raise lifts counters to the spend given, where it is more.
+func (b *Budgets) Raise(period store.Period, spend map[string]int64) {
+	start := PeriodStart(period, b.now())
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	for key, micros := range spend {
+		if c := b.currentLocked(key, start); micros > c.micros {
+			c.micros = micros
+			b.spend[key] = c
+		}
+	}
+}
+
 // Spent returns a key's spend in the current period, resetting it if the period
 // has rolled over.
 func (b *Budgets) Spent(keyID string, period store.Period) int64 {

internal/accounting/cost.go

@@ -18,6 +18,18 @@ import (
 type Price struct {
 	PromptPer1k     int64 `koanf:"prompt_per_1k"`
 	CompletionPer1k int64 `koanf:"completion_per_1k"`
+	// CachedPer1k is what a prompt token costs that the engine or provider
+	// read from its cache instead of computing: the part of prompt_tokens
+	// the answer counts as cached_tokens. Zero charges those at
+	// PromptPer1k, as before there was a rate for them.
+	CachedPer1k int64 `koanf:"cached_per_1k"`
+	// PerImage is what one picture costs, for a model that makes them: a
+	// picture's cost is the card's time, which its prompt's few tokens do
+	// not measure.
+	PerImage int64 `koanf:"per_image"`
+	// PerMesh is what one 3D model costs: minutes of a card, which no
+	// token count measures.
+	PerMesh int64 `koanf:"per_mesh"`
 }
 
 // Zero reports whether this model spends nothing.
@@ -25,14 +37,26 @@ type Price struct {
 // It matters beyond arithmetic: a zero-cost model stays available to a key
 // whose budget is spent, which is what makes the plan's budget fallback work
 // (§7.7).
-func (p Price) Zero() bool { return p.PromptPer1k == 0 && p.CompletionPer1k == 0 }
+func (p Price) Zero() bool {
+	return p.PromptPer1k == 0 && p.CompletionPer1k == 0 && p.CachedPer1k == 0 && p.PerImage == 0 && p.PerMesh == 0
+}
 
 // Cost computes what a request spent, rounding up.
 //
 // Up rather than to nearest: rounding down would make a stream of tiny requests
 // free, which is exactly the shape of traffic a quota exists to bound.
-func (p Price) Cost(promptTokens, completionTokens int64) int64 {
-	return divUp(promptTokens*p.PromptPer1k, 1000) + divUp(completionTokens*p.CompletionPer1k, 1000)
+//
+// cachedTokens are counted within promptTokens, as OpenAI's usage counts
+// them, and are charged at the cached rate instead of the prompt rate.
+func (p Price) Cost(promptTokens, cachedTokens, completionTokens, images, meshes int64) int64 {
+	cachedTokens = min(max(cachedTokens, 0), max(promptTokens, 0))
+	cachedRate := p.CachedPer1k
+	if cachedRate == 0 {
+		cachedRate = p.PromptPer1k
+	}
+	return divUp((promptTokens-cachedTokens)*p.PromptPer1k+cachedTokens*cachedRate, 1000) +
+		divUp(completionTokens*p.CompletionPer1k, 1000) +
+		max(images, 0)*p.PerImage + max(meshes, 0)*p.PerMesh
 }
 
 func divUp(n, d int64) int64 {

internal/accounting/limiter.go

@@ -25,7 +25,11 @@ type Limiter struct {
 type window struct {
 	// current and previous are counts in this minute and the one before.
 	current, previous int64
-	start             time.Time
+	// othersCurrent and othersPrevious are the same counts on the other
+	// control planes taking requests, as last heard: what makes a limit the
+	// cluster's rather than each process's.
+	othersCurrent, othersPrevious int64
+	start                         time.Time
 }
 
 // NewLimiter builds an empty limiter.
@@ -104,21 +108,24 @@ func (w *window) roll(now time.Time) {
 		return
 	case start.Sub(w.start) == windowLen:
 		w.previous, w.current = w.current, 0
+		w.othersPrevious, w.othersCurrent = w.othersCurrent, 0
 	default:
 		// More than one window has passed with no traffic.
 		w.previous, w.current = 0, 0
+		w.othersPrevious, w.othersCurrent = 0, 0
 	}
 	w.start = start
 }
 
 // rate weights the previous window by how much of it is still in view.
 func (w *window) rate(now time.Time) float64 {
+	cur, prev := w.current+w.othersCurrent, w.previous+w.othersPrevious
 	elapsed := now.Sub(w.start)
 	if elapsed >= windowLen {
-		return float64(w.current)
+		return float64(cur)
 	}
 	remaining := 1 - float64(elapsed)/float64(windowLen)
-	return float64(w.previous)*remaining + float64(w.current)
+	return float64(prev)*remaining + float64(cur)
 }
 
 // Forget drops a key's window. Used when a key is revoked; keeping it would

internal/accounting/shared.go

@@ -0,0 +1,110 @@
+package accounting
+
+import "time"
+
+// Counts are one limiter's counts by counter: this minute's and the one
+// before, as absolute figures, so reporting them twice changes nothing.
+type Counts struct {
+	Minute   time.Time        `json:"minute"`
+	Current  map[string]int64 `json:"current,omitempty"`
+	Previous map[string]int64 `json:"previous,omitempty"`
+}
+
+// Add sums other into c, aligned to c's minute: counts of a minute that is
+// c's previous one land there, older ones are dropped. Clocks a second apart
+// at a minute's turn are why.
+func (c *Counts) Add(other Counts) {
+	switch {
+	case other.Minute.Equal(c.Minute):
+		c.Current = addInto(c.Current, other.Current)
+		c.Previous = addInto(c.Previous, other.Previous)
+	case other.Minute.Add(windowLen).Equal(c.Minute):
+		c.Previous = addInto(c.Previous, other.Current)
+	}
+}
+
+func addInto(dst, src map[string]int64) map[string]int64 {
+	if len(src) == 0 {
+		return dst
+	}
+	if dst == nil {
+		dst = make(map[string]int64, len(src))
+	}
+	for k, v := range src {
+		dst[k] += v
+	}
+	return dst
+}
+
+// Counts reads this limiter's own counts, without the others'.
+func (l *Limiter) Counts() Counts {
+	now := l.now()
+	out := Counts{Minute: now.Truncate(windowLen), Current: map[string]int64{}, Previous: map[string]int64{}}
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	for k, w := range l.keys {
+		w.roll(now)
+		if w.current > 0 {
+			out.Current[k] = w.current
+		}
+		if w.previous > 0 {
+			out.Previous[k] = w.previous
+		}
+	}
+	return out
+}
+
+// SetOthers replaces what the other control planes have counted. A counter
+// they did not mention has nothing from them.
+func (l *Limiter) SetOthers(c Counts) {
+	now := l.now()
+	aligned := Counts{Minute: now.Truncate(windowLen)}
+	aligned.Add(c)
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	for k, w := range l.keys {
+		w.roll(now)
+		w.othersCurrent, w.othersPrevious = aligned.Current[k], aligned.Previous[k]
+		delete(aligned.Current, k)
+		delete(aligned.Previous, k)
+	}
+	// Counters only others have used yet: a key spread over the cluster is
+	// limited here from its first request.
+	for k := range aligned.Current {
+		w := l.windowLocked(k, now)
+		w.othersCurrent, w.othersPrevious = aligned.Current[k], aligned.Previous[k]
+		delete(aligned.Previous, k)
+	}
+	for k, v := range aligned.Previous {
+		l.windowLocked(k, now).othersPrevious = v
+	}
+}
+
+// Rates are an accountant's two limiters' counts.
+type Rates struct {
+	Requests Counts `json:"requests"`
+	Tokens   Counts `json:"tokens"`
+}
+
+// Add sums other into r.
+func (r *Rates) Add(other Rates) {
+	r.Requests.Add(other.Requests)
+	r.Tokens.Add(other.Tokens)
+}
+
+// NewRates is an empty sum for the minute now is in.
+func NewRates(now time.Time) Rates {
+	m := now.Truncate(windowLen)
+	return Rates{Requests: Counts{Minute: m}, Tokens: Counts{Minute: m}}
+}
+
+// Rates reads this process's own request and token counts.
+func (a *Accountant) Rates() Rates {
+	return Rates{Requests: a.requests.Counts(), Tokens: a.tokens.Counts()}
+}
+
+// SetOtherRates puts the other control planes' counts in force here.
+func (a *Accountant) SetOtherRates(r Rates) {
+	a.requests.SetOthers(r.Requests)
+	a.tokens.SetOthers(r.Tokens)
+}

internal/accounting/shared_test.go

@@ -0,0 +1,71 @@
+package accounting
+
+import (
+	"testing"
+	"time"
+)
+
+// Two control planes each allowing a key six of its ten a minute: once each
+// has heard the other's count, neither allows another.
+func TestALimitIsTheClustersOnceTheCountsAreShared(t *testing.T) {
+	now := time.Date(2026, 11, 1, 12, 0, 10, 0, time.UTC)
+	a, b := NewLimiter(), NewLimiter()
+	a.now, b.now = func() time.Time { return now }, func() time.Time { return now }
+	for range 6 {
+		if !a.Allow("k", 10, 1) || !b.Allow("k", 10, 1) {
+			t.Fatal("refused within the limit")
+		}
+	}
+	a.SetOthers(b.Counts())
+	b.SetOthers(a.Counts())
+	if a.Allow("k", 10, 1) || b.Allow("k", 10, 1) {
+		t.Error("twelve of ten allowed once both knew")
+	}
+	// Reporting the same counts again changes nothing: they are absolute.
+	a.SetOthers(b.Counts())
+	if got := a.Rate("k"); got != 12 {
+		t.Errorf("rate %v after the same counts twice", got)
+	}
+	// What the others counted is never charged back to them.
+	if c := a.Counts(); c.Current["k"] != 6 {
+		t.Errorf("own count %d", c.Current["k"])
+	}
+}
+
+// A key only the others have used is limited here from its first request,
+// and their counts move into the previous minute with the clock.
+func TestOthersCountsFollowTheMinute(t *testing.T) {
+	now := time.Date(2026, 11, 1, 12, 0, 50, 0, time.UTC)
+	a, b := NewLimiter(), NewLimiter()
+	a.now, b.now = func() time.Time { return now }, func() time.Time { return now }
+	for range 10 {
+		b.Allow("k", 10, 1)
+	}
+	a.SetOthers(b.Counts())
+	if a.Allow("k", 10, 1) {
+		t.Error("a key the cluster has used up was allowed where it had not been seen")
+	}
+	now = now.Add(20 * time.Second) // 12:01:10, a sixth into the next minute
+	if got := a.Rate("k"); got < 8 || got > 8.5 {
+		t.Errorf("rate %v: the others' last minute should weigh five sixths", got)
+	}
+	// Counts of a minute ago, sent late, land in the previous minute.
+	old := b.Counts()
+	old.Minute = now.Truncate(time.Minute).Add(-time.Minute)
+	old.Current, old.Previous = map[string]int64{"k": 3}, nil
+	a.SetOthers(old)
+	if got := a.Rate("k"); got < 2 || got > 3 {
+		t.Errorf("rate %v from late counts", got)
+	}
+}
+
+func TestRatesAddAlignsMinutes(t *testing.T) {
+	m := time.Date(2026, 11, 1, 12, 1, 0, 0, time.UTC)
+	sum := NewRates(m)
+	sum.Add(Rates{Requests: Counts{Minute: m, Current: map[string]int64{"k": 2}, Previous: map[string]int64{"k": 1}}})
+	sum.Add(Rates{Requests: Counts{Minute: m.Add(-time.Minute), Current: map[string]int64{"k": 5}}})
+	sum.Add(Rates{Requests: Counts{Minute: m.Add(-2 * time.Minute), Current: map[string]int64{"k": 100}}})
+	if sum.Requests.Current["k"] != 2 || sum.Requests.Previous["k"] != 6 {
+		t.Errorf("sum %+v", sum.Requests)
+	}
+}

internal/agent/agent.go

@@ -6,11 +6,13 @@ import (
 	"crypto/sha256"
 	"crypto/tls"
 	"encoding/hex"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"io"
 	"log/slog"
 	"maps"
+	"math"
 	"math/rand/v2"
 	"net"
 	"net/http"
@@ -34,6 +36,7 @@ import (
 	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/engines"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/provision"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/supervisor"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/train"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/weights"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/control/agentsvc"
 	"gitea.mynodes.xyz/ilker/llmgw/internal/firewall"
@@ -44,7 +47,9 @@ import (
 
 // Agent runs on a GPU node.
 type Agent struct {
-	cfg Config
+	// upgrading is true while an upgrade downloads.
+	upgrading atomic.Bool
+	cfg       Config
 	// answered says the current session heard from its control plane, which
 	// starts the reconnect backoff again when it ends.
 	answered atomic.Bool
@@ -136,6 +141,13 @@ type Agent struct {
 	// by mu.
 	kv     *kvevents.Mirror
 	kvSubs map[string]*kvevents.Subscriber
+
+	// settingsCheck says whether the cluster's settings for this agent
+	// would load beside its agent.yaml; see SetSettingsCheck.
+	settingsCheck func([]byte) error
+
+	// trainer runs LoRA training jobs, one at a time (BACKLOG F4).
+	trainer *train.Runner
 }
 
 // connectHTTPClient is the HTTP client connect uses; an interface so tests can
@@ -175,13 +187,25 @@ func New(cfg Config, httpClient connectHTTPClient, log *slog.Logger) (*Agent, er
 		logRing:      ring,
 		provKick:     make(chan struct{}, 1),
 		load:         discovery.NewLoadSampler(),
-		sensors:      discovery.NewSensorSampler(),
+		sensors:      discovery.NewSensorSampler().WithProfile(discovery.NewProfSampler(cfg.Discovery.DCGMExporter)),
 		wear:         discovery.NewHealthSampler(),
 		serverList:   loadControlPlanes(cfg.Servers, cfg.TLS.Enabled()),
 		kv:           kvevents.NewMirror(),
 		kvSubs:       map[string]*kvevents.Subscriber{},
 	}
 	a.prov = &provision.Installer{Root: cfg.EnginesDir, Log: log, OnStatus: a.provisionStatus}
+	runAs := cfg.Superv.RunAs
+	if runAs == "" {
+		runAs = supervisor.DefaultRunAs
+	}
+	// Where the engines are, as the installer has it when the file says
+	// nothing.
+	enginesDir := cfg.EnginesDir
+	if enginesDir == "" {
+		enginesDir = provision.DefaultRoot
+	}
+	a.trainer = train.New(train.Env{EnginesDir: enginesDir, WorkDir: filepath.Join(cfg.Superv.WorkDir, "trainer"),
+		RunAs: runAs, Log: log})
 	return a, nil
 }
 
@@ -371,8 +395,11 @@ func (a *Agent) settleAdopted(sh *supervisor.Shard, o supervisor.Orphan) {
 		_ = a.sup.Stop(ctx, o.ShardID)
 		return
 	}
-	// What it computes on, for one started before the agent kept that.
-	if a.sup.ComputeOf(o.ShardID) == nil {
+	// What it computes on, for one started before the agent kept that — or
+	// kept as the CPU for a GPU placement by an agent that did not follow
+	// an engine into a process group of its own, which the check above has
+	// just found on its GPUs.
+	if c := a.sup.ComputeOf(o.ShardID); c == nil || (len(o.Spec.GPUIndices) > 0 && c.Backend == "CPU") {
 		if c, ok := a.gpuCompute(ctx, o.Spec.GPUIndices, sh.Status().PID); ok {
 			a.sup.SetCompute(o.ShardID, c)
 		}
@@ -492,6 +519,8 @@ func (a *Agent) scan(ctx context.Context) error {
 	inv, err := discovery.Discover(ctx, opts, a.log)
 	host := discovery.DiscoverHost()
 	host.MinComputeCapability = discovery.MinComputeCapability(inv.GPUs)
+	bw := a.memBandwidth()
+	host.MemBandwidthGBs, host.MemBandwidthCoreGBs = bw.All, bw.Core
 	a.mu.Lock()
 	a.inv = inv
 	a.prov.Host = host
@@ -548,6 +577,10 @@ func (a *Agent) session(ctx context.Context, server string) error {
 	reg.Host = &llmgwv1.Host{
 		OsId: h.OSID, OsVersion: h.OSVersion, OsCodename: h.OSCodename,
 		Arch: h.Arch, CpuFlags: h.CPUFlags, NvidiaDriver: h.NVIDIADriver,
+		// What is measured by now, not what the startup scan found before
+		// the measurement ended: a reconnect would otherwise say 0.
+		MemBandwidthGbs:     a.memBandwidth().All,
+		MemBandwidthCoreGbs: a.memBandwidth().Core,
 	}
 
 	if err := stream.Send(&llmgwv1.AgentMessage{
@@ -645,6 +678,8 @@ func (a *Agent) heartbeat() *llmgwv1.AgentMessage {
 	// without anybody restarting anything.
 	hb.MemoryTotalMb = int64(discovery.MemoryTotalMB())
 	hb.CpuCount = int32(runtime.NumCPU())
+	bw := a.memBandwidth()
+	hb.MemBandwidthGbs, hb.MemBandwidthCoreGbs = bw.All, bw.Core
 
 	a.reprobeEngines()
 	a.mu.Lock()
@@ -766,6 +801,11 @@ func (a *Agent) sampleLoad() *llmgwv1.NodeLoad {
 				MemoryUsedMb: g.MemoryUsedMB, MemoryTotalMb: g.MemoryTotalMB,
 				SmClockMhz: g.SMClockMHz, EccUncorrected: g.ECCUncorrected,
 			}
+			if p := g.Profile; p != nil {
+				gs.HasProfile, gs.ProfileSource = true, p.Source
+				gs.SmActivePercent, gs.SmOccupancyPercent = p.SMActive, p.SMOccupancy
+				gs.TensorActivePercent, gs.DramActivePercent = p.TensorActive, p.DRAMActive
+			}
 			if w, ok := wear[g.Index]; ok {
 				gs.HasWear = true
 				gs.PcieReplays, gs.RetiredPagesSbe, gs.RetiredPagesDbe, gs.RetirementPending =
@@ -848,6 +888,7 @@ func (a *Agent) removeWeights(paths []string) {
 			a.log.Warn("could not remove cached weights", "path", p, "err", err)
 			continue
 		}
+		_ = os.Remove(weights.ChecksumPath(abs))
 		a.log.Info("removed cached weights no model asks for", "path", p)
 	}
 }
@@ -880,6 +921,12 @@ func (a *Agent) handle(ctx context.Context, msg *llmgwv1.ControlMessage, out cha
 		a.setControlPlanes(m.ControlPlanes)
 	case *llmgwv1.ControlMessage_Firewall:
 		go a.applyFirewall(m.Firewall)
+	case *llmgwv1.ControlMessage_AgentSettings:
+		go a.applySettings(m.AgentSettings)
+	case *llmgwv1.ControlMessage_TrainStart:
+		go a.trainJob(ctx, m.TrainStart, out)
+	case *llmgwv1.ControlMessage_TrainCancel:
+		a.trainer.Cancel(m.TrainCancel.GetJobId())
 	case *llmgwv1.ControlMessage_RestartAgent:
 		go func() {
 			a.log.Info("restarting, as the control plane asks; the engines keep serving",
@@ -1093,11 +1140,11 @@ func (a *Agent) startShard(ctx context.Context, m *llmgwv1.StartShard, out chan<
 		send(&llmgwv1.ShardStatus{
 			State: agentsvc.StateToWire(supervisor.Starting),
 			Fetch: &llmgwv1.FetchProgress{
-				Source: p.Source, BytesDone: p.BytesDone, BytesTotal: p.BytesTotal,
+				Source: p.Source, BytesDone: p.BytesDone, BytesTotal: p.BytesTotal, Phase: p.Phase,
 			},
 		})
 	})
-	if path, err := a.weights().Ensure(ctx, spec.Profile.ModelPath); err != nil {
+	if path, err := a.weights().EnsureChecked(ctx, spec.Profile.ModelPath, spec.Profile.AllowPickle, spec.Profile.ModelSHA256); err != nil {
 		stop()
 		report(supervisor.Failed, err.Error(), nil)
 		return
@@ -1105,11 +1152,37 @@ func (a *Agent) startShard(ctx context.Context, m *llmgwv1.StartShard, out chan<
 		stop()
 		spec.Profile.ModelPath = path
 	}
+	// A shared ComfyUI replica's members' weights, each fetched like the
+	// model's, and handed to the gate where they landed.
+	if raw := spec.Profile.Env[engines.SharedModelsEnv]; raw != "" {
+		var ms []map[string]string
+		if err := json.Unmarshal([]byte(raw), &ms); err != nil {
+			report(supervisor.Failed, "the shared replica's models are not readable: "+err.Error(), nil)
+			return
+		}
+		for i, m := range ms {
+			stop := a.watchFetch(m["source"], func(p weights.Progress) {
+				send(&llmgwv1.ShardStatus{State: agentsvc.StateToWire(supervisor.Starting),
+					Fetch: &llmgwv1.FetchProgress{Source: p.Source, BytesDone: p.BytesDone, BytesTotal: p.BytesTotal,
+						Phase: p.Phase}})
+			})
+			path, err := a.weights().Ensure(ctx, m["source"], m["allow_pickle"] == "true")
+			stop()
+			if err != nil {
+				report(supervisor.Failed, fmt.Sprintf("model %s: %v", m["name"], err), nil)
+				return
+			}
+			ms[i]["path"] = path
+		}
+		b, _ := json.Marshal(ms)
+		spec.Profile.Env = maps.Clone(spec.Profile.Env)
+		spec.Profile.Env[engines.SharedModelsEnv] = string(b)
+	}
 	// Adapters are fetched like the model, and before the engine starts: it
 	// loads them at start, and one missing is a replica that answers for the
 	// base model under the adapter's name.
 	for i, ad := range spec.Profile.Adapters {
-		path, err := a.weights().Ensure(ctx, ad.Path)
+		path, err := a.weights().Ensure(ctx, ad.Path, spec.Profile.AllowPickle)
 		if err != nil {
 			report(supervisor.Failed, fmt.Sprintf("adapter %s: %v", ad.Name, err), nil)
 			return
@@ -1365,11 +1438,14 @@ func Client(cfg tlsconf.Config, timeout time.Duration) (*http.Client, error) {
 // off — is never closed: nothing sends the FIN, and the agent's heartbeats sit
 // in the socket's buffer until the kernel gives up retransmitting, fifteen
 // minutes later. Found by switching off a leader: the standby took over in 14
-// seconds and the node stayed with the dead one. An HTTP/2 ping after ten
-// quiet seconds, given five to answer, finds it in a quarter of a minute.
+// seconds and the node stayed with the dead one. An HTTP/2 ping after five
+// quiet seconds, given three to answer, finds it in eight: by the time the
+// standby leads, as it does within the lease's ten seconds, the node is on
+// its way to it. At fifteen, with the lease at fifteen too, the new leader
+// went without its nodes, and so its judges, for half a minute.
 const (
-	streamIdle = 10 * time.Second
-	streamPing = 5 * time.Second
+	streamIdle = 5 * time.Second
+	streamPing = 3 * time.Second
 )
 
 // localizeSpec fills in what only this node knows.
@@ -1450,18 +1526,25 @@ func ResolveManagement(c Config) (addr, iface string) {
 // shard computes on as soon as that is known: an adopted engine's is found
 // after registration, which carried what was known then.
 func (a *Agent) reportCompute(ctx context.Context, out chan<- *llmgwv1.AgentMessage) {
-	sent := map[string]bool{}
+	// What was sent for each shard, so a compute that changes is sent again:
+	// an adopted shard's kept one is sent at once, and adopting it may then
+	// find it was kept wrong.
+	sent := map[string]string{}
 	t := time.NewTicker(5 * time.Second)
 	defer t.Stop()
 	for range 60 {
 		for _, st := range a.sup.Statuses() {
-			if st.State.Terminal() || sent[st.ID] {
+			if st.State.Terminal() {
 				continue
 			}
 			c := a.sup.ComputeOf(st.ID)
 			if c == nil {
 				continue
 			}
+			said := fmt.Sprint(*c)
+			if sent[st.ID] == said {
+				continue
+			}
 			sh, err := a.sup.Get(st.ID)
 			if err != nil {
 				continue
@@ -1472,7 +1555,7 @@ func (a *Agent) reportCompute(ctx context.Context, out chan<- *llmgwv1.AgentMess
 			}}}
 			select {
 			case out <- msg:
-				sent[st.ID] = true
+				sent[st.ID] = said
 			case <-ctx.Done():
 				return
 			}
@@ -1484,3 +1567,51 @@ func (a *Agent) reportCompute(ctx context.Context, out chan<- *llmgwv1.AgentMess
 		}
 	}
 }
+
+// memBandwidth is the node's memory read rate in GB/s, 0 until measured.
+// Measured in the background, the first time it is asked for: it reads half
+// a gigabyte, which must not hold up an agent reaching its control plane,
+// and the heartbeat carries it when it is known. Not at once but after a
+// minute or so, chosen at random, and again every hour, keeping the best:
+// VMs on one host that all restart together measure each other, and read a
+// third less than any of them gets alone. Once a process rather than an
+// agent, so tests running many agents in one read it once.
+func (a *Agent) memBandwidth() discovery.Bandwidth {
+	bwOnce.Do(func() {
+		go func() {
+			time.Sleep(bandwidthDelay())
+			for {
+				// The best of each seen so far: an hourly pass on a node
+				// busy serving reads less, and that is not the memory.
+				bw := measureBandwidth()
+				cur := a.memBandwidthNow()
+				if bw.All > cur.All || bw.Core > cur.Core {
+					bwBits.Store(math.Float64bits(max(bw.All, cur.All)))
+					coreBits.Store(math.Float64bits(max(bw.Core, cur.Core)))
+					a.log.Info("this node's memory", "read_gbs", math.Round(bw.All*10)/10,
+						"one_core_gbs", math.Round(bw.Core*10)/10)
+				}
+				time.Sleep(time.Hour)
+			}
+		}()
+	})
+	return a.memBandwidthNow()
+}
+
+// memBandwidthNow is what has been measured so far, without starting a
+// measurement.
+func (a *Agent) memBandwidthNow() discovery.Bandwidth {
+	return discovery.Bandwidth{All: math.Float64frombits(bwBits.Load()), Core: math.Float64frombits(coreBits.Load())}
+}
+
+// bandwidthDelay is how long an agent waits before measuring, so agents
+// started together measure apart.
+var bandwidthDelay = func() time.Duration { return 10*time.Second + rand.N(110*time.Second) }
+
+var (
+	bwOnce   sync.Once
+	bwBits   atomic.Uint64
+	coreBits atomic.Uint64
+	// measureBandwidth is replaced in tests, which have no use for it.
+	measureBandwidth = discovery.MeasureMemoryBandwidth
+)

internal/agent/agent_test.go

@@ -68,7 +68,7 @@ func (r *statusRecorder) OnShardCompute(replicaID, shardID string, c reconciler.
 	r.computes = append(r.computes, c)
 }
 
-func (r *statusRecorder) OnFetchProgress(replicaID, shardID, source string, done, total int64) {
+func (r *statusRecorder) OnFetchProgress(replicaID, shardID, source, phase string, done, total int64) {
 	r.mu.Lock()
 	defer r.mu.Unlock()
 	r.fetches = append(r.fetches, fetchReport{replicaID, shardID, source, done, total})
@@ -671,7 +671,7 @@ func TestAShardWithMissingWeightsFailsBeforeLaunching(t *testing.T) {
 	// nobody is watching — so it is caught here, where the message can say
 	// what is actually wrong.
 	a := &Agent{cfg: Config{WeightsCache: t.TempDir()}, log: slog.New(slog.DiscardHandler)}
-	_, err := a.weights().Ensure(t.Context(), filepath.Join(t.TempDir(), "not-there.gguf"))
+	_, err := a.weights().Ensure(t.Context(), filepath.Join(t.TempDir(), "not-there.gguf"), false)
 	if err == nil {
 		t.Fatal("a missing weights file was accepted")
 	}

internal/agent/config.go

@@ -103,6 +103,11 @@ type Config struct {
 	Tracing obs.Tracing `koanf:"tracing"`
 
 	Version string
+
+	// Settings is the cluster's settings for this agent it was started on,
+	// as the control plane sent them: what a newly sent version is
+	// compared with.
+	Settings string `koanf:"-"`
 }
 
 // DefaultConfig returns sensible defaults.

internal/agent/discovery/cudatoolkit.go

@@ -0,0 +1,93 @@
+package discovery
+
+import (
+	"encoding/json"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"regexp"
+	"slices"
+	"strconv"
+	"strings"
+)
+
+// CUDAToolkit is the root of an installed CUDA toolkit of this major version
+// — the directory whose bin holds nvcc — or "" when there is none. Engines
+// that compile their kernels where they run (FreeToken, ComfyUI's compiled
+// nodes) need one matching the CUDA their torch was built for, and a
+// toolkit of another major version compiles code that torch cannot load.
+//
+// CUDA_HOME is looked at first, then /usr/local/cuda, then every
+// /usr/local/cuda-<version>, newest first.
+func CUDAToolkit(major int) string {
+	return cudaToolkitIn(major, os.Getenv("CUDA_HOME"), "/usr/local")
+}
+
+func cudaToolkitIn(major int, home, root string) string {
+	var dirs []string
+	if home != "" {
+		dirs = append(dirs, home)
+	}
+	dirs = append(dirs, filepath.Join(root, "cuda"))
+	versioned, _ := filepath.Glob(filepath.Join(root, "cuda-*"))
+	slices.SortFunc(versioned, func(a, b string) int { return versionCompare(b, a) })
+	dirs = append(dirs, versioned...)
+	for _, d := range dirs {
+		if _, err := os.Stat(filepath.Join(d, "bin", "nvcc")); err != nil {
+			continue
+		}
+		if toolkitMajor(d) == major {
+			return d
+		}
+	}
+	return ""
+}
+
+// toolkitMajor is a toolkit's major version, from the version.json every
+// toolkit since 11.1 carries, or from nvcc where it has none; 0 when neither
+// says.
+func toolkitMajor(dir string) int {
+	if b, err := os.ReadFile(filepath.Join(dir, "version.json")); err == nil { //nolint:gosec // a toolkit's own file
+		var v struct {
+			CUDA struct {
+				Version string `json:"version"`
+			} `json:"cuda"`
+		}
+		if json.Unmarshal(b, &v) == nil && v.CUDA.Version != "" {
+			n, _ := strconv.Atoi(strings.SplitN(v.CUDA.Version, ".", 2)[0])
+			return n
+		}
+	}
+	//nolint:gosec // nvcc of a toolkit found on disk
+	out, err := exec.Command(filepath.Join(dir, "bin", "nvcc"), "--version").Output()
+	if err != nil {
+		return 0
+	}
+	if m := nvccRelease.FindSubmatch(out); m != nil {
+		n, _ := strconv.Atoi(string(m[1]))
+		return n
+	}
+	return 0
+}
+
+var nvccRelease = regexp.MustCompile(`release (\d+)\.`)
+
+// versionCompare orders two paths ending in dotted versions: cuda-13.10 is
+// newer than cuda-13.4, which comparing them as text gets wrong.
+func versionCompare(a, b string) int {
+	pa := strings.Split(strings.TrimPrefix(filepath.Base(a), "cuda-"), ".")
+	pb := strings.Split(strings.TrimPrefix(filepath.Base(b), "cuda-"), ".")
+	for i := range max(len(pa), len(pb)) {
+		var x, y int
+		if i < len(pa) {
+			x, _ = strconv.Atoi(pa[i])
+		}
+		if i < len(pb) {
+			y, _ = strconv.Atoi(pb[i])
+		}
+		if x != y {
+			return x - y
+		}
+	}
+	return 0
+}

internal/agent/discovery/cudatoolkit_test.go

@@ -0,0 +1,48 @@
+package discovery
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func toolkit(t *testing.T, root, name, version string) string {
+	t.Helper()
+	d := filepath.Join(root, name)
+	put(t, d, "bin/nvcc", "#!/bin/sh")
+	if version != "" {
+		put(t, d, "version.json", `{"cuda":{"name":"CUDA SDK","version":"`+version+`"}}`)
+	}
+	return d
+}
+
+func TestTheNewestToolkitOfTheMajorVersionAskedForIsFound(t *testing.T) {
+	root := t.TempDir()
+	toolkit(t, root, "cuda-12.8", "12.8.1")
+	toolkit(t, root, "cuda-13.4", "13.4.0")
+	want := toolkit(t, root, "cuda-13.10", "13.10.0")
+	// /usr/local/cuda is a 12: a toolkit of the wrong major version is not
+	// one torch built for 13 can load what it compiles.
+	toolkit(t, root, "cuda", "12.8.1")
+	if got := cudaToolkitIn(13, "", root); got != want {
+		t.Errorf("got %q, want %q", got, want)
+	}
+	if got := cudaToolkitIn(12, "", root); got != filepath.Join(root, "cuda") {
+		t.Errorf("12: got %q", got)
+	}
+	// CUDA_HOME first, when it is one.
+	home := toolkit(t, t.TempDir(), "mine", "13.0.2")
+	if got := cudaToolkitIn(13, home, root); got != home {
+		t.Errorf("CUDA_HOME: got %q", got)
+	}
+	if got := cudaToolkitIn(14, "", root); got != "" {
+		t.Errorf("14: got %q", got)
+	}
+	// A directory without nvcc is not a toolkit.
+	if err := os.MkdirAll(filepath.Join(root, "cuda-13.99"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if got := cudaToolkitIn(13, "", root); got != want {
+		t.Errorf("with an empty cuda-13.99: got %q", got)
+	}
+}

internal/agent/discovery/discover.go

@@ -28,6 +28,11 @@ type Options struct {
 	// SkipGPUs omits GPU discovery. A node without nvidia-smi is already
 	// found to have none, so this is for one whose driver is broken.
 	SkipGPUs bool `koanf:"skip_gpus"`
+	// DCGMExporter is the dcgm-exporter metrics URL read for how much of
+	// each card's chip is in use, on cards NVML's GPM does not cover (before
+	// Hopper). Empty is http://127.0.0.1:9400/metrics; "off" reads none. A
+	// node without one costs a refused connection every five minutes.
+	DCGMExporter string `koanf:"dcgm_exporter"`
 	// ModelDirs are scanned for weights. Empty means do not look.
 	ModelDirs []string `koanf:"-"`
 }

internal/agent/discovery/gpm_nonvml.go

@@ -0,0 +1,6 @@
+//go:build !nvml
+
+package discovery
+
+// newGPM has nothing to read GPM with in a build without the nvml tag.
+func newGPM() gpmSource { return nil }

internal/agent/discovery/gpm_nvml.go

@@ -0,0 +1,124 @@
+//go:build nvml
+
+package discovery
+
+import (
+	"sync"
+
+	"github.com/NVIDIA/go-nvml/pkg/nvml"
+)
+
+// nvmlGPM reads GPM, NVML's own profiling counters, on the cards that have
+// them: Hopper and later. Each read takes a sample and works out the metrics
+// against the one before, so NVML stays initialised between reads.
+type nvmlGPM struct {
+	mu     sync.Mutex
+	opened bool
+	failed bool
+	cards  map[int]*gpmCard
+}
+
+type gpmCard struct {
+	dev          nvml.Device
+	prev, cur    nvml.GpmSample
+	havePrevious bool
+}
+
+func newGPM() gpmSource { return &nvmlGPM{} }
+
+// The GPM metrics read, in the order they are asked for.
+var gpmMetrics = []struct {
+	id  nvml.GpmMetricId
+	set func(*GPUProfile, float64)
+}{
+	{nvml.GPM_METRIC_SM_UTIL, func(p *GPUProfile, v float64) { p.SMActive = v }},
+	{nvml.GPM_METRIC_SM_OCCUPANCY, func(p *GPUProfile, v float64) { p.SMOccupancy = v }},
+	{nvml.GPM_METRIC_ANY_TENSOR_UTIL, func(p *GPUProfile, v float64) { p.TensorActive = v }},
+	{nvml.GPM_METRIC_DRAM_BW_UTIL, func(p *GPUProfile, v float64) { p.DRAMActive = v }},
+}
+
+func (g *nvmlGPM) read() map[int]GPUProfile {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	if g.failed {
+		return nil
+	}
+	if !g.opened && !g.open() {
+		g.failed = true
+		return nil
+	}
+	out := map[int]GPUProfile{}
+	for idx, c := range g.cards {
+		if c.cur.Get(c.dev) != nvml.SUCCESS {
+			// A card in MIG mode, or one that has gone: no sample, and the
+			// next must not be measured against an old one.
+			c.havePrevious = false
+			continue
+		}
+		if c.havePrevious {
+			if p, ok := gpmBetween(c.prev, c.cur); ok {
+				out[idx] = p
+			}
+		}
+		c.prev, c.cur = c.cur, c.prev
+		c.havePrevious = true
+	}
+	return out
+}
+
+// open finds the cards with GPM. False is NVML not starting, or no card
+// having it, and the reader then never tries again.
+func (g *nvmlGPM) open() bool {
+	if nvml.Init() != nvml.SUCCESS {
+		return false
+	}
+	g.cards = map[int]*gpmCard{}
+	n, ret := nvml.DeviceGetCount()
+	if ret != nvml.SUCCESS {
+		n = 0
+	}
+	for i := range n {
+		dev, ret := nvml.DeviceGetHandleByIndex(i)
+		if ret != nvml.SUCCESS {
+			continue
+		}
+		sup, ret := dev.GpmQueryDeviceSupport()
+		if ret != nvml.SUCCESS || sup.IsSupportedDevice == 0 {
+			continue
+		}
+		a, ret := nvml.GpmSampleAlloc()
+		if ret != nvml.SUCCESS {
+			continue
+		}
+		b, ret := nvml.GpmSampleAlloc()
+		if ret != nvml.SUCCESS {
+			_ = a.Free()
+			continue
+		}
+		g.cards[i] = &gpmCard{dev: dev, prev: a, cur: b}
+	}
+	if len(g.cards) == 0 {
+		_ = nvml.Shutdown()
+		return false
+	}
+	g.opened = true
+	return true
+}
+
+// gpmBetween is the profile over the time between two samples.
+func gpmBetween(earlier, later nvml.GpmSample) (GPUProfile, bool) {
+	q := nvml.GpmMetricsGetType{NumMetrics: uint32(len(gpmMetrics)), Sample1: earlier, Sample2: later}
+	for i, m := range gpmMetrics {
+		q.Metrics[i].MetricId = uint32(m.id)
+	}
+	if nvml.GpmMetricsGet(&q) != nvml.SUCCESS {
+		return GPUProfile{}, false
+	}
+	p := GPUProfile{SMActive: -1, SMOccupancy: -1, TensorActive: -1, DRAMActive: -1, Source: "gpm"}
+	for i, m := range gpmMetrics {
+		if r := q.Metrics[i]; r.NvmlReturn == uint32(nvml.SUCCESS) && r.Value >= 0 {
+			m.set(&p, r.Value)
+		}
+	}
+	return p, true
+}

internal/agent/discovery/gpu.go

@@ -52,6 +52,8 @@ type GPU struct {
 	// memory the instance's share. Utilisation stays unreported, because no
 	// driver measures it per instance.
 	MIG *MIGDevice `json:"mig,omitempty"`
+	// PCIe is how the card is linked to its host, zero when not read.
+	PCIe PCIeLink `json:"pcie,omitempty"`
 }
 
 // GPUHealth records conditions that make a GPU unsafe to schedule onto.

internal/agent/discovery/gpu_nvml.go

@@ -67,10 +67,23 @@ func DiscoverGPUsNVML(_ context.Context) ([]GPU, error) {
 		if major, minor, ret := dev.GetCudaComputeCapability(); ret == nvml.SUCCESS {
 			g.ComputeCapability = fmt.Sprintf("%d.%d", major, minor)
 		}
+		if v, ret := dev.GetMaxPcieLinkGeneration(); ret == nvml.SUCCESS {
+			g.PCIe.Gen = v
+		}
+		if v, ret := dev.GetGpuMaxPcieLinkGeneration(); ret == nvml.SUCCESS {
+			g.PCIe.GPUGen = v
+		}
+		if v, ret := dev.GetCurrPcieLinkWidth(); ret == nvml.SUCCESS {
+			g.PCIe.Width = v
+		}
+		if v, ret := dev.GetMaxPcieLinkWidth(); ret == nvml.SUCCESS {
+			g.PCIe.WidthMax = v
+		}
 		if pci, ret := dev.GetPciInfo(); ret == nvml.SUCCESS {
 			g.PCIBusID = busIDString(pci.BusId[:])
 			byBusID[normalizeBusID(g.PCIBusID)] = i
 		}
+		g.PCIe = withSysfs(g.PCIe, "/sys/bus/pci/devices", g.PCIBusID)
 		if t, ret := dev.GetTemperature(nvml.TEMPERATURE_GPU); ret == nvml.SUCCESS {
 			g.Health.TemperatureC = int(t)
 		}

internal/agent/discovery/gpu_smi.go

@@ -64,6 +64,7 @@ func DiscoverGPUsSMI(ctx context.Context, binary string) ([]GPU, error) {
 	if err != nil || len(cards) == 0 {
 		return cards, err
 	}
+	readPCIe(ctx, binary, cards)
 	// The query lists whole cards, MIG mode or not. A card in MIG mode is
 	// replaced by its instances, which only -L names and only the full
 	// report sizes; asked only when some card could be in MIG mode, since

internal/agent/discovery/host.go

@@ -3,6 +3,7 @@ package discovery
 import (
 	"bufio"
 	"os"
+	"regexp"
 	"runtime"
 	"strconv"
 	"strings"
@@ -30,6 +31,14 @@ type Host struct {
 	// MinComputeCapability is the oldest GPU's, 0 when there are none or it
 	// is not known. Not read here but from the GPU inventory, by the agent.
 	MinComputeCapability float64 `json:"min_compute_capability,omitempty"`
+	// MemBandwidthGBs is how fast the memory reads, in GB/s, 0 when not
+	// measured: what decides a model's speed on the CPU. Not read here but
+	// measured by the agent when it starts (MeasureMemoryBandwidth).
+	MemBandwidthGBs float64 `json:"mem_bandwidth_gbs,omitempty"`
+	// MemBandwidthCoreGBs is what one core reads alone, in GB/s, 0 when not
+	// measured: a replica given n threads reads at about n times this, up to
+	// MemBandwidthGBs.
+	MemBandwidthCoreGBs float64 `json:"mem_bandwidth_core_gbs,omitempty"`
 }
 
 // Family is the packaging family: "deb", "rpm", or "" when neither is known.
@@ -142,17 +151,29 @@ func cpuFlagsFrom(path string) []string {
 // nvidiaDriverFrom reads the loaded kernel module's version, which exists
 // exactly when a driver is loaded — unlike nvidia-smi, which can be installed
 // on a machine whose module failed to load.
+var driverVersion = regexp.MustCompile(`^[0-9]+\.[0-9]+(\.[0-9]+)?$`)
+
 func nvidiaDriverFrom(path string) string {
 	b, err := os.ReadFile(path) //nolint:gosec // a fixed system path
 	if err != nil {
 		return ""
 	}
-	// "NVRM version: NVIDIA UNIX x86_64 Kernel Module  570.86.15  Thu ..."
+	// "NVRM version: NVIDIA UNIX x86_64 Kernel Module  570.86.15  Thu ...", or
+	// for the open module, the default since 560, "NVIDIA UNIX Open Kernel
+	// Module for x86_64  580.82.07  Release Build ...": the first version
+	// after "Module", not the word after it. Read as "for", the driver was
+	// unknown, and every build was tried whatever driver it needs.
 	fields := strings.Fields(string(b))
 	for i, f := range fields {
-		if f == "Module" && i+1 < len(fields) {
-			return fields[i+1]
+		if f != "Module" {
+			continue
 		}
+		for _, v := range fields[i+1:] {
+			if driverVersion.MatchString(v) {
+				return v
+			}
+		}
+		return ""
 	}
 	return ""
 }

internal/agent/discovery/host_driver_test.go

@@ -0,0 +1,25 @@
+package discovery
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestNVIDIADriverFrom(t *testing.T) {
+	for _, tc := range []struct{ text, want string }{
+		{"NVRM version: NVIDIA UNIX x86_64 Kernel Module  570.86.15  Thu Jan 23 21:30:38 UTC 2025\nGCC version:  gcc version 12.2.0\n", "570.86.15"},
+		{"NVRM version: NVIDIA UNIX Open Kernel Module for x86_64  580.82.07  Release Build  (dvs-builder@U16-I3-B03-4-3)  Wed Aug 27 00:51:12 UTC 2025\n", "580.82.07"},
+		{"NVRM version: NVIDIA UNIX Open Kernel Module for aarch64  580.95.05  Release Build\n", "580.95.05"},
+		{"NVRM version: NVIDIA UNIX x86_64 Kernel Module  535.54  Mon\n", "535.54"},
+		{"something else\n", ""},
+	} {
+		p := filepath.Join(t.TempDir(), "version")
+		if err := os.WriteFile(p, []byte(tc.text), 0o600); err != nil {
+			t.Fatal(err)
+		}
+		if got := nvidiaDriverFrom(p); got != tc.want {
+			t.Errorf("%q: got %q, want %q", tc.text, got, tc.want)
+		}
+	}
+}

internal/agent/discovery/membw.go

@@ -0,0 +1,169 @@
+package discovery
+
+import (
+	"bufio"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+)
+
+// How fast this machine reads its memory, which on a CPU is what decides how
+// fast a model answers: every token of a reply reads all the weights it uses
+// once, so a node that reads 40 GB/s writes a 4 GB model's tokens about ten
+// times a second, whatever its core count. Measured once when the agent
+// starts, in a fraction of a second, by every CPU summing a share of a buffer
+// too large for any cache — a VM sharing its host's memory with others gets
+// its share, which is the number that matters for it.
+
+// bandwidthBuffer is what is read at most; the least is eight times the
+// processor's last cache, and never under minBuffer. Smaller, it would be
+// the cache that was measured — some hold a gigabyte — and the figure kept
+// would be too high.
+const (
+	bandwidthBuffer = 2 << 30
+	minBuffer       = 256 << 20
+)
+
+// Bandwidth is how fast this machine's memory reads, in GB/s: with every
+// CPU reading, and with one alone. A replica given a few threads of a large
+// machine reads at a few times the second, not at the first: on a server of
+// many memory channels one core cannot keep more than a fraction of them
+// busy.
+type Bandwidth struct {
+	All, Core float64
+}
+
+// coreBuffer is what the one-core passes read: past any cache, and short
+// enough that one core reads it in a fraction of a second.
+const coreBuffer = 512 << 20
+
+// MeasureMemoryBandwidth is the memory's read rate, the best of a few passes
+// each way; zero when the machine has too little free memory to measure.
+func MeasureMemoryBandwidth() Bandwidth {
+	size := int64(bandwidthBuffer)
+	// Never more than a quarter of what is free: a node already holding
+	// models must not be pushed into swap by a measurement. Without room
+	// for a buffer past the cache, none is measured: the hourly retry
+	// measures it when there is.
+	if avail := memAvailable("/proc/meminfo"); avail > 0 {
+		size = min(size, avail/4)
+	}
+	if size < max(minBuffer, 8*lastCacheBytes("/sys/devices/system/cpu")) {
+		return Bandwidth{}
+	}
+	buf := make([]uint64, size/8)
+	for i := range buf {
+		buf[i] = uint64(i)
+	}
+	workers := runtime.NumCPU()
+	chunk := (len(buf) + workers - 1) / workers
+	best := 0.0
+	for pass := 0; pass < 4; pass++ {
+		var wg sync.WaitGroup
+		start := time.Now()
+		for w := 0; w < workers; w++ {
+			lo, hi := w*chunk, min((w+1)*chunk, len(buf))
+			if lo >= hi {
+				continue
+			}
+			wg.Add(1)
+			go func(part []uint64) {
+				defer wg.Done()
+				sink(sum(part))
+			}(buf[lo:hi])
+		}
+		wg.Wait()
+		if s := time.Since(start).Seconds(); s > 0 {
+			best = max(best, float64(size)/s/1e9)
+		}
+	}
+	// And one core alone, over a part of the same buffer.
+	part := buf[:min(len(buf), coreBuffer/8)]
+	core := 0.0
+	for pass := 0; pass < 3; pass++ {
+		start := time.Now()
+		sink(sum(part))
+		if s := time.Since(start).Seconds(); s > 0 {
+			core = max(core, float64(len(part)*8)/s/1e9)
+		}
+	}
+	return Bandwidth{All: best, Core: min(core, best)}
+}
+
+// sum reads every word, eight at a time so the loop is not what limits it.
+func sum(p []uint64) uint64 {
+	var a, b, c, d, e, f, g, h uint64
+	i := 0
+	for ; i+8 <= len(p); i += 8 {
+		a += p[i]
+		b += p[i+1]
+		c += p[i+2]
+		d += p[i+3]
+		e += p[i+4]
+		f += p[i+5]
+		g += p[i+6]
+		h += p[i+7]
+	}
+	for ; i < len(p); i++ {
+		a += p[i]
+	}
+	return a + b + c + d + e + f + g + h
+}
+
+// sinkValue keeps the sums alive, so the reads are not optimised away.
+var sinkValue struct {
+	sync.Mutex
+	v uint64
+}
+
+func sink(v uint64) {
+	sinkValue.Lock()
+	sinkValue.v += v
+	sinkValue.Unlock()
+}
+
+// memAvailable is MemAvailable in bytes, 0 when unknown.
+func memAvailable(path string) int64 {
+	f, err := os.Open(path) //nolint:gosec // a fixed system path
+	if err != nil {
+		return 0
+	}
+	defer f.Close()
+	sc := bufio.NewScanner(f)
+	for sc.Scan() {
+		fields := strings.Fields(sc.Text())
+		if len(fields) >= 2 && fields[0] == "MemAvailable:" {
+			kb, _ := strconv.ParseInt(fields[1], 10, 64)
+			return kb << 10
+		}
+	}
+	return 0
+}
+
+// lastCacheBytes is the largest cache any CPU reports, 0 when unknown.
+func lastCacheBytes(root string) int64 {
+	paths, _ := filepath.Glob(filepath.Join(root, "cpu*", "cache", "index*", "size"))
+	var most int64
+	for _, p := range paths {
+		b, err := os.ReadFile(p) //nolint:gosec // under /sys
+		if err != nil {
+			continue
+		}
+		v := strings.TrimSpace(string(b))
+		mult := int64(1)
+		switch {
+		case strings.HasSuffix(v, "K"):
+			mult, v = 1<<10, strings.TrimSuffix(v, "K")
+		case strings.HasSuffix(v, "M"):
+			mult, v = 1<<20, strings.TrimSuffix(v, "M")
+		}
+		if n, err := strconv.ParseInt(v, 10, 64); err == nil {
+			most = max(most, n*mult)
+		}
+	}
+	return most
+}

internal/agent/discovery/membw_test.go

@@ -0,0 +1,32 @@
+package discovery
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestMemoryBandwidthIsMeasured(t *testing.T) {
+	if testing.Short() {
+		t.Skip("reads half a gigabyte")
+	}
+	bw := MeasureMemoryBandwidth()
+	// Any machine running this reads more than a gigabyte a second and less
+	// than a few terabytes, and one core no more than all of them.
+	if bw.All < 1 || bw.All > 5000 || bw.Core < 0.5 || bw.Core > bw.All {
+		t.Errorf("measured %.1f GB/s, %.1f on one core", bw.All, bw.Core)
+	}
+	t.Logf("%.1f GB/s, %.1f on one core", bw.All, bw.Core)
+}
+
+func TestTheLastCacheIsReadFromSys(t *testing.T) {
+	root := t.TempDir()
+	for path, size := range map[string]string{"cpu0/cache/index0/size": "48K", "cpu0/cache/index3/size": "96M"} {
+		p := filepath.Join(root, path)
+		_ = os.MkdirAll(filepath.Dir(p), 0o755)
+		_ = os.WriteFile(p, []byte(size+"\n"), 0o644)
+	}
+	if got := lastCacheBytes(root); got != 96<<20 {
+		t.Errorf("%d", got)
+	}
+}

internal/agent/discovery/pcie.go

@@ -0,0 +1,181 @@
+package discovery
+
+import (
+	"context"
+	"encoding/csv"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strconv"
+	"strings"
+)
+
+// A card in the wrong slot: an x16 card in a slot wired x8 or x4, behind a
+// riser, or in a slot of an older PCIe generation than the card's, links
+// slower than it was built to, and nothing says so. A model on the card
+// still decodes at the card's own memory speed, but it loads slower, a
+// tensor-parallel replica over PCIe is slower, and so is anything that
+// spills to the host. The driver says how the link stands and what the
+// card could do; the difference is reported where the node's warnings are.
+
+// PCIeLink is how a card is linked to its host.
+type PCIeLink struct {
+	// Gen is the fastest generation the card and the slot together allow —
+	// not the current one, which drops while the card idles, to save power.
+	// GPUGen is the card's own fastest, 0 when the driver does not say.
+	Gen    int `json:"gen,omitempty"`
+	GPUGen int `json:"gpu_gen,omitempty"`
+	// Width is the lanes in use now, WidthMax the card's most.
+	Width    int `json:"width,omitempty"`
+	WidthMax int `json:"width_max,omitempty"`
+}
+
+// laneGBs is one lane's usable rate in GB/s, by generation.
+var laneGBs = map[int]float64{1: 0.25, 2: 0.5, 3: 0.985, 4: 1.969, 5: 3.938, 6: 7.877}
+
+// GBs is what a link of gen and width carries, in GB/s each way.
+func pcieGBs(gen, width int) float64 { return laneGBs[gen] * float64(width) }
+
+// Problem says how a card's link falls short of what the card could do,
+// empty when it does not or the driver did not say.
+func (p PCIeLink) Problem() string {
+	gen := p.Gen
+	if p.GPUGen > 0 && gen == 0 {
+		gen = p.GPUGen
+	}
+	slowGen := p.GPUGen > 0 && p.Gen > 0 && p.Gen < p.GPUGen
+	narrow := p.WidthMax > 0 && p.Width > 0 && p.Width < p.WidthMax
+	if !slowGen && !narrow {
+		return ""
+	}
+	could := pcieGBs(max(p.GPUGen, gen), p.WidthMax)
+	has := pcieGBs(gen, p.Width)
+	var why []string
+	if narrow {
+		why = append(why, fmt.Sprintf("x%d of its x%d lanes", p.Width, p.WidthMax))
+	}
+	if slowGen {
+		why = append(why, fmt.Sprintf("PCIe Gen%d of its Gen%d", p.Gen, p.GPUGen))
+	}
+	return fmt.Sprintf("linked at %s: %.1f GB/s where it could have %.1f. "+
+		"Check the slot (wired fewer lanes, or an older generation), a riser, or the BIOS's PCIe setting",
+		strings.Join(why, " and "), has, could)
+}
+
+// pcieFields are asked apart from the rest, and the card's own fastest
+// generation apart again: a driver that does not know a field refuses the
+// whole query, and these must not cost a node its cards.
+var pcieFields = [][]string{
+	{"index", "pcie.link.gen.max", "pcie.link.width.current", "pcie.link.width.max", "pcie.link.gen.gpumax"},
+	{"index", "pcie.link.gen.max", "pcie.link.width.current", "pcie.link.width.max"},
+}
+
+// readPCIe fills each card's link from nvidia-smi, leaving it empty when
+// the driver says nothing of it.
+func readPCIe(ctx context.Context, binary string, cards []GPU) {
+	for _, fields := range pcieFields {
+		out, err := exec.CommandContext(ctx, binary, "--query-gpu="+strings.Join(fields, ","),
+			"--format=csv,noheader,nounits").Output()
+		if err != nil {
+			continue
+		}
+		links := parsePCIe(string(out))
+		for i := range cards {
+			if l, ok := links[cards[i].Index]; ok && cards[i].MIG == nil {
+				cards[i].PCIe = l
+			}
+		}
+		break
+	}
+	for i := range cards {
+		if cards[i].MIG == nil {
+			cards[i].PCIe = withSysfs(cards[i].PCIe, "/sys/bus/pci/devices", cards[i].PCIBusID)
+		}
+	}
+}
+
+// withSysfs takes the card's own lane count, and its own fastest
+// generation, from the kernel: the driver's width.max is what card and slot
+// allow together, so a x16 card in a slot wired x8 reads x8 of x8 there,
+// while the device's max_link_width is the card's, 16.
+func withSysfs(l PCIeLink, root, busID string) PCIeLink {
+	dev := sysfsBusID(busID)
+	if dev == "" {
+		return l
+	}
+	read := func(name string) string {
+		b, err := os.ReadFile(filepath.Join(root, dev, name)) //nolint:gosec // under /sys
+		if err != nil {
+			return ""
+		}
+		return strings.TrimSpace(string(b))
+	}
+	if n, err := strconv.Atoi(read("max_link_width")); err == nil && n > 0 {
+		l.WidthMax = n
+	}
+	if n, err := strconv.Atoi(read("current_link_width")); err == nil && n > 0 {
+		l.Width = n
+	}
+	if g := genOfSpeed(read("max_link_speed")); g > 0 && l.GPUGen == 0 {
+		l.GPUGen = g
+	}
+	return l
+}
+
+// sysfsBusID is nvidia-smi's 00000000:3B:00.0 as the kernel names it,
+// 0000:3b:00.0.
+func sysfsBusID(id string) string {
+	id = strings.ToLower(strings.TrimSpace(id))
+	dom, rest, ok := strings.Cut(id, ":")
+	if !ok || len(dom) < 4 {
+		return ""
+	}
+	return dom[len(dom)-4:] + ":" + rest
+}
+
+// genOfSpeed is the PCIe generation of a link speed as sysfs writes it.
+func genOfSpeed(s string) int {
+	switch {
+	case strings.HasPrefix(s, "2.5 GT/s"):
+		return 1
+	case strings.HasPrefix(s, "5.0 GT/s"), strings.HasPrefix(s, "5 GT/s"):
+		return 2
+	case strings.HasPrefix(s, "8.0 GT/s"), strings.HasPrefix(s, "8 GT/s"):
+		return 3
+	case strings.HasPrefix(s, "16.0 GT/s"), strings.HasPrefix(s, "16 GT/s"):
+		return 4
+	case strings.HasPrefix(s, "32.0 GT/s"), strings.HasPrefix(s, "32 GT/s"):
+		return 5
+	case strings.HasPrefix(s, "64.0 GT/s"), strings.HasPrefix(s, "64 GT/s"):
+		return 6
+	}
+	return 0
+}
+
+func parsePCIe(out string) map[int]PCIeLink {
+	r := csv.NewReader(strings.NewReader(out))
+	r.TrimLeadingSpace = true
+	r.FieldsPerRecord = -1
+	rows, _ := r.ReadAll()
+	num := func(s string) int {
+		n, _ := strconv.Atoi(strings.TrimSpace(s))
+		return n
+	}
+	links := map[int]PCIeLink{}
+	for _, row := range rows {
+		if len(row) < 4 {
+			continue
+		}
+		idx, err := strconv.Atoi(strings.TrimSpace(row[0]))
+		if err != nil {
+			continue
+		}
+		l := PCIeLink{Gen: num(row[1]), Width: num(row[2]), WidthMax: num(row[3])}
+		if len(row) > 4 {
+			l.GPUGen = num(row[4])
+		}
+		links[idx] = l
+	}
+	return links
+}

internal/agent/discovery/pcie_test.go

@@ -0,0 +1,54 @@
+package discovery
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+func TestACardInTheWrongSlotIsSaidSo(t *testing.T) {
+	for _, c := range []struct {
+		link PCIeLink
+		want string // empty: nothing to say
+	}{
+		// The PoC's Quadro RTX 6000, idle: current Gen1, but its fastest is 3,
+		// as the card's is, on all 16 lanes. Nothing wrong.
+		{PCIeLink{Gen: 3, GPUGen: 3, Width: 16, WidthMax: 16}, ""},
+		// An x16 card in a slot wired x8.
+		{PCIeLink{Gen: 4, GPUGen: 4, Width: 8, WidthMax: 16}, "x8 of its x16 lanes: 15.8 GB/s where it could have 31.5"},
+		// A Gen5 card in a Gen4 slot.
+		{PCIeLink{Gen: 4, GPUGen: 5, Width: 16, WidthMax: 16}, "PCIe Gen4 of its Gen5: 31.5 GB/s where it could have 63.0"},
+		// Both, behind a riser.
+		{PCIeLink{Gen: 3, GPUGen: 4, Width: 4, WidthMax: 16}, "x4 of its x16 lanes and PCIe Gen3 of its Gen4"},
+		// A driver that says nothing.
+		{PCIeLink{}, ""},
+	} {
+		got := c.link.Problem()
+		if (c.want == "") != (got == "") || !strings.Contains(got, c.want) {
+			t.Errorf("%+v: %q, want %q", c.link, got, c.want)
+		}
+	}
+}
+
+func TestTheLinkIsReadFromNvidiaSMI(t *testing.T) {
+	links := parsePCIe("0, 3, 16, 16, 3\n1, 4, 8, 16, [N/A]\n")
+	if links[0] != (PCIeLink{Gen: 3, GPUGen: 3, Width: 16, WidthMax: 16}) || links[1] != (PCIeLink{Gen: 4, Width: 8, WidthMax: 16}) {
+		t.Errorf("%+v", links)
+	}
+}
+
+func TestTheCardsOwnLanesComeFromTheKernel(t *testing.T) {
+	// The driver says x8 of x8 for a x16 card in a slot wired x8; the
+	// device's own max_link_width says 16.
+	root := t.TempDir()
+	dev := filepath.Join(root, "0000:3b:00.0")
+	_ = os.MkdirAll(dev, 0o755)
+	for f, v := range map[string]string{"max_link_width": "16", "current_link_width": "8", "max_link_speed": "16.0 GT/s PCIe"} {
+		_ = os.WriteFile(filepath.Join(dev, f), []byte(v+"\n"), 0o644)
+	}
+	l := withSysfs(PCIeLink{Gen: 4, Width: 8, WidthMax: 8}, root, "00000000:3B:00.0")
+	if l.WidthMax != 16 || l.Width != 8 || l.GPUGen != 4 || !strings.Contains(l.Problem(), "x8 of its x16") {
+		t.Errorf("%+v: %q", l, l.Problem())
+	}
+}

internal/agent/discovery/prof.go

@@ -0,0 +1,176 @@
+package discovery
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// GPUProfile is how much of a card's chip the work on it uses. nvidia-smi's
+// utilization.gpu cannot say: it is the share of time any kernel ran, and a
+// kernel keeping one SM of a hundred busy reads 100%. Measured on a Quadro
+// RTX 6000 under decode, it read 92% while the SMs had work 69% of the time
+// and issued an instruction on 18% of their cycles.
+//
+// These come from the card's profiling counters, which only two things read
+// without stopping the work: NVML's GPM, on Hopper and later, and DCGM, which
+// a site may run as dcgm-exporter. Percentages; -1 is not reported.
+type GPUProfile struct {
+	// SMActive is the share of time the SMs had work, averaged over all of
+	// them; SMOccupancy how full those SMs were with warps.
+	SMActive, SMOccupancy float64
+	// TensorActive is the tensor cores. Decode at batch one hardly uses them
+	// even when nothing is wrong: a quantised matrix-vector product runs on
+	// the integer units. Prefill and batches are where they show.
+	TensorActive float64
+	// DRAMActive is the memory interface, which decode is bound by.
+	DRAMActive float64
+	// Source is "gpm" or "dcgm".
+	Source string
+}
+
+// DefaultDCGMExporter is where dcgm-exporter listens unless told otherwise.
+const DefaultDCGMExporter = "http://127.0.0.1:9400/metrics"
+
+// dcgmRetry is how long a dcgm-exporter that did not answer is left alone.
+// Most nodes run none, and asking every ten seconds would be asking nobody.
+const dcgmRetry = 5 * time.Minute
+
+// gpmSource reads GPM, where this build and the card have it.
+type gpmSource interface {
+	read() map[int]GPUProfile
+}
+
+// ProfSampler reads each card's GPUProfile: from GPM where the card has it,
+// otherwise from a dcgm-exporter on the node.
+type ProfSampler struct {
+	gpm  gpmSource
+	dcgm string
+	get  func(ctx context.Context, url string) ([]byte, error)
+	now  func() time.Time
+
+	dcgmAfter time.Time
+}
+
+// NewProfSampler reads this machine. dcgmURL is the dcgm-exporter to ask:
+// empty is DefaultDCGMExporter, "off" asks none.
+func NewProfSampler(dcgmURL string) *ProfSampler {
+	switch dcgmURL {
+	case "":
+		dcgmURL = DefaultDCGMExporter
+	case "off":
+		dcgmURL = ""
+	}
+	return &ProfSampler{gpm: newGPM(), dcgm: dcgmURL, get: httpGet, now: time.Now}
+}
+
+// Sample is each card's profile by index; a card missing has none. GPM needs
+// two samples, so its cards appear from the second call.
+func (p *ProfSampler) Sample(ctx context.Context) map[int]GPUProfile {
+	out := map[int]GPUProfile{}
+	if p.gpm != nil {
+		for i, v := range p.gpm.read() {
+			out[i] = v
+		}
+	}
+	if p.dcgm == "" || p.now().Before(p.dcgmAfter) {
+		return out
+	}
+	body, err := p.get(ctx, p.dcgm)
+	if err != nil {
+		p.dcgmAfter = p.now().Add(dcgmRetry)
+		return out
+	}
+	for i, v := range parseDCGM(body) {
+		if _, ok := out[i]; !ok {
+			out[i] = v
+		}
+	}
+	return out
+}
+
+func httpGet(ctx context.Context, url string) ([]byte, error) {
+	ctx, cancel := context.WithTimeout(ctx, 2*time.Second)
+	defer cancel()
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
+	if err != nil {
+		return nil, err
+	}
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		return nil, err
+	}
+	defer func() { _ = resp.Body.Close() }()
+	if resp.StatusCode != http.StatusOK {
+		return nil, io.ErrUnexpectedEOF
+	}
+	return io.ReadAll(io.LimitReader(resp.Body, 4<<20))
+}
+
+// The dcgm-exporter fields read, as ratios of one. Its default field list
+// has them on every card DCGM can profile.
+var dcgmFields = map[string]func(*GPUProfile, float64){
+	"DCGM_FI_PROF_SM_ACTIVE":          func(p *GPUProfile, v float64) { p.SMActive = v },
+	"DCGM_FI_PROF_SM_OCCUPANCY":       func(p *GPUProfile, v float64) { p.SMOccupancy = v },
+	"DCGM_FI_PROF_PIPE_TENSOR_ACTIVE": func(p *GPUProfile, v float64) { p.TensorActive = v },
+	"DCGM_FI_PROF_DRAM_ACTIVE":        func(p *GPUProfile, v float64) { p.DRAMActive = v },
+}
+
+// parseDCGM reads dcgm-exporter's text exposition: a line per field and
+// card, the card in the gpu label. A card with none of the fields — DCGM
+// cannot profile it — is left out. A MIG instance's lines carry a
+// GPU_I_ID and are left out too: these are whole cards.
+func parseDCGM(body []byte) map[int]GPUProfile {
+	out := map[int]GPUProfile{}
+	for _, line := range strings.Split(string(body), "\n") {
+		if line == "" || line[0] == '#' {
+			continue
+		}
+		open := strings.IndexByte(line, '{')
+		end := strings.LastIndexByte(line, '}')
+		if open < 0 || end < open {
+			continue
+		}
+		set, ok := dcgmFields[line[:open]]
+		if !ok {
+			continue
+		}
+		labels := line[open+1 : end]
+		if strings.Contains(labels, `GPU_I_ID="`) && !strings.Contains(labels, `GPU_I_ID=""`) {
+			continue
+		}
+		idx, err := strconv.Atoi(labelValue(labels, "gpu"))
+		if err != nil {
+			continue
+		}
+		val := strings.Fields(line[end+1:])
+		if len(val) == 0 {
+			continue
+		}
+		v, err := strconv.ParseFloat(val[0], 64)
+		if err != nil || v < 0 {
+			continue
+		}
+		p, seen := out[idx]
+		if !seen {
+			p = GPUProfile{SMActive: -1, SMOccupancy: -1, TensorActive: -1, DRAMActive: -1, Source: "dcgm"}
+		}
+		set(&p, 100*v)
+		out[idx] = p
+	}
+	return out
+}
+
+// labelValue is one label's value from a Prometheus label set.
+func labelValue(labels, name string) string {
+	for _, kv := range strings.Split(labels, ",") {
+		k, v, ok := strings.Cut(strings.TrimSpace(kv), "=")
+		if ok && k == name {
+			return strings.Trim(v, `"`)
+		}
+	}
+	return ""
+}

internal/agent/discovery/prof_test.go

@@ -0,0 +1,107 @@
+package discovery
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+)
+
+// dcgmBody is dcgm-exporter's output for two cards and a MIG instance:
+// card 0 profiled, card 1 a model DCGM cannot profile, which has the device
+// fields and none of the PROF ones.
+const dcgmBody = `# HELP DCGM_FI_DEV_GPU_UTIL GPU utilization (in %).
+# TYPE DCGM_FI_DEV_GPU_UTIL gauge
+DCGM_FI_DEV_GPU_UTIL{gpu="0",UUID="GPU-a",device="nvidia0",modelName="NVIDIA L40S",Hostname="n1"} 100
+DCGM_FI_DEV_GPU_UTIL{gpu="1",UUID="GPU-b",device="nvidia1",modelName="Quadro RTX 6000",Hostname="n1"} 92
+# HELP DCGM_FI_PROF_SM_ACTIVE The ratio of cycles an SM has at least 1 warp assigned.
+# TYPE DCGM_FI_PROF_SM_ACTIVE gauge
+DCGM_FI_PROF_SM_ACTIVE{gpu="0",UUID="GPU-a",device="nvidia0",modelName="NVIDIA L40S",Hostname="n1"} 0.007042
+DCGM_FI_PROF_SM_ACTIVE{gpu="2",UUID="GPU-c",device="nvidia2",modelName="NVIDIA A100",GPU_I_PROFILE="3g.40gb",GPU_I_ID="1",Hostname="n1"} 0.5
+DCGM_FI_PROF_SM_OCCUPANCY{gpu="0",UUID="GPU-a",device="nvidia0",modelName="NVIDIA L40S",Hostname="n1"} 0.25
+DCGM_FI_PROF_PIPE_TENSOR_ACTIVE{gpu="0",UUID="GPU-a",device="nvidia0",modelName="NVIDIA L40S",Hostname="n1"} 0
+`
+
+func TestDCGMExporterSaysHowMuchOfEachChipIsInUse(t *testing.T) {
+	got := parseDCGM([]byte(dcgmBody))
+	if len(got) != 1 {
+		t.Fatalf("want card 0 alone (card 1 unprofiled, card 2 a MIG instance): %+v", got)
+	}
+	p := got[0]
+	// The LinkedIn L40S: 100% busy, one SM of 142 at work.
+	if p.SMActive < 0.70 || p.SMActive > 0.71 || p.SMOccupancy != 25 || p.TensorActive != 0 || p.Source != "dcgm" {
+		t.Errorf("card 0 %+v", p)
+	}
+	// A field the exporter does not send is unknown, not zero.
+	if p.DRAMActive != -1 {
+		t.Errorf("DRAM active not sent, read as %v", p.DRAMActive)
+	}
+}
+
+type fakeGPM map[int]GPUProfile
+
+func (f fakeGPM) read() map[int]GPUProfile { return f }
+
+func TestGPMIsPreferredAndDCGMFillsTheCardsItDoesNotCover(t *testing.T) {
+	calls := 0
+	p := &ProfSampler{
+		gpm:  fakeGPM{1: {SMActive: 40, Source: "gpm"}},
+		dcgm: DefaultDCGMExporter,
+		get: func(context.Context, string) ([]byte, error) {
+			calls++
+			return []byte(`DCGM_FI_PROF_SM_ACTIVE{gpu="0"} 0.5` + "\n" + `DCGM_FI_PROF_SM_ACTIVE{gpu="1"} 0.9` + "\n"), nil
+		},
+		now: time.Now,
+	}
+	got := p.Sample(t.Context())
+	if got[0].Source != "dcgm" || got[0].SMActive != 50 {
+		t.Errorf("card 0 %+v", got[0])
+	}
+	if got[1].Source != "gpm" || got[1].SMActive != 40 {
+		t.Errorf("GPM's card was taken from DCGM: %+v", got[1])
+	}
+	if calls != 1 {
+		t.Errorf("%d calls", calls)
+	}
+}
+
+func TestANodeWithoutDCGMIsNotAskedEverySample(t *testing.T) {
+	now := time.Now()
+	calls := 0
+	p := &ProfSampler{dcgm: DefaultDCGMExporter, now: func() time.Time { return now },
+		get: func(context.Context, string) ([]byte, error) { calls++; return nil, errors.New("connection refused") }}
+	for range 5 {
+		if got := p.Sample(t.Context()); len(got) != 0 {
+			t.Fatalf("%+v", got)
+		}
+		now = now.Add(10 * time.Second)
+	}
+	if calls != 1 {
+		t.Errorf("asked %d times in 50 s", calls)
+	}
+	now = now.Add(dcgmRetry)
+	p.Sample(t.Context())
+	if calls != 2 {
+		t.Errorf("not asked again after %v", dcgmRetry)
+	}
+	if off := NewProfSampler("off"); off.dcgm != "" {
+		t.Errorf("off still asks %q", off.dcgm)
+	}
+}
+
+func TestEachSampleCarriesTheProfileOfTheCardsThatHaveOne(t *testing.T) {
+	smi := func(context.Context, string) ([]byte, error) {
+		return []byte("0, 71, 239, 260, 0x0, 92, 30, 23000, 24576, 1878, 0, 0\n" +
+			"1, 30, 20, 260, 0x0, 0, 0, 512, 24576, 300, 0, 0\n"), nil
+	}
+	now := time.Now()
+	s := sampler(t.TempDir(), smi, &now).WithProfile(&ProfSampler{
+		gpm: fakeGPM{0: {SMActive: 69, SMOccupancy: 45, TensorActive: 1, DRAMActive: 33, Source: "gpm"}}, now: time.Now})
+	got := s.Sample(t.Context()).GPUs
+	if got[0].Profile == nil || got[0].Profile.SMActive != 69 || got[0].UtilPercent != 92 {
+		t.Errorf("card 0 %+v", got[0])
+	}
+	if got[1].Profile != nil {
+		t.Errorf("card 1 has a profile nobody read: %+v", got[1].Profile)
+	}
+}

internal/agent/discovery/sensors.go

@@ -49,6 +49,9 @@ type GPUSensor struct {
 	MemoryUsedMB, MemoryTotalMB float64
 	SMClockMHz                  float64
 	ECCUncorrected              float64
+	// Profile is how much of the chip the work uses, where something on the
+	// node can say: nil on most cards. UtilPercent is not that.
+	Profile *GPUProfile
 }
 
 // sensorEvery bounds how often nvidia-smi is run. Heartbeats come more often
@@ -61,6 +64,8 @@ type SensorSampler struct {
 	sys string // /sys, or a test's copy of it
 	smi func(ctx context.Context, fields string) ([]byte, error)
 	now func() time.Time
+	// prof reads the cards' profiling counters; nil reads none.
+	prof *ProfSampler
 
 	mu         sync.Mutex
 	last       Sensors
@@ -74,6 +79,19 @@ func NewSensorSampler() *SensorSampler {
 	return &SensorSampler{sys: "/sys", smi: runSMI, now: time.Now}
 }
 
+// WithProfile has each sample carry the cards' profiles as well.
+func (s *SensorSampler) WithProfile(p *ProfSampler) *SensorSampler {
+	s.prof = p
+	return s
+}
+
+// SampleGPUs reads every card's heat, power and how busy it is now, without
+// a sampler's ten seconds of memory: for a screen that asks every few seconds
+// and wants what it asked for. Nil on a machine without nvidia-smi or a card.
+func SampleGPUs(ctx context.Context) []GPUSensor {
+	return (&SensorSampler{smi: runSMI}).gpus(ctx)
+}
+
 func runSMI(ctx context.Context, fields string) ([]byte, error) {
 	ctx, cancel := context.WithTimeout(ctx, 5*time.Second)
 	defer cancel()
@@ -94,6 +112,14 @@ func (s *SensorSampler) Sample(ctx context.Context) Sensors {
 	out.CPUTempC, out.InletTempC = s.hwmon()
 	out.CPUPowerW = s.rapl(now)
 	out.GPUs = s.gpus(ctx)
+	if s.prof != nil {
+		profs := s.prof.Sample(ctx)
+		for i := range out.GPUs {
+			if p, ok := profs[out.GPUs[i].Index]; ok {
+				out.GPUs[i].Profile = &p
+			}
+		}
+	}
 	s.last, s.lastAt = out, now
 	return out
 }

internal/agent/engines/comfyui.go

@@ -0,0 +1,151 @@
+package engines
+
+import (
+	"context"
+	"fmt"
+	"net"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strconv"
+	"sync/atomic"
+)
+
+// ComfyUIConfig locates a ComfyUI installation.
+type ComfyUIConfig struct {
+	// Dir is ComfyUI's directory, where main.py is. Readable by the user
+	// engines run as.
+	Dir string `koanf:"dir"`
+	// Python is the interpreter of its environment; Dir's .venv when unset.
+	Python string `koanf:"python"`
+}
+
+// ComfyUI runs ComfyUI as an engine, behind llmgw's gate (llmgw
+// comfyui-gate): ComfyUI asks for no key, so it listens on the loopback and
+// the gate, which does, on the replica's port. Its replicas make pictures
+// for /v1/images/generations.
+type ComfyUI struct {
+	cfg       ComfyUIConfig
+	installed atomic.Pointer[string]
+}
+
+// Name identifies the engine in the catalog.
+func (*ComfyUI) Name() string { return "comfyui" }
+
+// HealthPath is the gate's, which answers once ComfyUI does.
+func (*ComfyUI) HealthPath() string { return "/health" }
+
+// MetricsPath is empty: ComfyUI publishes no metrics, and the router reads
+// none from it.
+func (*ComfyUI) MetricsPath() string { return "" }
+
+// ServesHTTP is true: a ComfyUI replica is one process on one node.
+func (*ComfyUI) ServesHTTP(ShardSpec) bool { return true }
+
+// UseInstalled points the engine at the environment the agent installed:
+// its interpreter, with ComfyUI's code beside its bin, in comfyui/.
+func (c *ComfyUI) UseInstalled(python string) { c.installed.Store(&python) }
+
+// Configured reports whether the configuration names a ComfyUI.
+func (c *ComfyUI) Configured() bool { return c.cfg.Dir != "" }
+
+func (c *ComfyUI) dir() string {
+	if p := c.installed.Load(); p != nil {
+		return filepath.Join(filepath.Dir(filepath.Dir(*p)), "comfyui")
+	}
+	return c.cfg.Dir
+}
+
+func (c *ComfyUI) python() string {
+	if p := c.installed.Load(); p != nil {
+		return *p
+	}
+	if c.cfg.Python != "" {
+		return c.cfg.Python
+	}
+	return filepath.Join(c.cfg.Dir, ".venv", "bin", "python")
+}
+
+var comfyVersion = regexp.MustCompile(`__version__\s*=\s*"([^"]+)"`)
+
+// Probe reports ComfyUI's version, from the file its build writes it to:
+// importing ComfyUI to ask would load torch and look for cards.
+func (c *ComfyUI) Probe(context.Context) (string, error) {
+	dir := c.dir()
+	if dir == "" {
+		return "", fmt.Errorf("no ComfyUI directory is configured")
+	}
+	if _, err := os.Stat(filepath.Join(dir, "main.py")); err != nil {
+		return "", fmt.Errorf("%s is not ComfyUI: %w", dir, err)
+	}
+	b, err := os.ReadFile(filepath.Join(dir, "comfyui_version.py"))
+	if err != nil {
+		return Unknown, nil //nolint:nilerr // installed, of a version it does not say
+	}
+	if m := comfyVersion.FindSubmatch(b); m != nil {
+		return string(m[1]), nil
+	}
+	return Unknown, nil
+}
+
+// Validate checks the profile: one card, on one node. ComfyUI runs a
+// workflow on one device and has no tensor parallelism to spread it by.
+func (*ComfyUI) Validate(p Profile, gpus []GPUInfo) error {
+	if p.Nodes > 1 || p.TP > 1 || p.PP > 1 || p.GPUsPerNode > 1 {
+		return fmt.Errorf("profile %q: ComfyUI runs on one GPU of one node", p.Name)
+	}
+	return validateCommon(p, gpus)
+}
+
+// Commands is the gate, which starts ComfyUI itself.
+func (c *ComfyUI) Commands(s ShardSpec) ([]Command, error) {
+	dir := c.dir()
+	if dir == "" {
+		return nil, fmt.Errorf("comfyui: no ComfyUI directory is configured (engines.comfyui.dir)")
+	}
+	bin, err := os.Executable()
+	if err != nil {
+		return nil, err
+	}
+	p := s.Profile
+	args := []string{"comfyui-gate",
+		"--listen", net.JoinHostPort(s.Host(), strconv.Itoa(s.HTTPPort)),
+		"--model", p.ServedModelName,
+		"--weights", p.ModelPath,
+		// Relative to where engines run — their user's home — so it is
+		// theirs to write.
+		"--state", filepath.Join("comfyui", s.ReplicaID),
+		"--python", c.python(),
+		"--comfyui", dir,
+		"--",
+	}
+	args = append(args, p.Args...)
+	env := baseEnv(s)
+	if s.APIKey != "" {
+		env["LLMGW_ENGINE_KEY"] = s.APIKey
+	}
+	// A model's custom nodes and the repository a node loads its weights by,
+	// which only the catalog sets — a profile's env may not name them — and
+	// so taken before baseEnv drops them. Without the repository the gate
+	// links no Hugging Face cache, and sf3d's node went to the network for
+	// its gated repository and was refused with a 401.
+	for _, k := range []string{comfyNodesEnv, comfyHFRepoEnv, SharedModelsEnv} {
+		if v := p.Env[k]; v != "" {
+			env[k] = v
+		}
+	}
+	return []Command{{Name: "comfyui", Path: bin, Args: args, Env: materializeEnv(env)}}, nil
+}
+
+// comfyNodesEnv is the custom nodes the catalog gives a ComfyUI model, as
+// JSON; catalog.NodesEnv and comfy.NodesEnv by the same name.
+const comfyNodesEnv = "LLMGW_COMFYUI_NODES"
+
+// comfyHFRepoEnv is the repository a custom node loads the weights by name
+// from; catalog.HFRepoEnv and comfy.HFRepoEnv by the same name.
+const comfyHFRepoEnv = "LLMGW_COMFYUI_HF_REPO"
+
+// SharedModelsEnv is a shared replica's members — name, source, and the
+// path the agent fetched it to — as JSON; catalog.SharedModelsEnv and
+// comfy.SharedModelsEnv by the same name.
+const SharedModelsEnv = "LLMGW_COMFYUI_MODELS"

internal/agent/engines/comfyui_test.go

@@ -0,0 +1,99 @@
+package engines
+
+import (
+	"os"
+	"path/filepath"
+	"slices"
+	"strings"
+	"testing"
+)
+
+func TestComfyUIRunsBehindTheGateWithItsKeyOutOfSight(t *testing.T) {
+	c := &ComfyUI{cfg: ComfyUIConfig{Dir: "/opt/comfyui"}}
+	cmds, err := c.Commands(ShardSpec{
+		ReplicaID: "r-flux-1", HTTPPort: 8100, BindAddr: "10.0.0.147", APIKey: "sekret", GPUIndices: []int{0},
+		Profile: Profile{Name: "p", Engine: "comfyui", ModelPath: "/w/flux.safetensors", ServedModelName: "flux",
+			Args: []string{"--disable-smart-memory"}},
+	})
+	if err != nil || len(cmds) != 1 {
+		t.Fatalf("%v %v", cmds, err)
+	}
+	args := strings.Join(cmds[0].Args, " ")
+	for _, want := range []string{"comfyui-gate", "--listen 10.0.0.147:8100", "--model flux", "--weights /w/flux.safetensors",
+		"--state comfyui/r-flux-1", "--python /opt/comfyui/.venv/bin/python", "--comfyui /opt/comfyui", "-- --disable-smart-memory"} {
+		if !strings.Contains(args, want) {
+			t.Errorf("%q not in %q", want, args)
+		}
+	}
+	if strings.Contains(args, "sekret") {
+		t.Error("the key is on the command line, where every user can read it")
+	}
+	if !slices.Contains(cmds[0].Env, "LLMGW_ENGINE_KEY=sekret") || !slices.Contains(cmds[0].Env, "CUDA_VISIBLE_DEVICES=0") {
+		t.Error("the key or the card is not in the environment")
+	}
+}
+
+func TestTheGateIsToldTheNodesAndTheRepositoryTheCatalogGives(t *testing.T) {
+	// The repository was dropped on the way: sf3d's node went to the network
+	// for its gated repository, with no cache linked, and got a 401.
+	c := &ComfyUI{cfg: ComfyUIConfig{Dir: "/opt/comfyui"}}
+	cmds, err := c.Commands(ShardSpec{
+		ReplicaID: "r-sf3d-1", HTTPPort: 8100, BindAddr: "10.0.0.147", GPUIndices: []int{0},
+		Profile: Profile{Name: "p", Engine: "comfyui", ModelPath: "/w/sf3d/", ServedModelName: "sf3d",
+			Env: map[string]string{comfyNodesEnv: `[{"repo":"x"}]`, comfyHFRepoEnv: "stabilityai/stable-fast-3d"}},
+	})
+	if err != nil || len(cmds) != 1 {
+		t.Fatalf("%v %v", cmds, err)
+	}
+	for _, want := range []string{comfyNodesEnv + `=[{"repo":"x"}]`, comfyHFRepoEnv + "=stabilityai/stable-fast-3d"} {
+		if !slices.Contains(cmds[0].Env, want) {
+			t.Errorf("%q not in the gate's environment", want)
+		}
+	}
+}
+
+func TestComfyUIRunsOnOneCard(t *testing.T) {
+	c := &ComfyUI{}
+	if err := c.Validate(Profile{Name: "p", ModelPath: "/w", TP: 2}, h100(2)); err == nil {
+		t.Error("a two-card ComfyUI profile was accepted")
+	}
+	if err := c.Validate(Profile{Name: "p", ModelPath: "/w", TP: 1}, h100(1)); err != nil {
+		t.Error(err)
+	}
+	if _, err := c.Commands(ShardSpec{Profile: Profile{ModelPath: "/w"}}); err == nil {
+		t.Error("commands without a ComfyUI directory")
+	}
+}
+
+func TestComfyUIIsOfferedByTheVersionItsBuildWrote(t *testing.T) {
+	dir := t.TempDir()
+	c := &ComfyUI{cfg: ComfyUIConfig{Dir: dir}}
+	if _, err := c.Probe(t.Context()); err == nil {
+		t.Error("a directory without main.py passed for ComfyUI")
+	}
+	_ = os.WriteFile(filepath.Join(dir, "main.py"), nil, 0o644)
+	_ = os.WriteFile(filepath.Join(dir, "comfyui_version.py"), []byte("__version__ = \"0.39.2\"\n"), 0o644)
+	if v, err := c.Probe(t.Context()); err != nil || v != "0.39.2" {
+		t.Errorf("%q %v", v, err)
+	}
+	if got := NewRegistry(Config{ComfyUI: ComfyUIConfig{Dir: dir}}).ProbeConfigured(t.Context()); got["comfyui"] != "0.39.2" {
+		t.Errorf("a configured ComfyUI is reported as %v", got)
+	}
+	if err := UnsafeArgs("comfyui", []string{"--enable-manager"}); err == nil {
+		t.Error("a profile may install ComfyUI nodes")
+	}
+}
+
+func TestAnInstalledComfyUIIsFoundBesideItsInterpreter(t *testing.T) {
+	c := &ComfyUI{cfg: ComfyUIConfig{Dir: "/opt/by-hand"}}
+	c.UseInstalled("/opt/llmgw/engines/comfyui/0.39.2-cu13/bin/python")
+	cmds, err := c.Commands(ShardSpec{ReplicaID: "r", HTTPPort: 1, Profile: Profile{ModelPath: "/w", ServedModelName: "m"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	args := strings.Join(cmds[0].Args, " ")
+	if !strings.Contains(args, "--python /opt/llmgw/engines/comfyui/0.39.2-cu13/bin/python") ||
+		!strings.Contains(args, "--comfyui /opt/llmgw/engines/comfyui/0.39.2-cu13/comfyui") {
+		t.Errorf("the installed ComfyUI is not what runs: %s", args)
+	}
+}

internal/agent/engines/engine.go

@@ -47,6 +47,13 @@ type Profile struct {
 	// KVEvents has the engine publish its KV cache events to the agent,
 	// for the router's precise index. vLLM only.
 	KVEvents bool `json:"kv_events,omitempty"`
+	// AllowPickle lets the node fetch weights in a pickle format, which run
+	// code when loaded; the catalog's allow_pickle.
+	AllowPickle bool `json:"allow_pickle,omitempty"`
+	// ModelSHA256 is the sha256 the registry published for the file
+	// ModelPath names; the node refuses a copy that is not it. Empty when
+	// nothing is known to check against.
+	ModelSHA256 string `json:"model_sha256,omitempty"`
 }
 
 // ShardSpec is one shard of one replica on this node.
@@ -214,6 +221,9 @@ func NewRegistry(cfg Config) *Registry {
 	r.Add(&LlamaCpp{cfg: cfg.LlamaCpp})
 	r.Add(&TRTLLM{cfg: cfg.TRTLLM})
 	r.Add(&Sim{cfg: cfg.Sim})
+	r.Add(&ComfyUI{cfg: cfg.ComfyUI})
+	r.Add(&Strata{cfg: cfg.Strata})
+	r.Add(&FreeToken{cfg: cfg.FreeToken})
 	return r
 }
 
@@ -262,6 +272,15 @@ type Config struct {
 	// offers it only when its interpreter is named here.
 	TRTLLM TRTLLMConfig `koanf:"trtllm"`
 	Sim    SimConfig    `koanf:"sim"`
+	// ComfyUI makes pictures; offered by a node whose ComfyUI is named here.
+	ComfyUI ComfyUIConfig `koanf:"comfyui"`
+	// Strata serves a few mixture-of-experts models from RAM and one card;
+	// offered by a node whose Strata is named here, or that installed it.
+	Strata StrataConfig `koanf:"strata"`
+	// FreeToken serves mixture-of-experts models larger than a card from
+	// RAM and one card of 8.0 or newer; installed by the agent where the
+	// cards allow it, or named here.
+	FreeToken FreeTokenConfig `koanf:"freetoken"`
 }
 
 // --- shared helpers ---

internal/agent/engines/engines_test.go

@@ -477,14 +477,14 @@ func TestValidateRequiresModelPath(t *testing.T) {
 
 func TestRegistry(t *testing.T) {
 	r := NewRegistry(Config{})
-	if got := r.Names(); !slices.Equal(got, []string{"llamacpp", "sglang", "sim", "trtllm", "vllm"}) {
+	if got := r.Names(); !slices.Equal(got, []string{"comfyui", "freetoken", "llamacpp", "sglang", "sim", "strata", "trtllm", "vllm"}) {
 		t.Errorf("registry contents: %v", got)
 	}
 	if _, err := r.Get("vllm"); err != nil {
 		t.Errorf("vllm missing: %v", err)
 	}
 	err := func() error { _, err := r.Get("nope"); return err }()
-	if err == nil || !strings.Contains(err.Error(), "llamacpp, sglang, sim, trtllm, vllm") {
+	if err == nil || !strings.Contains(err.Error(), "llamacpp, sglang, sim, strata, trtllm, vllm") {
 		t.Errorf("an unknown engine should list what is available: %v", err)
 	}
 }

internal/agent/engines/freetoken.go

@@ -0,0 +1,137 @@
+package engines
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"sync/atomic"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/discovery"
+)
+
+// FreeTokenConfig locates a FreeToken installation.
+type FreeTokenConfig struct {
+	// Binary is its ft command, in the environment it was installed into.
+	// The agent installs one itself on a node whose GPUs it runs on; this
+	// is for one installed by hand.
+	Binary string `koanf:"binary"`
+}
+
+// freetokenToolkit finds the CUDA 13 toolkit; a variable for tests.
+var freetokenToolkit = func() string { return discovery.CUDAToolkit(13) }
+
+// FreeToken runs FreeToken (github.com/FlashML-org/FreeToken): a
+// mixture-of-experts engine that keeps a model's experts in host RAM, the
+// ones used most in VRAM, and streams the rest over PCIe, so a model several
+// times a card's size serves from one card. It reads Hugging Face
+// safetensors checkpoints as they are published (FP8, NVFP4, MXFP4, BF16)
+// and serves the OpenAI and Anthropic APIs. One card of compute capability
+// 8.0 or newer; it compiles some of its kernels on first start, with the
+// CUDA 13 toolkit.
+type FreeToken struct {
+	cfg       FreeTokenConfig
+	installed atomic.Pointer[string]
+}
+
+// Name identifies the engine in the catalog.
+func (*FreeToken) Name() string { return "freetoken" }
+
+// HealthPath answers before the model is loaded too, with its progress in
+// the body; the supervisor waits for the status code, so it counts as ready
+// only once FreeToken's own startup — loading, warm-up, CUDA graphs — is
+// over and the API server is accepting.
+func (*FreeToken) HealthPath() string { return "/health" }
+
+// MetricsPath is empty: FreeToken serves its numbers as JSON (/v1/stats),
+// not in Prometheus' text, so the router counts its requests itself.
+func (*FreeToken) MetricsPath() string { return "" }
+
+// ServesHTTP is true: a FreeToken replica is one process on one node.
+func (*FreeToken) ServesHTTP(ShardSpec) bool { return true }
+
+// UseInstalled points the engine at the ft the agent installed.
+func (f *FreeToken) UseInstalled(path string) { f.installed.Store(&path) }
+
+// Configured reports whether the configuration names a FreeToken.
+func (f *FreeToken) Configured() bool { return f.cfg.Binary != "" }
+
+func (f *FreeToken) binary() string {
+	if p := f.installed.Load(); p != nil {
+		return *p
+	}
+	return f.cfg.Binary
+}
+
+// HelpCommand is ft serve's own help.
+func (f *FreeToken) HelpCommand(Command) (string, []string) {
+	return f.binary(), []string{"serve", "--help"}
+}
+
+// Probe reports the FreeToken a configuration names, by what it says its
+// version is.
+func (f *FreeToken) Probe(ctx context.Context) (string, error) {
+	bin := f.binary()
+	if bin == "" {
+		return "", fmt.Errorf("no FreeToken is configured")
+	}
+	out, err := exec.CommandContext(ctx, bin, "--version").Output() //nolint:gosec // the configured engine
+	if err != nil {
+		return "", fmt.Errorf("%s does not run: %w", bin, err)
+	}
+	// "freetoken version 0.2.1"
+	if fs := strings.Fields(string(out)); len(fs) > 0 {
+		return fs[len(fs)-1], nil
+	}
+	return Unknown, nil
+}
+
+// Validate checks the profile: one card of 8.0 or newer, on one node.
+// FreeToken has no tensor parallel; it picks one device.
+func (*FreeToken) Validate(p Profile, gpus []GPUInfo) error {
+	if p.Nodes > 1 || p.TP > 1 || p.PP > 1 || p.GPUsPerNode > 1 {
+		return fmt.Errorf("profile %q: freetoken runs on one GPU of one node", p.Name)
+	}
+	if err := requireComputeCapability(p, gpus, 8.0, "freetoken"); err != nil {
+		return err
+	}
+	return validateCommon(p, gpus)
+}
+
+// Commands is ft serve. The card is the shard's, by CUDA_VISIBLE_DEVICES,
+// so it is device 0 to FreeToken; its key goes in the environment, where it
+// reads it from, rather than on a command line every user can read.
+func (f *FreeToken) Commands(sh ShardSpec) ([]Command, error) {
+	bin := f.binary()
+	if bin == "" {
+		return nil, fmt.Errorf("freetoken: no FreeToken is installed or configured (engines.freetoken.binary)")
+	}
+	toolkit := freetokenToolkit()
+	if toolkit == "" {
+		return nil, fmt.Errorf("freetoken compiles its kernels on first start and this node has no CUDA 13 toolkit " +
+			"(nvcc in CUDA_HOME, /usr/local/cuda or /usr/local/cuda-13.*)")
+	}
+	p := sh.Profile
+	args := []string{"serve",
+		"--model", p.ModelPath,
+		"--host", sh.Host(),
+		"--port", strconv.Itoa(sh.HTTPPort),
+		"--gpu", "0",
+	}
+	if p.ServedModelName != "" {
+		args = append(args, "--served-model-name", p.ServedModelName)
+	}
+	args = append(args, p.Args...)
+	env := baseEnv(sh)
+	env["CUDA_HOME"] = toolkit
+	// nvcc for the kernels it compiles, and the environment's own bin
+	// first, where ninja is.
+	env["PATH"] = strings.Join([]string{filepath.Dir(bin), filepath.Join(toolkit, "bin"), os.Getenv("PATH")}, ":")
+	if sh.APIKey != "" {
+		env["FREETOKEN_API_KEY"] = sh.APIKey
+	}
+	return []Command{{Name: "freetoken", Path: bin, Args: args, Env: materializeEnv(env)}}, nil
+}

internal/agent/engines/freetoken_test.go

@@ -0,0 +1,112 @@
+package engines
+
+import (
+	"slices"
+	"strings"
+	"testing"
+)
+
+func withToolkit(t *testing.T, dir string) {
+	t.Helper()
+	old := freetokenToolkit
+	freetokenToolkit = func() string { return dir }
+	t.Cleanup(func() { freetokenToolkit = old })
+}
+
+func TestFreeTokenServesOnTheShardsCardWithTheKeyInTheEnvironment(t *testing.T) {
+	withToolkit(t, "/usr/local/cuda-13.4")
+	f := &FreeToken{}
+	f.UseInstalled("/opt/llmgw/engines/freetoken/0.2.1-cu13/bin/ft")
+	spec := ShardSpec{
+		ReplicaID: "r1", HTTPPort: 18091, BindAddr: "10.0.0.160", GPUIndices: []int{3},
+		APIKey: "secret-key",
+		Profile: Profile{Name: "p", Engine: "freetoken", TP: 1, ModelPath: "/var/lib/llmgw-weights/glm-5.2-nvfp4",
+			ServedModelName: "glm-5.2", Args: []string{"--max-running-requests", "8"}},
+	}
+	cmds, err := f.Commands(spec)
+	if err != nil {
+		t.Fatal(err)
+	}
+	c := cmds[0]
+	if c.Path != "/opt/llmgw/engines/freetoken/0.2.1-cu13/bin/ft" {
+		t.Errorf("path %s", c.Path)
+	}
+	args := strings.Join(c.Args, " ")
+	for _, want := range []string{
+		"serve --model /var/lib/llmgw-weights/glm-5.2-nvfp4", "--host 10.0.0.160", "--port 18091",
+		// The shard's card is device 0 once CUDA_VISIBLE_DEVICES pins it.
+		"--gpu 0", "--served-model-name glm-5.2", "--max-running-requests 8",
+	} {
+		if !strings.Contains(args, want) {
+			t.Errorf("args lack %q: %s", want, args)
+		}
+	}
+	if strings.Contains(args, "secret-key") {
+		t.Error("the key is on the command line, which every user on the node can read")
+	}
+	for _, want := range []string{"FREETOKEN_API_KEY=secret-key", "CUDA_VISIBLE_DEVICES=3", "CUDA_HOME=/usr/local/cuda-13.4"} {
+		if !slices.Contains(c.Env, want) {
+			t.Errorf("env lacks %s", want)
+		}
+	}
+	var path string
+	for _, e := range c.Env {
+		if p, ok := strings.CutPrefix(e, "PATH="); ok {
+			path = p
+		}
+	}
+	if !strings.HasPrefix(path, "/opt/llmgw/engines/freetoken/0.2.1-cu13/bin:/usr/local/cuda-13.4/bin:") {
+		t.Errorf("nvcc and the environment's tools are not first on PATH: %s", path)
+	}
+}
+
+func TestFreeTokenWithoutTheToolkitSaysSoInsteadOfFailingOnItsFirstModel(t *testing.T) {
+	withToolkit(t, "")
+	f := &FreeToken{cfg: FreeTokenConfig{Binary: "/x/bin/ft"}}
+	_, err := f.Commands(ShardSpec{HTTPPort: 1, Profile: Profile{Name: "p", ModelPath: "/m"}})
+	if err == nil || !strings.Contains(err.Error(), "CUDA 13 toolkit") {
+		t.Errorf("got %v", err)
+	}
+}
+
+func TestFreeTokenRunsOnOneAmpereOrNewerCard(t *testing.T) {
+	f := &FreeToken{}
+	p := Profile{Name: "p", Engine: "freetoken", TP: 1, ModelPath: "/m"}
+	if err := f.Validate(p, []GPUInfo{{Name: "NVIDIA RTX A6000", ComputeCapability: "8.6", MemoryTotalMB: 49140}}); err != nil {
+		t.Errorf("an Ampere card was refused: %v", err)
+	}
+	// Tried on cuda01: no kernel image for the card, then an illegal memory
+	// access in fp16.
+	if err := f.Validate(p, []GPUInfo{{Name: "Quadro RTX 6000", ComputeCapability: "7.5", MemoryTotalMB: 24576}}); err == nil {
+		t.Error("a Turing card was accepted")
+	}
+	p.TP = 2
+	if err := f.Validate(p, h100(2)); err == nil {
+		t.Error("a two-card profile was accepted; FreeToken has no tensor parallel")
+	}
+}
+
+func TestFreeTokenFlagsThatReadFilesOrMoveWhatTheAgentSetsAreRefused(t *testing.T) {
+	for _, a := range []string{
+		"--gpu", "--served-model-name=x", "--mmproj", "--allowed-local-media-path",
+		// argparse takes a beginning for the whole flag.
+		"--allowed-local", "--api-key", "--port", "--hf-overrides",
+	} {
+		if UnsafeArgs("freetoken", []string{a, "v"}) == nil {
+			t.Errorf("%s was allowed", a)
+		}
+	}
+	for _, a := range []string{"--max-running-requests", "--moe-strategy", "--dtype", "--max-seq-len-override", "--moe-cache-auto"} {
+		if err := UnsafeArgs("freetoken", []string{a, "1"}); err != nil {
+			t.Errorf("%s was refused: %v", a, err)
+		}
+	}
+	for k, refuse := range map[string]bool{
+		"FREETOKEN_PIN_BUDGET_GB": false, "FREETOKEN_API_KEY": true, "FREETOKEN_API_LOG_DIR": true,
+		"FREETOKEN_MMPROJ": true, "FREETOKEN_HOST": true,
+	} {
+		if got, why := UnsafeEnv(k); got != refuse {
+			t.Errorf("%s: refused %v (%s), want %v", k, got, why, refuse)
+		}
+	}
+}

internal/agent/engines/pickle_test.go

@@ -0,0 +1,58 @@
+package engines
+
+import (
+	"slices"
+	"testing"
+)
+
+func TestAModelDownloadedByTheEngineIsReadAsSafetensorsOnly(t *testing.T) {
+	for path, want := range map[string]bool{
+		"Qwen/Qwen2.5-7B-Instruct":    true,
+		"/var/lib/llmgw/weights/abc/": false, "./local": false,
+		"http://m/q/v/": false, "q.gguf": false,
+	} {
+		got := slices.Contains(safetensorsOnly(Profile{ModelPath: path}), "safetensors")
+		if got != want {
+			t.Errorf("%q: safetensors only = %v, want %v", path, got, want)
+		}
+	}
+	if len(safetensorsOnly(Profile{ModelPath: "Qwen/Qwen2.5-7B-Instruct", AllowPickle: true})) != 0 {
+		t.Error("allow_pickle still forced safetensors")
+	}
+}
+
+func TestAForbiddenFlagIsKnownHoweverItIsSpelled(t *testing.T) {
+	for engine, args := range map[string][]string{
+		"vllm":   {"--trust_remote_code"},
+		"sglang": {"--load-format", "pt"},
+	} {
+		if UnsafeArgs(engine, args) == nil {
+			t.Errorf("%s %v was let through", engine, args)
+		}
+	}
+	for _, a := range [][]string{{"--load_format", "pt"}, {"--trust-remote"}, {"--load-form=pt"}} {
+		if UnsafeArgs("vllm", a) == nil {
+			t.Errorf("vllm %v was let through", a)
+		}
+	}
+	// What is not one of them, nor the start of one, still passes.
+	for _, a := range [][]string{{"--max-model-len", "8192"}, {"--tokenizer-mode", "mistral"}, {"--gpu-memory-utilization=0.9"}} {
+		if err := UnsafeArgs("vllm", a); err != nil {
+			t.Errorf("vllm %v was refused: %v", a, err)
+		}
+	}
+	// Short flags are matched as written; llama.cpp takes no abbreviations,
+	// so its --chat-template is not taken for --chat-template-file.
+	for _, a := range [][]string{{"-m", "x"}, {"-hf", "o/r"}, {"-mu", "http://x"}} {
+		if UnsafeArgs("llamacpp", a) == nil {
+			t.Errorf("llamacpp %v was let through", a)
+		}
+	}
+	if err := UnsafeArgs("llamacpp", []string{"--chat-template", "chatml"}); err != nil {
+		t.Errorf("llama.cpp's own --chat-template was refused: %v", err)
+	}
+	// ComfyUI's argparse reads --li as --listen.
+	if UnsafeArgs("comfyui", []string{"--li", "0.0.0.0"}) == nil {
+		t.Error("comfyui --li was let through")
+	}
+}

internal/agent/engines/sglang.go

@@ -102,6 +102,7 @@ func (g *SGLang) serverCommand(s ShardSpec) (Command, error) {
 	if p.ServedModelName != "" {
 		args = append(args, "--served-model-name", p.ServedModelName)
 	}
+	args = append(args, safetensorsOnly(p)...)
 	if s.NNodes > 1 {
 		args = append(args,
 			"--nnodes", strconv.Itoa(s.NNodes),

internal/agent/engines/strata.go

@@ -0,0 +1,184 @@
+package engines
+
+import (
+	"context"
+	"fmt"
+	"net"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strconv"
+	"sync/atomic"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/strata"
+)
+
+// StrataConfig locates a Strata installation.
+type StrataConfig struct {
+	// Dir is Strata's source directory, where setup.py and serve/ are.
+	Dir string `koanf:"dir"`
+	// Python is the interpreter of its environment; Dir's .venv when unset.
+	Python string `koanf:"python"`
+	// Engine is the directory of its prebuilt engine, where the strata
+	// binary and the CUDA libraries it ships are; Dir's engine/ when unset.
+	Engine string `koanf:"engine"`
+}
+
+// Strata runs Strata (github.com/Niko1221/Strata): a mixture-of-experts
+// engine that keeps a model's experts in RAM and the ones used most in VRAM,
+// so a 125B model serves from one 24 GB card. It runs a few model families
+// only, each prepared from its GGUF files by Strata's own setup, so the agent
+// starts llmgw strata-serve, which prepares the model and then becomes
+// Strata's server.
+type Strata struct {
+	cfg       StrataConfig
+	installed atomic.Pointer[string]
+}
+
+// Name identifies the engine in the catalog.
+func (*Strata) Name() string { return "strata" }
+
+// HealthPath answers 200 only once the model is loaded.
+func (*Strata) HealthPath() string { return "/health" }
+
+// MetricsPath is Strata's Prometheus text, in vLLM's names; its /metrics is
+// JSON for its own monitor unless asked for this format.
+func (*Strata) MetricsPath() string { return "/metrics?format=prometheus" }
+
+// ServesHTTP is true: a Strata replica is one process on one node.
+func (*Strata) ServesHTTP(ShardSpec) bool { return true }
+
+// UseInstalled points the engine at the environment the agent installed:
+// its interpreter, with Strata's code beside its bin in strata/ and the
+// engine in strata-engine/.
+func (s *Strata) UseInstalled(python string) { s.installed.Store(&python) }
+
+// Configured reports whether the configuration names a Strata.
+func (s *Strata) Configured() bool { return s.cfg.Dir != "" }
+
+// StrataSourceDir and StrataEngineDir are where an installed environment
+// keeps Strata's code and its prebuilt engine, beside bin/.
+const (
+	StrataSourceDir = "strata"
+	StrataEngineDir = "strata-engine"
+)
+
+func (s *Strata) env() string {
+	if p := s.installed.Load(); p != nil {
+		return filepath.Dir(filepath.Dir(*p))
+	}
+	return ""
+}
+
+func (s *Strata) dir() string {
+	if e := s.env(); e != "" {
+		return filepath.Join(e, StrataSourceDir)
+	}
+	return s.cfg.Dir
+}
+
+func (s *Strata) engineDir() string {
+	if e := s.env(); e != "" {
+		return filepath.Join(e, StrataEngineDir)
+	}
+	if s.cfg.Engine != "" {
+		return s.cfg.Engine
+	}
+	return filepath.Join(s.cfg.Dir, "engine")
+}
+
+func (s *Strata) python() string {
+	if p := s.installed.Load(); p != nil {
+		return *p
+	}
+	if s.cfg.Python != "" {
+		return s.cfg.Python
+	}
+	return filepath.Join(s.cfg.Dir, ".venv", "bin", "python")
+}
+
+var strataVersion = regexp.MustCompile(`PRODUCT_VERSION\s*=\s*"([^"]+)"`)
+
+// Probe reports the Strata a configuration names, by the version its source
+// states: a directory with its setup and server, an interpreter, and the
+// engine binary. Without it a Strata named by hand was never reported, and
+// so never placed on.
+func (s *Strata) Probe(context.Context) (string, error) {
+	dir := s.dir()
+	if dir == "" {
+		return "", fmt.Errorf("no Strata directory is configured")
+	}
+	for _, f := range []string{filepath.Join(dir, "setup.py"), filepath.Join(dir, "serve", "server.py"),
+		s.python(), filepath.Join(s.engineDir(), "strata")} {
+		if _, err := os.Stat(f); err != nil {
+			return "", fmt.Errorf("%s is not a whole Strata: %w", dir, err)
+		}
+	}
+	b, err := os.ReadFile(filepath.Join(dir, "strata_cli", "__init__.py"))
+	if err != nil {
+		return Unknown, nil //nolint:nilerr // whole, of a version it does not say
+	}
+	if m := strataVersion.FindSubmatch(b); m != nil {
+		return string(m[1]), nil
+	}
+	return Unknown, nil
+}
+
+// Validate checks the profile: one card, on one node. Strata splits a model
+// over a node's cards by layer, but that is not driven from here yet.
+func (*Strata) Validate(p Profile, gpus []GPUInfo) error {
+	if p.Nodes > 1 || p.TP > 1 || p.PP > 1 || p.GPUsPerNode > 1 {
+		return fmt.Errorf("profile %q: strata runs on one GPU of one node", p.Name)
+	}
+	// Pascal and Volta run its CUDA 12 build, which the install picks for
+	// them.
+	if err := requireComputeCapability(p, gpus, 6.0, "strata"); err != nil {
+		return err
+	}
+	// A family whose engine needs more: said now, rather than by Strata's
+	// setup after the replica was placed here.
+	if fam, ok := strata.FamilyNamed(strata.ArgValue(p.Args, "--family")); ok && fam.MinCC > 0 {
+		if err := requireComputeCapability(p, gpus, fam.MinCC, "strata's "+fam.Title+" engine"); err != nil {
+			return err
+		}
+	}
+	return validateCommon(p, gpus)
+}
+
+// Commands is llmgw strata-serve, which prepares the model with Strata's
+// setup and then execs Strata's server.
+func (s *Strata) Commands(sh ShardSpec) ([]Command, error) {
+	dir := s.dir()
+	if dir == "" {
+		return nil, fmt.Errorf("strata: no Strata directory is configured (engines.strata.dir)")
+	}
+	bin, err := os.Executable()
+	if err != nil {
+		return nil, err
+	}
+	p := sh.Profile
+	args := []string{"strata-serve",
+		"--listen", net.JoinHostPort(sh.Host(), strconv.Itoa(sh.HTTPPort)),
+		"--model", p.ServedModelName,
+		"--weights", p.ModelPath,
+		// Relative to where engines run — their user's home — so they are
+		// theirs to write. The data is shared by the node's replicas: a
+		// prepared model and its draft layer are gigabytes, the same for
+		// every replica of it.
+		"--state", filepath.Join("strata", sh.ReplicaID),
+		"--data", filepath.Join("strata", "data"),
+		"--python", s.python(),
+		"--strata", dir,
+		"--engine", s.engineDir(),
+		"--",
+	}
+	args = append(args, p.Args...)
+	env := baseEnv(sh)
+	if sh.APIKey != "" {
+		// Strata's server reads its key from here, and strata-serve passes
+		// it on in the environment: not on a command line every user on the
+		// node can read.
+		env["STRATA_API_KEY"] = sh.APIKey
+	}
+	return []Command{{Name: "strata", Path: bin, Args: args, Env: materializeEnv(env)}}, nil
+}

internal/agent/engines/strata_test.go

@@ -0,0 +1,113 @@
+package engines
+
+import (
+	"os"
+	"path/filepath"
+	"slices"
+	"strings"
+	"testing"
+)
+
+func TestStrataStartsStrataServeWithTheKeyInTheEnvironment(t *testing.T) {
+	s := &Strata{}
+	s.UseInstalled("/opt/llmgw/engines/strata/0.2.7-cuda13/bin/python")
+	spec := ShardSpec{
+		ReplicaID: "r1", HTTPPort: 18090, BindAddr: "10.0.0.147", GPUIndices: []int{0},
+		APIKey: "secret-key",
+		Profile: Profile{Name: "p", Engine: "strata", TP: 1, ModelPath: "/var/lib/llmgw/models/q38",
+			ServedModelName: "qwen3.8-flash-next", Args: []string{"--family", "qwen", "--size", "IQ2_XS"}},
+	}
+	cmds, err := s.Commands(spec)
+	if err != nil {
+		t.Fatal(err)
+	}
+	c := cmds[0]
+	args := strings.Join(c.Args, " ")
+	env := "/opt/llmgw/engines/strata/0.2.7-cuda13"
+	for _, want := range []string{
+		"strata-serve", "--listen 10.0.0.147:18090", "--model qwen3.8-flash-next",
+		"--weights /var/lib/llmgw/models/q38", "--state " + filepath.Join("strata", "r1"),
+		"--python " + env + "/bin/python", "--strata " + env + "/strata", "--engine " + env + "/strata-engine",
+		"-- --family qwen --size IQ2_XS",
+	} {
+		if !strings.Contains(args, want) {
+			t.Errorf("args lack %q: %s", want, args)
+		}
+	}
+	if strings.Contains(args, "secret-key") {
+		t.Error("the key is on the command line, which every user on the node can read")
+	}
+	if !slices.Contains(c.Env, "STRATA_API_KEY=secret-key") {
+		t.Error("the key is not in the environment, where Strata's server reads it")
+	}
+	if !slices.Contains(c.Env, "CUDA_VISIBLE_DEVICES=0") {
+		t.Error("the shard is not pinned to its GPU")
+	}
+}
+
+func TestStrataRunsOnOneCard(t *testing.T) {
+	s := &Strata{}
+	p := Profile{Name: "p", Engine: "strata", TP: 1, GPUsPerNode: 2, ModelPath: "/m"}
+	if err := s.Validate(p, nil); err == nil {
+		t.Error("a two-card profile was accepted")
+	}
+	p.GPUsPerNode = 1
+	if err := s.Validate(p, []GPUInfo{{Name: "Quadro RTX 6000", ComputeCapability: "7.5"}}); err != nil {
+		t.Errorf("one Turing card was refused: %v", err)
+	}
+}
+
+func TestStrataSetupFlagsThatMoveFilesAreRefused(t *testing.T) {
+	for _, bad := range []string{"--gguf-dir", "--gguf", "--data-dir=/etc", "--model", "--mod", "--feed", "--host", "--gpus"} {
+		if err := UnsafeArgs("strata", []string{bad, "x"}); err == nil {
+			t.Errorf("%s was accepted", bad)
+		}
+	}
+	if err := UnsafeArgs("strata", []string{"--family", "qwen", "--size", "IQ2_XS", "--context", "32768",
+		"--parallel", "2", "--vram-reserve-mib", "2048", "--vision", "no"}); err != nil {
+		t.Errorf("an ordinary profile was refused: %v", err)
+	}
+}
+
+func TestAStrataNamedByHandIsReportedWithItsVersion(t *testing.T) {
+	dir := t.TempDir()
+	for _, f := range []string{"setup.py", "serve/server.py", ".venv/bin/python", "engine/strata"} {
+		p := filepath.Join(dir, f)
+		if err := os.MkdirAll(filepath.Dir(p), 0o755); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(p, nil, 0o755); err != nil {
+			t.Fatal(err)
+		}
+	}
+	s := &Strata{cfg: StrataConfig{Dir: dir}}
+	if v, err := s.Probe(t.Context()); err != nil || v != Unknown {
+		t.Fatalf("%q %v", v, err)
+	}
+	_ = os.MkdirAll(filepath.Join(dir, "strata_cli"), 0o755)
+	_ = os.WriteFile(filepath.Join(dir, "strata_cli", "__init__.py"), []byte("PRODUCT_VERSION = \"0.2.7\"\n"), 0o644)
+	if v, _ := s.Probe(t.Context()); v != "0.2.7" {
+		t.Errorf("version %q", v)
+	}
+	r := NewRegistry(Config{Strata: StrataConfig{Dir: dir}})
+	if got := r.ProbeConfigured(t.Context()); got["strata"] != "0.2.7" {
+		t.Errorf("not reported: %v", got)
+	}
+	_ = os.Remove(filepath.Join(dir, "engine", "strata"))
+	if _, err := s.Probe(t.Context()); err == nil {
+		t.Error("reported without its engine")
+	}
+}
+
+func TestStrataRefusesDeepSeekOnACardItsEngineWasNotBuiltFor(t *testing.T) {
+	s := &Strata{}
+	p := Profile{Name: "p", Engine: "strata", ModelPath: "/m", TP: 1, GPUsPerNode: 1, Args: []string{"--family", "deepseek41", "--size", "IST-DAS-1.75bpw"}}
+	gpus := []GPUInfo{{Index: 0, Name: "Quadro RTX 6000", ComputeCapability: "7.5"}}
+	if err := s.Validate(p, gpus); err == nil || !strings.Contains(err.Error(), "8.6") {
+		t.Errorf("DeepSeek on 7.5: %v", err)
+	}
+	p.Args = []string{"--family", "qwen", "--size", "IQ2_XS"}
+	if err := s.Validate(p, gpus); err != nil {
+		t.Errorf("Qwen on 7.5: %v", err)
+	}
+}

internal/agent/engines/unsafeenv.go

@@ -2,6 +2,7 @@ package engines
 
 import (
 	"fmt"
+	"os"
 	"strings"
 )
 
@@ -24,7 +25,7 @@ var envFamilies = []string{
 	"CUDA_", "NCCL_", "NVIDIA_", "NVTE_", "GLOO_", "UCX_", "TORCH_", "PYTORCH_",
 	"VLLM_", "SGLANG_", "GGML_", "LLAMA_", "OMP_", "MKL_", "OPENBLAS_", "KMP_",
 	"HF_", "TRANSFORMERS_", "TOKENIZERS_", "TRITON_", "FLASHINFER_",
-	"TRTLLM_", "TLLM_",
+	"TRTLLM_", "TLLM_", "FREETOKEN_",
 }
 
 // envExact are single variables that are tuning too.
@@ -32,6 +33,14 @@ var envExact = map[string]bool{
 	"PYTHONUNBUFFERED": true, "PYTHONHASHSEED": true, "TZ": true,
 }
 
+// envNeverExact are variables in an allowed family that no profile may set,
+// whose names the patterns below do not catch: what the agent sets itself,
+// and FreeToken's way of loading a projector file.
+var envNeverExact = map[string]string{
+	"FREETOKEN_API_KEY": "is set by the agent", "FREETOKEN_HOST": "is set by the agent",
+	"FREETOKEN_MMPROJ": "loads a file",
+}
+
 // envNever are names in an allowed family that load code or move where the
 // engine reads and writes.
 var envNever = []string{
@@ -58,6 +67,9 @@ func UnsafeEnv(key string) (bool, string) {
 		return true, "only the variables engines are tuned with are allowed (" +
 			strings.Join(envFamilies, "*, ") + "*)"
 	}
+	if why, ok := envNeverExact[k]; ok {
+		return true, "it " + why
+	}
 	for _, n := range envNever {
 		if strings.Contains(k, n) && !envExact[k] {
 			return true, "it names a path, a plugin, a library or an address, which is a way to make the engine " +
@@ -101,6 +113,13 @@ var flagsNever = map[string]map[string]string{
 		"--hf-token": "is a credential", "--alias": "is set by the agent",
 		"--model-draft": "loads a file", "-md": "loads a file",
 		"--prompt-cache": "writes a file", "--logdir": "writes files",
+		// The short spellings of the same.
+		"-mm": "loads a file; the projector is fetched with the model", "--mmproj-url": "downloads a file",
+		"-mmu": "downloads a file", "-hfr": "downloads a model", "-hff": "downloads a model",
+		"--hf-repo-draft": "downloads a model", "-hfd": "downloads a model", "-hfrd": "downloads a model",
+		"--hf-repo-v": "downloads a model", "-hfv": "downloads a model", "--hf-file-v": "downloads a model",
+		"-hffv": "downloads a model", "-hft": "is a credential", "-jf": "reads a file", "-a": "is set by the agent",
+		"--model-vocoder": "loads a file", "-mv": "loads a file",
 	},
 	"vllm": {
 		"--download-dir": "writes files", "--tokenizer": "reads another repository or path",
@@ -118,6 +137,46 @@ var flagsNever = map[string]map[string]string{
 		"--file-storage-path": "writes files", "--log-requests": "logs prompts",
 		"--model-path": "is the model's source, set by the agent", "--served-model-name": "is set by the agent",
 		"--lora-paths": "loads files", "--download-dir": "writes files",
+		"--load-format": "changes how weights are loaded, some ways running code",
+	},
+	// ComfyUI's main.py, behind the gate that sets where it listens and
+	// where its files are.
+	"comfyui": {
+		"--listen": "is set by the agent", "--base-directory": "is set by the agent",
+		"--extra-model-paths-config": "is set by the agent",
+		"--output-directory":         "writes files", "--temp-directory": "writes files",
+		"--input-directory": "reads files", "--user-directory": "writes files",
+		"--database-url": "writes a database elsewhere", "--front-end-root": "serves a directory",
+		"--enable-cors-header": "lets other sites' pages call it",
+		"--enable-manager":     "installs nodes, which run Python",
+		"--tls-keyfile":        "reads a file", "--tls-certfile": "reads a file",
+		"--comfy-api-base": "sends work to another service",
+	},
+	// Strata's setup.py, which a profile's arguments go to: llmgw
+	// strata-serve sets where the model is, where its files go and where
+	// the server listens, and the engine is the one the agent installed.
+	// Its own --model, the size, is spelt --size in a profile, since
+	// --model is the source everywhere else.
+	"strata": {
+		"--gguf-dir": "is the model's source, set by the agent", "--models-dir": "is set by the agent",
+		"--data-dir": "is set by the agent", "--move-data": "moves files", "--set-default-data-dir": "writes settings",
+		"--feed": "downloads an engine", "--prebuilt": "loads an engine", "--build": "compiles an engine",
+		"--update": "downloads an engine", "--rollback-engine": "changes the engine",
+		"--dsv41-package": "loads a file", "--inspect": "reads another repository or path",
+		"--source": "downloads from elsewhere", "--setup": "is not for a server",
+		"--browser": "opens a browser", "--gpu": "is set by the agent", "--gpus": "is set by the agent",
+		"--gpu-roles": "is set by the agent", "--backend": "is set by the agent",
+		"--experimental-speed-projection": "changes what the model refuses",
+	},
+	// ft serve: where it listens, its key, its model and its card are the
+	// agent's, and what reads files or fetches is refused.
+	"freetoken": {
+		"--gpu": "is set by the agent", "--served-model-name": "is set by the agent",
+		"--mmproj": "loads a file", "--mm-encoder-weights": "loads a file",
+		"--allowed-local-media-path": "serves local files to requests",
+		"--allowed-media-domains":    "lets requests have the engine fetch from those sites",
+		"--moe-cache-split":          "reads a profile file",
+		"--hf-overrides":             "rewrites the checkpoint's configuration, which can name code to load",
 	},
 	// trtllm-serve spells its flags with underscores, so the shared list's
 	// hyphenated names do not reach them.
@@ -150,14 +209,43 @@ var flagsNever = map[string]map[string]string{
 // unsafeFlag is why an engine may not be given a flag, or "".
 func unsafeFlag(engine, flag, _ string) string {
 	name, _, _ := strings.Cut(flag, "=")
+	if !strings.HasPrefix(name, "-") {
+		return ""
+	}
+	// Written as the engine would read it: vLLM takes --load_format for
+	// --load-format, TensorRT-LLM the other way round. Short flags (-m,
+	// -hf) are matched as written.
+	norm := func(f string) string {
+		if strings.HasPrefix(f, "--") {
+			return "--" + strings.ReplaceAll(f[2:], "_", "-")
+		}
+		return f
+	}
+	given := norm(name)
+	// argparse also takes any unambiguous beginning of a long flag for the
+	// flag — --trust-remote is --trust-remote-code, --li is ComfyUI's
+	// --listen — so for the engines that parse with it, a beginning of a
+	// refused flag is refused. llama.cpp and TensorRT-LLM take no such
+	// thing, and --chat-template there is a flag of its own.
+	abbreviates := argparseEngines[engine] && strings.HasPrefix(given, "--") && len(given) >= 4
 	for _, set := range []string{"", engine} {
-		if why, ok := flagsNever[set][name]; ok && why != "" {
-			return why
+		for never, why := range flagsNever[set] {
+			if why == "" {
+				continue
+			}
+			n := norm(never)
+			if n == given || (abbreviates && strings.HasPrefix(n, "--") && strings.HasPrefix(n, given)) {
+				return why
+			}
 		}
 	}
 	return ""
 }
 
+// argparseEngines read their flags with Python's argparse, which takes a
+// long flag's beginning for the flag.
+var argparseEngines = map[string]bool{"vllm": true, "sglang": true, "comfyui": true, "strata": true, "freetoken": true}
+
 // UnsafeArgs refuses the first of a profile's arguments an engine may not be
 // given, saying why.
 func UnsafeArgs(engine string, args []string) error {
@@ -172,3 +260,29 @@ func UnsafeArgs(engine string, args []string) error {
 	}
 	return nil
 }
+
+// safetensorsOnly has vLLM or SGLang read only safetensors from a model it
+// downloads itself, by its Hugging Face id: left to choose, either takes
+// the repository's pickle files (pytorch_model.bin) when it has no
+// safetensors, and loading those runs whatever they hold. A path was
+// fetched by the agent, which checked it, or put there by the operator.
+func safetensorsOnly(p Profile) []string {
+	if p.AllowPickle || !hubID(p.ModelPath) {
+		return nil
+	}
+	return []string{"--load-format", "safetensors"}
+}
+
+// hubID reports whether a model path is a Hugging Face id, org/name, rather
+// than a file or directory.
+func hubID(s string) bool {
+	if s == "" || strings.HasPrefix(s, "/") || strings.HasPrefix(s, ".") || strings.Contains(s, "://") {
+		return false
+	}
+	// org/name:Q4_K_M is a GGUF in a repository, which vLLM reads as one.
+	if strings.Count(s, "/") != 1 || strings.Contains(s, ":") {
+		return false
+	}
+	_, err := os.Stat(s)
+	return err != nil
+}

internal/agent/engines/vllm.go

@@ -119,6 +119,7 @@ func (v *VLLM) serverCommand(s ShardSpec) (Command, error) {
 	if p.ServedModelName != "" {
 		args = append(args, "--served-model-name", p.ServedModelName)
 	}
+	args = append(args, safetensorsOnly(p)...)
 
 	env := baseEnv(s)
 

internal/agent/gpucheck.go

@@ -20,6 +20,7 @@ var (
 	gpuProcesses = discovery.GPUProcesses
 	gpuMemory    = discovery.GPUMemoryByProcess
 	processGroup = pgidOf
+	parentOf     = ppidOf
 	// 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
@@ -116,19 +117,44 @@ func (a *Agent) foreignGPUMemory(ctx context.Context, shards []int) (foreign, us
 	return foreign, used
 }
 
-// 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.
+// holds reports whether any of pids is the shard's process, in its group,
+// or a process it started: an engine that computes in a child process is
+// still the shard, and Strata's server starts its engine in a group of the
+// engine's own.
 func holds(pids []int, pid, group int) bool {
 	for _, p := range pids {
-		if p == pid || (group > 0 && processGroup(p) == group) {
+		if p == pid || (group > 0 && processGroup(p) == group) || descends(p, pid) {
 			return true
 		}
 	}
 	return false
 }
 
+// descends reports whether process p was started, at any remove, by pid.
+func descends(p, pid int) bool {
+	// Bounded: a chain this deep is no engine's, and a /proc read racing an
+	// exit must not loop.
+	for range 16 {
+		p = parentOf(p)
+		switch {
+		case p == pid:
+			return true
+		case p <= 1:
+			return false
+		}
+	}
+	return false
+}
+
 // pgidOf reads a process's group from /proc; zero when it cannot.
-func pgidOf(pid int) int {
+func pgidOf(pid int) int { return statField(pid, 2) }
+
+// ppidOf reads a process's parent from /proc; zero when it cannot.
+func ppidOf(pid int) int { return statField(pid, 1) }
+
+// statField is one of the numbers after the command name in /proc's stat
+// line: 1 is the parent, 2 the group.
+func statField(pid, field int) int {
 	b, err := os.ReadFile("/proc/" + strconv.Itoa(pid) + "/stat")
 	if err != nil {
 		return 0
@@ -141,10 +167,10 @@ func pgidOf(pid int) int {
 		return 0
 	}
 	f := strings.Fields(s[i+1:])
-	if len(f) < 3 {
+	if len(f) <= field {
 		return 0
 	}
-	n, _ := strconv.Atoi(f[2])
+	n, _ := strconv.Atoi(f[field])
 	return n
 }
 

internal/agent/gpucheck_test.go

@@ -123,6 +123,30 @@ func TestForeignGPUMemoryLeavesOutTheAgentsOwnShards(t *testing.T) {
 	}
 }
 
+func TestAnEngineStartedInAGroupOfItsOwnIsStillTheShards(t *testing.T) {
+	// cuda01: Strata's server (259709) starts its engine (261072) in a
+	// group of the engine's own, and the engine holds the card. It read as
+	// a shard on the CPU, and as 22 GB of somebody else's memory.
+	oldMem, oldGroup, oldParent := gpuMemory, processGroup, parentOf
+	t.Cleanup(func() { gpuMemory, processGroup, parentOf = oldMem, oldGroup, oldParent })
+	processGroup = func(pid int) int { return pid }
+	parentOf = func(pid int) int {
+		return map[int]int{261072: 259709, 259709: 258734, 258734: 1, 999: 1}[pid]
+	}
+	gpuMemory = func(context.Context, string) ([]discovery.GPUMemory, error) {
+		return []discovery.GPUMemory{{Index: 0, UsedMB: 23000,
+			Procs: []discovery.GPUProcess{{PID: 261072, UsedMB: 22226}, {PID: 999, UsedMB: 700}}}}, nil
+	}
+	a := &Agent{log: slog.Default()}
+	if c, ok := a.gpuCompute(context.Background(), []int{0}, 259709); !ok || c.Backend != "CUDA" ||
+		len(c.Devices) != 1 || c.Devices[0] != "GPU 0: 21.7 GB" {
+		t.Errorf("the engine's child on GPU 0: %+v %v", c, ok)
+	}
+	if foreign, _ := a.foreignGPUMemory(context.Background(), []int{259709}); !slices.Equal(foreign, []int32{774}) {
+		t.Errorf("foreign %v; want only what the other process and nobody holds, [774]", foreign)
+	}
+}
+
 func TestLlamaCppsOwnWordOnTheCPUIsEnoughToRefuse(t *testing.T) {
 	eng := fallbackEngine{}
 	// Without nvidia-smi to ask, the engine's own output says it.

internal/agent/main_test.go

@@ -0,0 +1,17 @@
+package agent
+
+import (
+	"os"
+	"testing"
+	"time"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/discovery"
+)
+
+func TestMain(m *testing.M) {
+	// Half a gigabyte read per test binary, and contending with the tests
+	// for the CPUs: these never look at the figure.
+	measureBandwidth = func() discovery.Bandwidth { return discovery.Bandwidth{} }
+	bandwidthDelay = func() time.Duration { return 0 }
+	os.Exit(m.Run())
+}

internal/agent/provision/comfyui_test.go

@@ -0,0 +1,177 @@
+package provision
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/discovery"
+)
+
+// comfyuiSource is a ComfyUI release archive as GitHub makes one, and the
+// recipe pointed at it.
+func comfyuiSource(t *testing.T, in *Installer, rel *release, digest string) {
+	t.Helper()
+	archive := tarGz(t, map[string]string{
+		"ComfyUI-0.39.2/main.py":            "print('comfy')\n",
+		"ComfyUI-0.39.2/comfyui_version.py": "__version__ = \"0.39.2\"\n",
+		"ComfyUI-0.39.2/requirements.txt": "comfyui-frontend-package==1.53.10\ntorch\ntorchvision\n" +
+			"numpy>=1.25.0\n\n#non essential dependencies:\nkornia>=0.7.1\n",
+	}, nil)
+	rel.assets["v0.39.2.tar.gz"] = archive
+	sum := sha256.Sum256(archive)
+	if digest == "" {
+		digest = hex.EncodeToString(sum[:])
+	}
+	old := comfyuiReleases
+	comfyuiReleases = map[string]struct {
+		sha256    string
+		published string
+	}{"0.39.2": {digest, "2026-10-07T21:58:07Z"}}
+	t.Cleanup(func() { comfyuiReleases = old })
+}
+
+var turingNode = discovery.Host{OSID: "ubuntu", OSVersion: "24.04", Arch: "amd64",
+	NVIDIADriver: "580.82.07", MinComputeCapability: 7.5}
+
+func TestComfyUIIsInstalledFromItsSourceCheckedAgainstTheRecipe(t *testing.T) {
+	sys := &fakePython{fakeSystem: fakeSystem{installed: map[string]bool{}}}
+	in, rel := pythonInstaller(t, turingNode, sys)
+	comfyuiSource(t, in, rel, "")
+
+	got := in.Converge(t.Context(), []Want{{Engine: "comfyui", Version: "0.39.2"}})
+	if got[0].State != Ready {
+		t.Fatalf("not ready: %+v", got[0])
+	}
+	dir := filepath.Join(in.Root, "comfyui", "0.39.2-cu13")
+	if got[0].Binary != filepath.Join(dir, "bin", "python") {
+		t.Errorf("binary %s, want the environment's interpreter", got[0].Binary)
+	}
+	if _, err := os.Stat(filepath.Join(dir, "comfyui", "main.py")); err != nil {
+		t.Errorf("ComfyUI's code is not beside the interpreter: %v", err)
+	}
+	all := strings.Join(sys.commands, "\n")
+	for _, want := range []string{
+		// Not PyPI's date, which ComfyUI has none of: the recipe's, a day on.
+		"--exclude-newer 2026-10-08T21:58:07Z",
+		"comfyui-frontend-package==1.53.10", "numpy>=1.25.0", "kornia>=0.7.1", "torch",
+		"--require-hashes",
+	} {
+		if !strings.Contains(all, want) {
+			t.Errorf("no %q in:\n%s", want, all)
+		}
+	}
+	if strings.Contains(all, "/pypi/") {
+		t.Error("PyPI was asked when ComfyUI was published")
+	}
+}
+
+func TestComfyUIOnAnOlderDriverTakesPyTorchsCUDA12Torch(t *testing.T) {
+	sys := &fakePython{fakeSystem: fakeSystem{installed: map[string]bool{}}}
+	host := turingNode
+	host.NVIDIADriver = "550.163.01"
+	in, rel := pythonInstaller(t, host, sys)
+	comfyuiSource(t, in, rel, "")
+	got := in.Converge(t.Context(), []Want{{Engine: "comfyui", Version: "0.39.2"}})
+	if got[0].State != Ready || got[0].Variant != "cu129" {
+		t.Fatalf("got %+v", got[0])
+	}
+	var project string
+	for _, c := range sys.commands {
+		if strings.HasPrefix(c, "pyproject: ") {
+			project = c
+		}
+	}
+	// torch from PyTorch's index alone, named once, as the build names it.
+	if !strings.Contains(project, `torch = { index = "extra" }`) || strings.Count(project, `"torch"`) != 1 {
+		t.Errorf("torch is not from PyTorch's index alone:\n%s", project)
+	}
+}
+
+// A V100 on driver 580: CUDA 13 runs on the driver, but neither its torch
+// nor CUDA 12.9's has kernels for Volta. Tried in turn, each was installed,
+// 10 GB at a time, and failed on the card.
+func TestComfyUIOnAVoltaTakesPyTorchsCUDA126Torch(t *testing.T) {
+	sys := &fakePython{fakeSystem: fakeSystem{installed: map[string]bool{}}}
+	host := turingNode
+	host.MinComputeCapability = 7.0
+	in, rel := pythonInstaller(t, host, sys)
+	comfyuiSource(t, in, rel, "")
+	got := in.Converge(t.Context(), []Want{{Engine: "comfyui", Version: "0.39.2"}})
+	if got[0].State != Ready || got[0].Variant != "cu126" {
+		t.Fatalf("got %+v", got[0])
+	}
+	all := strings.Join(sys.commands, "\n")
+	if !strings.Contains(all, pytorchCU126) || strings.Contains(all, pytorchCU129) {
+		t.Errorf("torch is not CUDA 12.6's:\n%s", all)
+	}
+}
+
+func TestComfyUIWhoseSourceIsNotTheOneTheRecipeNamesIsNotInstalled(t *testing.T) {
+	sys := &fakePython{fakeSystem: fakeSystem{installed: map[string]bool{}}}
+	in, rel := pythonInstaller(t, turingNode, sys)
+	comfyuiSource(t, in, rel, strings.Repeat("ab", 32))
+	got := in.Converge(t.Context(), []Want{{Engine: "comfyui", Version: "0.39.2"}})
+	if got[0].State != Failed || !strings.Contains(got[0].Detail, "does not match") {
+		t.Fatalf("got %s %q", got[0].State, got[0].Detail)
+	}
+	for _, c := range sys.commands {
+		if strings.Contains(c, "pip install") {
+			t.Errorf("installed from a source that did not match: %s", c)
+		}
+	}
+}
+
+func TestComfyUIIsInstalledOnlyWhereItCanRun(t *testing.T) {
+	sys := &fakePython{fakeSystem: fakeSystem{installed: map[string]bool{}}}
+	in, _ := pythonInstaller(t, ubuntu, sys)
+	if got := in.Converge(t.Context(), []Want{{Engine: "comfyui", Version: "0.39.2"}}); got[0].State != Skipped {
+		t.Errorf("a node with no GPU: %+v", got[0])
+	}
+	in, _ = pythonInstaller(t, turingNode, sys)
+	got := in.Converge(t.Context(), []Want{{Engine: "comfyui", Version: "0.1.0"}})
+	if got[0].State != Unsupported || !strings.Contains(got[0].Detail, "knows the source digest of ComfyUI 0.39.2") {
+		t.Errorf("a version without a digest: %s %q", got[0].State, got[0].Detail)
+	}
+}
+
+func TestARequirementsFileIsReadForPackagesAndNothingElse(t *testing.T) {
+	got, err := requirementsIn("torch\ntorchvision>=0.20\nPillow\n# a comment\naiohttp>=3.11 # inline\n", pytorchPackages)
+	if err != nil || strings.Join(got, ",") != "Pillow,aiohttp>=3.11" {
+		t.Errorf("%v %v", got, err)
+	}
+	for _, bad := range []string{"-r other.txt", "--index-url https://evil", "x @ https://evil/x.whl", "-e ."} {
+		if _, err := requirementsIn(bad+"\n", nil); err == nil {
+			t.Errorf("%q was accepted", bad)
+		}
+	}
+}
+
+// A card every build runs on keeps the order it had: a Turing node's
+// installed cu13 is still the first found, and is not installed again
+// because a build for older cards was added after it.
+func TestComfyUIBuildsAreTriedBestFirst(t *testing.T) {
+	variants := func(h discovery.Host) string {
+		builds, err := comfyuiCandidates("0.39.2", "", h)
+		if err != nil {
+			return err.Error()
+		}
+		var v []string
+		for _, b := range builds {
+			v = append(v, b.Variant)
+		}
+		return strings.Join(v, ",")
+	}
+	volta := turingNode
+	volta.MinComputeCapability = 7.0
+	older := turingNode
+	older.NVIDIADriver = "550.163.01"
+	for h, want := range map[*discovery.Host]string{&turingNode: "cu13,cu129,cu126", &volta: "cu126", &older: "cu129,cu126"} {
+		if got := variants(*h); got != want {
+			t.Errorf("cc %.1f driver %s: %s, want %s", h.MinComputeCapability, h.NVIDIADriver, got, want)
+		}
+	}
+}

internal/agent/provision/freetoken_test.go

@@ -0,0 +1,53 @@
+package provision
+
+import (
+	"errors"
+	"slices"
+	"strings"
+	"testing"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/discovery"
+)
+
+func TestFreeTokenIsInstalledOnlyWhereItsKernelsRun(t *testing.T) {
+	old := freetokenToolkit
+	toolkit := "/usr/local/cuda-13.4"
+	freetokenToolkit = func() string { return toolkit }
+	t.Cleanup(func() { freetokenToolkit = old })
+
+	ampere := discovery.Host{Arch: "amd64", NVIDIADriver: "580.95.05", MinComputeCapability: 8.6}
+	got, err := freetokenCandidates("0.2.1", "", ampere)
+	if err != nil {
+		t.Fatal(err)
+	}
+	b := got[0]
+	if b.Binary != "ft" || !slices.Equal(b.Requirements, []string{"freetoken[accel]==0.2.1"}) || !b.GPUCheck || b.Python == "" {
+		t.Errorf("build %+v", b)
+	}
+
+	// Not meant for these, and not reported as a fault: no GPU, a Turing
+	// card (cuda01, where 0.2.1 had no kernel image for it), an Arm node.
+	for name, h := range map[string]discovery.Host{
+		"no gpu":  {Arch: "amd64"},
+		"turing":  {Arch: "amd64", NVIDIADriver: "615.71.09", MinComputeCapability: 7.5},
+		"aarch64": {Arch: "arm64", NVIDIADriver: "580.95.05", MinComputeCapability: 9.0},
+	} {
+		_, err := freetokenCandidates("0.2.1", "", h)
+		var nh NotHere
+		if !errors.As(err, &nh) {
+			t.Errorf("%s: %v, want NotHere", name, err)
+		}
+	}
+
+	// Meant for these, and said why not.
+	if _, err := freetokenCandidates("0.2.1", "", discovery.Host{Arch: "amd64", NVIDIADriver: "570.1", MinComputeCapability: 9.0}); err == nil ||
+		!strings.Contains(err.Error(), "driver 580") {
+		t.Errorf("old driver: %v", err)
+	}
+	toolkit = ""
+	_, err = freetokenCandidates("0.2.1", "", ampere)
+	var nh NotHere
+	if err == nil || errors.As(err, &nh) || !strings.Contains(err.Error(), "CUDA 13 toolkit") {
+		t.Errorf("no toolkit: %v", err)
+	}
+}

internal/agent/provision/install.go

@@ -174,7 +174,7 @@ func (in *Installer) one(ctx context.Context, w Want) Status {
 		}
 		return st
 	}
-	if w.Offline != nil && recipe.Repo == "" {
+	if w.Offline != nil && (recipe.Repo == "" || builds[0].Python != "") {
 		// Said, not attempted: uv would spend minutes failing to reach PyPI,
 		// and the failure would read like a network fault rather than the
 		// design of an offline site.

internal/agent/provision/python.go

@@ -66,12 +66,17 @@ func (in *Installer) installPython(ctx context.Context, r Recipe, w Want, b Buil
 		return fmt.Errorf("installing uv, which installs %s: %s", w.Engine, uv.Detail)
 	}
 
-	in.progress(st, "asking PyPI when %s %s was published", r.PyPI, w.Version)
-	published, err := in.published(ctx, r.PyPI, w.Version)
-	if err != nil {
-		return err
+	var asOf string
+	if !b.AsOf.IsZero() {
+		asOf = b.AsOf.UTC().Format(time.RFC3339)
+	} else {
+		in.progress(st, "asking PyPI when %s %s was published", r.PyPI, w.Version)
+		published, err := in.published(ctx, r.PyPI, w.Version)
+		if err != nil {
+			return err
+		}
+		asOf = published.Add(24 * time.Hour).UTC().Format(time.RFC3339)
 	}
-	asOf := published.Add(24 * time.Hour).UTC().Format(time.RFC3339)
 
 	cache := filepath.Join(in.root(), ".uv-cache")
 	env := []string{
@@ -96,7 +101,19 @@ func (in *Installer) installPython(ctx context.Context, r Recipe, w Want, b Buil
 	}
 	python := filepath.Join(tmp, "bin", "python")
 
-	reqs := b.Requirements
+	reqs := slices.Clone(b.Requirements)
+	if b.Source != nil {
+		more, err := in.unpackSource(ctx, w, b, tmp, st)
+		if err != nil {
+			return err
+		}
+		reqs = append(reqs, more...)
+	}
+	if len(b.Assets) > 0 {
+		if err := in.unpackAssets(ctx, r, w, b, tmp, st); err != nil {
+			return err
+		}
+	}
 	if b.Wheel != nil {
 		// The engine from its GitHub release, pinned to the digest GitHub
 		// publishes for the file: the same trust as an archive, and uv
@@ -320,3 +337,110 @@ func (in *Installer) gpuCheck(ctx context.Context, python string, w Want, b Buil
 	}
 	return nil
 }
+
+// unpackSource fetches an engine's source archive, checks it against the
+// digest the recipe holds, puts its code in the environment, and returns the
+// requirements its requirements file names — less the packages a second
+// index gives, which the build names itself.
+func (in *Installer) unpackSource(ctx context.Context, w Want, b Build, env string, st *Status) ([]string, error) {
+	src := b.Source
+	url := src.URL
+	if strings.HasPrefix(url, "/") {
+		url = strings.TrimSuffix(in.downloadBase(), "/") + url
+	}
+	scratch := env + ".source"
+	_ = os.RemoveAll(scratch)
+	if err := os.MkdirAll(scratch, 0o755); err != nil {
+		return nil, err
+	}
+	defer os.RemoveAll(scratch) //nolint:errcheck // scratch
+	archive := filepath.Join(scratch, "source.tar.gz")
+	name := w.Engine + " " + w.Version + "'s source"
+	if err := in.download(ctx, url, name, archive, src.SHA256, st); err != nil {
+		return nil, err
+	}
+	in.progress(st, "unpacking %s", name)
+	if err := untarGz(archive, scratch); err != nil {
+		return nil, fmt.Errorf("unpacking %s: %w", name, err)
+	}
+	dst := filepath.Join(env, src.Dir)
+	if err := os.Rename(filepath.Join(scratch, src.Top), dst); err != nil {
+		return nil, fmt.Errorf("%s has no %s at its top: %w", name, src.Top, err)
+	}
+	if src.Requirements == "" {
+		return nil, nil
+	}
+	f, err := os.ReadFile(filepath.Join(dst, src.Requirements))
+	if err != nil {
+		return nil, fmt.Errorf("%s: %w", name, err)
+	}
+	return requirementsIn(string(f), b.IndexPackages)
+}
+
+// unpackAssets fetches a Python engine's release binaries, each checked
+// against the digest GitHub publishes for it, and unpacks them into the
+// environment's AssetDir.
+func (in *Installer) unpackAssets(ctx context.Context, r Recipe, w Want, b Build, env string, st *Status) error {
+	tag := b.AssetTag
+	if tag == "" {
+		tag = w.Version
+	}
+	in.progress(st, "asking GitHub for the checksums of %s %s", r.Repo, tag)
+	sums, err := in.checksums(ctx, r.Repo, tag)
+	if err != nil {
+		return err
+	}
+	dst := filepath.Join(env, b.AssetDir)
+	if err := os.MkdirAll(dst, 0o755); err != nil {
+		return err
+	}
+	for _, asset := range b.Assets {
+		want, ok := sums[asset]
+		if !ok {
+			// Refused rather than installed unchecked: this is a binary that
+			// will run on every node.
+			return fmt.Errorf("%s %s publishes no checksum for %s, so it cannot be verified", r.Repo, tag, asset)
+		}
+		url := strings.TrimSuffix(in.downloadBase(), "/") + "/" + r.Repo + "/releases/download/" + tag + "/" + asset
+		archive := filepath.Join(env+".source", asset)
+		if err := os.MkdirAll(filepath.Dir(archive), 0o755); err != nil {
+			return err
+		}
+		if err := in.download(ctx, url, asset, archive, want, st); err != nil {
+			return err
+		}
+		in.progress(st, "unpacking %s", asset)
+		err := untarGz(archive, dst)
+		_ = os.RemoveAll(filepath.Dir(archive))
+		if err != nil {
+			return fmt.Errorf("unpacking %s: %w", asset, err)
+		}
+	}
+	return nil
+}
+
+// requirementsIn reads a requirements file's lines. Options (-r, -e, --index-url
+// and the like) are refused: they would fetch from somewhere other than PyPI,
+// or install what the lock never saw.
+func requirementsIn(text string, skip []string) ([]string, error) {
+	var out []string
+	for _, line := range strings.Split(text, "\n") {
+		line, _, _ = strings.Cut(line, "#")
+		line = strings.TrimSpace(line)
+		if line == "" {
+			continue
+		}
+		if strings.HasPrefix(line, "-") || strings.Contains(line, "://") || strings.Contains(line, " @ ") {
+			return nil, fmt.Errorf("the requirements name %q, which is not a package from PyPI", line)
+		}
+		name := strings.ToLower(line)
+		if i := strings.IndexAny(name, "<>=!~;[ "); i >= 0 {
+			name = name[:i]
+		}
+		if slices.Contains(skip, name) {
+			continue
+		}
+		out = append(out, line)
+	}
+	return out, nil
+}

internal/agent/provision/python_test.go

@@ -53,6 +53,7 @@ func (f *fakePython) run(ctx context.Context, env []string, name string, args ..
 			return []byte("error: Hash mismatch for torch\n"), errExit
 		}
 		venv := filepath.Dir(filepath.Dir(args[slices.Index(args, "--python")+1]))
+		_ = os.WriteFile(filepath.Join(venv, "bin", "python"), []byte("#!/bin/sh\n"), 0o755)
 		return nil, os.WriteFile(filepath.Join(venv, "bin", "vllm"), []byte("#!/bin/sh\n"), 0o755)
 	case strings.HasSuffix(name, "/bin/vllm") && args[0] == "--version":
 		return []byte("0.30.0\n"), nil

internal/agent/provision/recipe.go

@@ -13,6 +13,7 @@ import (
 	"fmt"
 	"slices"
 	"strings"
+	"time"
 
 	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/discovery"
 )
@@ -68,12 +69,36 @@ type Build struct {
 	// Note says why this build, when it is not the plain one: shown beside
 	// it wherever the installation is.
 	Note string
+	// Source, for a Python engine that is not on PyPI, is its own code: an
+	// archive pinned by its digest, unpacked into the environment, whose
+	// requirements file is installed with Requirements. AsOf is then the
+	// date its dependencies are resolved as of — the day after its version
+	// was published, which PyPI cannot be asked for.
+	Source *Source
+	AsOf   time.Time
+	// AssetTag and AssetDir, for a Python engine that also ships a binary
+	// in its releases, are the release tag Assets are taken from, checked
+	// against the digests GitHub publishes, and the directory in the
+	// environment they are unpacked into.
+	AssetTag, AssetDir string
 	// FreeGB is the disk space the installation needs before it starts: a
 	// Python engine is several gigabytes, and one that fills the disk halfway
 	// leaves the node worse off than one that never began.
 	FreeGB int
 }
 
+// Source is an engine's code as an archive.
+type Source struct {
+	// URL is fetched as is; SHA256 is what it must hash to, written into
+	// the recipe because a project's source archives come with no published
+	// digest of their own.
+	URL, SHA256 string
+	// Top is the archive's one top directory, put in the environment as Dir.
+	Top, Dir string
+	// Requirements is the requirements file, in Dir.
+	Requirements string
+}
+
 // Wheel is a package published as a GitHub release asset: the asset whose
 // name starts with Prefix and ends with Suffix.
 type Wheel struct {
@@ -101,10 +126,13 @@ type Recipe struct {
 // from PyPI by uv, which is itself installed like llama.cpp: a GitHub release
 // checked against its published checksum.
 var Recipes = map[string]Recipe{
-	"llamacpp": {Repo: "ggml-org/llama.cpp", Candidates: llamaCppCandidates},
-	"vllm":     {PyPI: "vllm", Candidates: vllmCandidates},
-	"sglang":   {PyPI: "sglang", Candidates: sglangCandidates},
-	"uv":       {Repo: "astral-sh/uv", Candidates: uvCandidates},
+	"llamacpp":  {Repo: "ggml-org/llama.cpp", Candidates: llamaCppCandidates},
+	"vllm":      {PyPI: "vllm", Candidates: vllmCandidates},
+	"sglang":    {PyPI: "sglang", Candidates: sglangCandidates},
+	"comfyui":   {Candidates: comfyuiCandidates},
+	"strata":    {Repo: StrataRepo, Candidates: strataCandidates},
+	"freetoken": {PyPI: "freetoken", Candidates: freetokenCandidates},
+	"uv":        {Repo: "astral-sh/uv", Candidates: uvCandidates},
 }
 
 // Installable are the engines this agent can install, as it tells the
@@ -151,6 +179,9 @@ func Builds(engine, version, arch string) ([]Build, error) {
 			continue
 		}
 		for _, b := range bs {
+			if b.Python != "" {
+				return nil, fmt.Errorf("%s is a Python environment from PyPI, which an offline bundle does not carry", engine)
+			}
 			if !seen[b.Variant] {
 				seen[b.Variant] = true
 				out = append(out, b)
@@ -556,3 +587,228 @@ func orUnknown(s string) string {
 	}
 	return s
 }
+
+// --- ComfyUI ---
+
+// comfyuiReleases are the ComfyUI versions this agent installs: each source
+// archive's digest, taken when it was added here and confirmed from a second
+// machine, and when the version was published. GitHub publishes no digest
+// for a tag's archive, so a version not listed is refused rather than
+// installed unchecked.
+var comfyuiReleases = map[string]struct {
+	sha256    string
+	published string
+}{
+	"0.39.2": {"1d655e22cf546700fc1d6d2536e27701695c5832a51f63bb16c41866ac09be0b", "2026-10-07T21:58:07Z"},
+}
+
+// ComfyUIVersion is the newest ComfyUI this agent installs: what a cluster
+// is set to when a model first needs it.
+func ComfyUIVersion() string {
+	newest := ""
+	for v := range comfyuiReleases {
+		if newest == "" || versionBefore(newest, v) {
+			newest = v
+		}
+	}
+	return newest
+}
+
+// ComfyUIRepo is where ComfyUI's source is published.
+const ComfyUIRepo = "Comfy-Org/ComfyUI"
+
+// comfyuiDir is where ComfyUI's code is put in its environment: main.py is
+// <environment>/comfyui/main.py, beside bin/python.
+const comfyuiDir = "comfyui"
+
+// pytorchCU126 is PyTorch's CUDA 12.6 build: the last that carries kernels
+// for Volta and Pascal. From torch 2.8 its CUDA 12.8 and 12.9 builds start
+// at Turing (7.5), as CUDA 13 does, so a V100 has no torch but this one.
+const pytorchCU126 = "https://download.pytorch.org/whl/cu126"
+
+// comfyuiCandidates are the ComfyUI builds for a machine: its source, in a
+// Python environment with the torch for the node's driver and cards —
+// PyPI's, built for CUDA 13; PyTorch's CUDA 12.9 build for a driver CUDA 13
+// does not run on; and its CUDA 12.6 build for cards older than Turing.
+func comfyuiCandidates(version, variant string, h discovery.Host) ([]Build, error) {
+	if _, err := gpuArch("ComfyUI", h); err != nil {
+		return nil, err
+	}
+	rel, ok := comfyuiReleases[version]
+	if !ok {
+		known := make([]string, 0, len(comfyuiReleases))
+		for v := range comfyuiReleases {
+			known = append(known, v)
+		}
+		slices.Sort(known)
+		return nil, fmt.Errorf("this agent knows the source digest of ComfyUI %s, not of %s; "+
+			"install that one by hand and name it in agent.yaml (engines.comfyui.dir)",
+			strings.Join(known, ", "), version)
+	}
+	published, err := time.Parse(time.RFC3339, rel.published)
+	if err != nil {
+		return nil, err
+	}
+	build := func(variant string) Build {
+		return Build{
+			Variant: variant,
+			Python:  pythonForEngines,
+			Source: &Source{
+				URL:    "/" + ComfyUIRepo + "/archive/refs/tags/v" + version + ".tar.gz",
+				SHA256: rel.sha256, Top: "ComfyUI-" + version, Dir: comfyuiDir,
+				Requirements: "requirements.txt",
+			},
+			AsOf: published.Add(24 * time.Hour),
+			// ComfyUI is started as a script, so what the agent is given is
+			// the environment's interpreter; the proof is importing what it
+			// imports first, from where its code is.
+			Binary: "python",
+			VersionArgs: []string{"-c", `import os, sys, torch, aiohttp, safetensors
+d = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(sys.executable))), "` + comfyuiDir + `")
+assert os.path.isfile(os.path.join(d, "main.py")), d + " has no main.py"
+print("comfyui version", "` + version + `", "on torch", torch.__version__)`},
+			Packages: compilerPackages,
+			FreeGB:   15,
+			GPUCheck: true,
+		}
+	}
+	cu129 := build("cu129")
+	cu129.ExtraIndex, cu129.IndexPackages = pytorchCU129, pytorchPackages
+	cu126 := build("cu126")
+	cu126.ExtraIndex, cu126.IndexPackages = pytorchCU126, pytorchPackages
+	return pick("ComfyUI "+version, variant, []candidate{
+		{build("cu13"), gpuNeeds{minDriverForCUDA13, 7.5}, "ComfyUI " + version + " on CUDA 13's torch"},
+		{cu129, gpuNeeds{minDriverForCUDA12, 7.5}, "ComfyUI " + version + " on CUDA 12.9's torch"},
+		{cu126, gpuNeeds{minDriverForCUDA12, 5.0}, "ComfyUI " + version + " on CUDA 12.6's torch"},
+	}, h)
+}
+
+// --- Strata ---
+
+// StrataRepo is where Strata is published: its source, and in each release
+// the prebuilt engine for each CUDA.
+const StrataRepo = "Niko1221/Strata"
+
+// strataReleases are the Strata versions this agent installs: the source
+// archive's digest, taken when it was added here and confirmed from a second
+// machine, when the version was published, and the engine its release
+// carries, whose files GitHub publishes digests for.
+var strataReleases = map[string]struct {
+	sha256    string
+	published string
+	engine    string
+}{
+	"0.2.7": {"41ae861478f5e5d9506864f97dfb5dd7cfe6b83193300e2b0d4f04c0380e3af9", "2026-10-27T14:54:42Z", "0.1.52"},
+}
+
+// strataCandidates are the Strata builds for a machine: its source, in a
+// Python environment, with the prebuilt engine for the node's driver and
+// cards — built for CUDA 13, Turing to Blackwell; or for CUDA 12.9, Pascal
+// to Ada, which is what a V100, or a driver older than 580, gets.
+func strataCandidates(version, variant string, h discovery.Host) ([]Build, error) {
+	if _, err := gpuArch("Strata", h); err != nil {
+		return nil, err
+	}
+	if h.Arch != "amd64" {
+		return nil, NotHere{"Strata publishes its Linux CUDA engine for x86-64 only"}
+	}
+	rel, ok := strataReleases[version]
+	if !ok {
+		known := make([]string, 0, len(strataReleases))
+		for v := range strataReleases {
+			known = append(known, v)
+		}
+		slices.Sort(known)
+		return nil, fmt.Errorf("this agent knows the source digest of Strata %s, not of %s; "+
+			"install that one by hand and name it in agent.yaml (engines.strata.dir)",
+			strings.Join(known, ", "), version)
+	}
+	published, err := time.Parse(time.RFC3339, rel.published)
+	if err != nil {
+		return nil, err
+	}
+	build := func(cuda string) Build {
+		return Build{
+			Variant: cuda,
+			Python:  pythonForEngines,
+			Source: &Source{
+				URL:    "/" + StrataRepo + "/archive/refs/tags/v" + version + ".tar.gz",
+				SHA256: rel.sha256, Top: "Strata-" + version, Dir: "strata",
+				Requirements: "requirements.txt",
+			},
+			AsOf:     published.Add(24 * time.Hour),
+			Assets:   []string{"strata-engine-" + rel.engine + "-linux-x86_64-cuda-" + cuda + ".tar.gz"},
+			AssetTag: "v" + version,
+			AssetDir: "strata-engine",
+			// Strata is started as a script, so what the agent is given is
+			// the environment's interpreter; the proof is importing what its
+			// server needs, and finding its engine beside it.
+			Binary: "python",
+			VersionArgs: []string{"-c", `import os, sys, numpy, jinja2, regex
+e = os.path.dirname(os.path.dirname(os.path.abspath(sys.executable)))
+for f in ("strata/setup.py", "strata/serve/server.py", "strata-engine/strata"):
+    assert os.path.isfile(os.path.join(e, f)), f + " is missing"
+print("strata version", "` + version + `", "engine", "` + rel.engine + `")`},
+			FreeGB: 5,
+		}
+	}
+	return pick("Strata "+version, variant, []candidate{
+		{build("cuda13"), gpuNeeds{minDriverForCUDA13, 7.5}, "Strata " + version + " built for CUDA 13"},
+		// 12.9 rather than 12: its machine code is for 6.0 to 8.9 and its
+		// PTX was compiled by 12.9, so the driver must be 575.
+		{build("cuda12"), gpuNeeds{575, 6.0}, "Strata " + version + " built for CUDA 12.9"},
+	}, h)
+}
+
+// --- FreeToken ---
+
+// FreeTokenRepo is where FreeToken is developed; its builds are on PyPI.
+const FreeTokenRepo = "FlashML-org/FreeToken"
+
+// freetokenMinCompute is the oldest GPU FreeToken runs on. It says RTX 30
+// and newer, and its kernels agree: on a Quadro RTX 6000 (7.5) version 0.2.1
+// loaded gpt-oss-20b and failed its first prefill with no kernel image for
+// the card, and in fp16 with fused experts its first request hit an illegal
+// memory access. Its FP8 kernels need 8.9 and its NVFP4 ones 12.0, which
+// the model's quantisation decides.
+const freetokenMinCompute = 8.0
+
+// freetokenToolkit finds the CUDA toolkit FreeToken compiles its kernels
+// with; a variable for tests.
+var freetokenToolkit = func() string { return discovery.CUDAToolkit(13) }
+
+// freetokenCandidates are the FreeToken builds for a machine: its PyPI
+// wheels with the kernel extras (FlashInfer, SGLang's kernels), CUDA 13,
+// x86-64 only. A node of older cards is not one it is meant for, and is not
+// reported as a fault; one of new cards without the toolkit is.
+func freetokenCandidates(version, variant string, h discovery.Host) ([]Build, error) {
+	if _, err := gpuArch("FreeToken", h); err != nil {
+		return nil, err
+	}
+	if h.Arch != "amd64" {
+		return nil, NotHere{"FreeToken publishes wheels for x86-64 only; this node is " + h.Arch}
+	}
+	if cc := h.MinComputeCapability; cc > 0 && cc < freetokenMinCompute {
+		return nil, NotHere{fmt.Sprintf("FreeToken runs on GPUs of compute capability %.1f and newer "+
+			"(RTX 30, A100 and later); this node has %.1f", freetokenMinCompute, cc)}
+	}
+	// It compiles kernels the first time a model loads, with nvcc, as its
+	// own install guide says; without a toolkit the installation looks
+	// finished and the first replica fails.
+	if freetokenToolkit() == "" {
+		return nil, fmt.Errorf("FreeToken compiles its kernels where it runs and needs the CUDA 13 toolkit's nvcc " +
+			"(in CUDA_HOME, /usr/local/cuda or /usr/local/cuda-13.*); install cuda-toolkit-13 from NVIDIA's repository")
+	}
+	b := Build{
+		Variant:      "cu13",
+		Python:       pythonForEngines,
+		Requirements: []string{"freetoken[accel]==" + version},
+		Binary:       "ft",
+		Packages:     compilerPackages,
+		FreeGB:       20,
+		GPUCheck:     true,
+	}
+	return pick("FreeToken "+version, variant, []candidate{
+		{b, gpuNeeds{minDriverForCUDA13, freetokenMinCompute}, "FreeToken " + version},
+	}, h)
+}

internal/agent/provision/strata_test.go

@@ -0,0 +1,118 @@
+package provision
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/discovery"
+)
+
+// strataRelease is Strata's source archive as GitHub makes one, and its
+// release's engine for CUDA 13, with the recipe pointed at the archive.
+func strataRelease(t *testing.T, rel *release) {
+	t.Helper()
+	archive := tarGz(t, map[string]string{
+		"Strata-0.2.7/setup.py":         "print('setup')\n",
+		"Strata-0.2.7/serve/server.py":  "print('server')\n",
+		"Strata-0.2.7/requirements.txt": "numpy==2.5.3; python_version >= \"3.12\"\njinja2==3.1.6\n",
+	}, nil)
+	rel.assets["v0.2.7.tar.gz"] = archive
+	sum := sha256.Sum256(archive)
+	engine := tarGz(t, map[string]string{"strata": "\x7fELF", "libcudart.so.13": "lib"}, nil)
+	name := "strata-engine-0.1.52-linux-x86_64-cuda-cuda13.tar.gz"
+	rel.assets[name] = engine
+	esum := sha256.Sum256(engine)
+	rel.digests[name] = "sha256:" + hex.EncodeToString(esum[:])
+
+	old := strataReleases
+	strataReleases = map[string]struct {
+		sha256    string
+		published string
+		engine    string
+	}{"0.2.7": {hex.EncodeToString(sum[:]), "2026-10-27T14:54:42Z", "0.1.52"}}
+	t.Cleanup(func() { strataReleases = old })
+}
+
+func TestStrataIsInstalledWithItsEngineBesideItsCode(t *testing.T) {
+	sys := &fakePython{fakeSystem: fakeSystem{installed: map[string]bool{}}}
+	in, rel := pythonInstaller(t, turingNode, sys)
+	strataRelease(t, rel)
+
+	got := in.Converge(t.Context(), []Want{{Engine: "strata", Version: "0.2.7"}})
+	if got[0].State != Ready {
+		t.Fatalf("not ready: %+v", got[0])
+	}
+	dir := filepath.Join(in.Root, "strata", "0.2.7-cuda13")
+	if got[0].Binary != filepath.Join(dir, "bin", "python") {
+		t.Errorf("binary %s, want the environment's interpreter", got[0].Binary)
+	}
+	for _, f := range []string{"strata/setup.py", "strata/serve/server.py", "strata-engine/strata", "strata-engine/libcudart.so.13"} {
+		if _, err := os.Stat(filepath.Join(dir, f)); err != nil {
+			t.Errorf("%s is not in the environment: %v", f, err)
+		}
+	}
+	all := strings.Join(sys.commands, "\n")
+	for _, want := range []string{"--exclude-newer 2026-10-28T14:54:42Z", "jinja2==3.1.6"} {
+		if !strings.Contains(all, want) {
+			t.Errorf("the installation never used %q:\n%s", want, all)
+		}
+	}
+}
+
+func TestStrataRefusesAnEngineWhoseDigestDoesNotMatch(t *testing.T) {
+	sys := &fakePython{fakeSystem: fakeSystem{installed: map[string]bool{}}}
+	in, rel := pythonInstaller(t, turingNode, sys)
+	strataRelease(t, rel)
+	rel.digests["strata-engine-0.1.52-linux-x86_64-cuda-cuda13.tar.gz"] = "sha256:" + strings.Repeat("0", 64)
+
+	got := in.Converge(t.Context(), []Want{{Engine: "strata", Version: "0.2.7", Variant: "cuda13"}})
+	if got[0].State != Failed || !strings.Contains(got[0].Detail, "does not match the checksum") {
+		t.Fatalf("an engine that does not match its digest was not refused: %+v", got[0])
+	}
+}
+
+func TestStrataPicksTheEngineForTheCardsAndDriver(t *testing.T) {
+	for _, tc := range []struct {
+		name string
+		host discovery.Host
+		want []string
+		err  string
+	}{
+		{"turing on 580", turingNode, []string{"cuda13", "cuda12"}, ""},
+		{"volta", discovery.Host{Arch: "amd64", NVIDIADriver: "580.82.07", MinComputeCapability: 7.0}, []string{"cuda12"}, ""},
+		{"driver 575", discovery.Host{Arch: "amd64", NVIDIADriver: "575.51.03", MinComputeCapability: 8.6}, []string{"cuda12"}, ""},
+		{"driver 570", discovery.Host{Arch: "amd64", NVIDIADriver: "570.86.10", MinComputeCapability: 8.6}, nil, "575"},
+		{"no gpu", discovery.Host{Arch: "amd64"}, nil, "no"},
+		{"arm", discovery.Host{Arch: "arm64", NVIDIADriver: "580.82.07", MinComputeCapability: 9.0}, nil, "x86-64"},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			bs, err := strataCandidates("0.2.7", "", tc.host)
+			if tc.err != "" {
+				if err == nil || !strings.Contains(err.Error(), tc.err) {
+					t.Fatalf("got %v, %v; want an error saying %q", bs, err, tc.err)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatal(err)
+			}
+			var got []string
+			for _, b := range bs {
+				got = append(got, b.Variant)
+			}
+			if strings.Join(got, ",") != strings.Join(tc.want, ",") {
+				t.Errorf("builds %v, want %v", got, tc.want)
+			}
+		})
+	}
+}
+
+func TestStrataIsNotInAnOfflineBundle(t *testing.T) {
+	if _, err := Builds("strata", "0.2.7", "amd64"); err == nil {
+		t.Fatal("an offline bundle would carry Strata, which needs PyPI")
+	}
+}

internal/agent/settings.go

@@ -0,0 +1,109 @@
+package agent
+
+import (
+	"context"
+	"crypto/sha256"
+	"os"
+	"path/filepath"
+	"time"
+
+	llmgwv1 "gitea.mynodes.xyz/ilker/llmgw/api/gen/llmgw/v1"
+)
+
+// The cluster's settings for this agent (BACKLOG E3): what agent.yaml can
+// say but the agent's id, control planes, address and certificate, kept by
+// the control planes and sent over the stream. Kept here beside agent.yaml,
+// which then holds only what the agent needs to start; at start the two are
+// read together, these over the file.
+
+// SettingsFile keeps the settings the control plane last sent.
+var SettingsFile = "/var/lib/llmgw-agent/agent-settings.yaml"
+
+// ReadSettings is what SettingsFile holds, nil when nothing.
+func ReadSettings() []byte {
+	b, err := os.ReadFile(SettingsFile)
+	if err != nil {
+		return nil
+	}
+	return b
+}
+
+// SetSettingsCheck says how to tell whether settings would load beside this
+// agent's own file; settings that would not are not taken.
+func (a *Agent) SetSettingsCheck(f func([]byte) error) { a.settingsCheck = f }
+
+// applySettings takes the settings the control plane sent: nothing when they
+// are what this agent runs on, and otherwise kept and started again on.
+func (a *Agent) applySettings(m *llmgwv1.AgentSettings) {
+	next := m.GetYaml()
+	if next == a.cfg.Settings {
+		return
+	}
+	if a.settingsCheck != nil {
+		if err := a.settingsCheck([]byte(next)); err != nil {
+			// Kept running on what it has: settings that will not load
+			// here must not take the node out of service.
+			a.log.Error("the cluster's settings for this agent do not load beside its agent.yaml; running on what it has",
+				"err", err)
+			return
+		}
+	}
+	if err := os.MkdirAll(filepath.Dir(SettingsFile), 0o755); err != nil {
+		a.log.Error("could not keep the cluster's settings for this agent", "err", err)
+		return
+	}
+	tmp := SettingsFile + ".new"
+	if err := os.WriteFile(tmp, []byte(next), 0o600); err != nil {
+		a.log.Error("could not keep the cluster's settings for this agent", "err", err)
+		return
+	}
+	if err := os.Rename(tmp, SettingsFile); err != nil {
+		a.log.Error("could not keep the cluster's settings for this agent", "err", err)
+		return
+	}
+	a.log.Info("the cluster's settings for this agent changed; starting again on them, the engines keep serving")
+	time.Sleep(300 * time.Millisecond)
+	exitForUpgrade()
+}
+
+// WatchFile starts the agent again when its configuration file changes and
+// still loads: agent.yaml is read only at start, and a change to it that
+// waited for somebody to restart the agent was a change nobody saw take
+// effect.
+func (a *Agent) WatchFile(ctx context.Context, path string, loads func() error) {
+	if path == "" {
+		return
+	}
+	sum := func() [32]byte {
+		b, _ := os.ReadFile(path) //nolint:gosec // the agent's own configuration
+		return sha256.Sum256(b)
+	}
+	was := sum()
+	t := time.NewTicker(watchEvery)
+	defer t.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-t.C:
+		}
+		now := sum()
+		if now == was {
+			continue
+		}
+		if err := loads(); err != nil {
+			// Once per change: an editor saving halfway must not stop the
+			// agent, and a broken file is reported, not restarted into.
+			a.log.Error("agent.yaml changed and does not load; running on what it had", "path", path, "err", err)
+			was = now
+			continue
+		}
+		a.log.Info("agent.yaml changed; starting again on it, the engines keep serving", "path", path)
+		time.Sleep(300 * time.Millisecond)
+		exitForUpgrade()
+		return
+	}
+}
+
+// watchEvery is how often agent.yaml is looked at.
+var watchEvery = 10 * time.Second

internal/agent/settings_test.go

@@ -0,0 +1,88 @@
+package agent
+
+import (
+	"errors"
+	"os"
+	"path/filepath"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	llmgwv1 "gitea.mynodes.xyz/ilker/llmgw/api/gen/llmgw/v1"
+)
+
+func settingsAgent(t *testing.T, running string) (*Agent, chan struct{}) {
+	t.Helper()
+	oldFile, oldExit := SettingsFile, exitForUpgrade
+	SettingsFile = filepath.Join(t.TempDir(), "agent-settings.yaml")
+	exited := make(chan struct{}, 1)
+	exitForUpgrade = func() { exited <- struct{}{} }
+	t.Cleanup(func() { SettingsFile, exitForUpgrade = oldFile, oldExit })
+	cfg := Config{Server: "127.0.0.1:7443", NodeID: "n1", WeightsCache: t.TempDir(), EnginesDir: t.TempDir(), Settings: running}
+	cfg.Superv.WorkDir = t.TempDir()
+	a, err := New(cfg, nil, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return a, exited
+}
+
+// exitedSoon is whether the agent exited within a second: it waits a
+// moment for its log line to leave first.
+func exitedSoon(ch chan struct{}) bool {
+	select {
+	case <-ch:
+		return true
+	case <-time.After(time.Second):
+		return false
+	}
+}
+
+func TestAnAgentOnItsSettingsStaysAndOneOffThemStartsAgain(t *testing.T) {
+	a, exited := settingsAgent(t, "heartbeat_interval: 5s\n")
+	a.applySettings(&llmgwv1.AgentSettings{Yaml: "heartbeat_interval: 5s\n"})
+	if exitedSoon(exited) || ReadSettings() != nil {
+		t.Error("settings it already runs on restarted it")
+	}
+	a.applySettings(&llmgwv1.AgentSettings{Yaml: "heartbeat_interval: 2s\n"})
+	if !exitedSoon(exited) || string(ReadSettings()) != "heartbeat_interval: 2s\n" {
+		t.Errorf("new settings: kept %q", ReadSettings())
+	}
+}
+
+func TestSettingsThatWouldNotLoadAreNotTaken(t *testing.T) {
+	a, exited := settingsAgent(t, "")
+	a.SetSettingsCheck(func([]byte) error { return errors.New("fabric ib-a lists no HCAs") })
+	a.applySettings(&llmgwv1.AgentSettings{Yaml: "fabrics: [{name: ib-a}]\n"})
+	if exitedSoon(exited) || ReadSettings() != nil {
+		t.Error("settings that do not load were taken")
+	}
+}
+
+func TestAChangedAgentYamlThatLoadsStartsTheAgentAgain(t *testing.T) {
+	a, exited := settingsAgent(t, "")
+	old := watchEvery
+	watchEvery = 10 * time.Millisecond
+	t.Cleanup(func() { watchEvery = old })
+	p := filepath.Join(t.TempDir(), "agent.yaml")
+	_ = os.WriteFile(p, []byte("node_id: n1\n"), 0o600)
+	var loads atomic.Bool
+	check := func() error {
+		if loads.Load() {
+			return nil
+		}
+		return errors.New("broken")
+	}
+	go a.WatchFile(t.Context(), p, check)
+	time.Sleep(30 * time.Millisecond)
+	// Saved broken: reported, not restarted into.
+	_ = os.WriteFile(p, []byte("node_id: [\n"), 0o600)
+	if exitedSoon(exited) {
+		t.Fatal("a broken agent.yaml restarted the agent")
+	}
+	loads.Store(true)
+	_ = os.WriteFile(p, []byte("node_id: n1\nlabels: {a: b}\n"), 0o600)
+	if !exitedSoon(exited) {
+		t.Error("a changed agent.yaml did not restart the agent")
+	}
+}

internal/agent/supervisor/health.go

@@ -54,9 +54,13 @@ func NewHTTPHealth(timeout time.Duration, opts ...HTTPHealthOption) *HTTPHealth
 
 // Ready reports whether the engine at baseURL is serving model.
 func (h *HTTPHealth) Ready(ctx context.Context, baseURL, model, apiKey string) error {
-	if err := h.get(ctx, baseURL+h.healthPath, apiKey, nil); err != nil {
+	var body healthBody
+	if err := h.get(ctx, baseURL+h.healthPath, apiKey, &body); err != nil {
 		return fmt.Errorf("health: %w", err)
 	}
+	if body.notReady() {
+		return fmt.Errorf("health: the engine says %q", body.Status)
+	}
 	if !h.requireName || model == "" {
 		return nil
 	}
@@ -98,5 +102,29 @@ func (h *HTTPHealth) get(ctx context.Context, url, apiKey string, out any) error
 	if out == nil {
 		return nil
 	}
+	if hb, ok := out.(*healthBody); ok {
+		// Most engines answer /health with nothing, or with text: only a
+		// JSON status says anything, and a body that is not one is not a
+		// failure.
+		_ = json.NewDecoder(io.LimitReader(resp.Body, 64<<10)).Decode(hb)
+		return nil
+	}
 	return json.NewDecoder(resp.Body).Decode(out)
 }
+
+// healthBody is what an engine's /health may say in its body. Some answer
+// 200 from the moment their HTTP server is up and say in the body that the
+// model is still loading — FreeToken's {"status":"loading","phase":...} —
+// and a model that lists itself in /v1/models meanwhile would be routed to.
+type healthBody struct {
+	Status string `json:"status"`
+}
+
+// notReady is a body that says the engine cannot serve yet, or has failed.
+func (b healthBody) notReady() bool {
+	switch b.Status {
+	case "loading", "starting", "initializing", "warming_up", "error", "unhealthy":
+		return true
+	}
+	return false
+}

internal/agent/supervisor/health_test.go

@@ -103,3 +103,29 @@ func TestTheHealthCheckSendsTheEnginesKey(t *testing.T) {
 		t.Error("without the key the engine refused, and that read as ready")
 	}
 }
+
+func TestNotReadyWhileTheHealthBodySaysLoading(t *testing.T) {
+	// FreeToken answers 200 from the moment its HTTP server is up and lists
+	// its model meanwhile; only the body says the weights are still loading.
+	body := `{"status":"loading","phase":"weights","progress":{"done_bytes":1,"total_bytes":9},"model":"glm"}`
+	mux := http.NewServeMux()
+	mux.HandleFunc("GET /health", func(w http.ResponseWriter, _ *http.Request) { io.WriteString(w, body) })
+	mux.HandleFunc("GET /v1/models", func(w http.ResponseWriter, _ *http.Request) {
+		io.WriteString(w, `{"data":[{"id":"glm"}]}`)
+	})
+	srv := httptest.NewServer(mux)
+	t.Cleanup(srv.Close)
+	h := NewHTTPHealth(time.Second)
+	if err := h.Ready(t.Context(), srv.URL, "glm", ""); err == nil || !strings.Contains(err.Error(), "loading") {
+		t.Errorf("a loading engine was reported ready, or not said to be loading: %v", err)
+	}
+	body = `{"status":"ok","model":"glm","uptime_s":3}`
+	if err := h.Ready(t.Context(), srv.URL, "glm", ""); err != nil {
+		t.Errorf("a serving engine was reported unready: %v", err)
+	}
+	// Text, as some engines answer: not a status, and not a failure.
+	body = "OK"
+	if err := h.Ready(t.Context(), srv.URL, "glm", ""); err != nil {
+		t.Errorf("a plain-text health answer was taken for a failure: %v", err)
+	}
+}

internal/agent/supervisor/outputfile.go

@@ -0,0 +1,139 @@
+package supervisor
+
+import (
+	"bytes"
+	"os"
+	"path/filepath"
+	"time"
+)
+
+// An engine's output goes to a file of its own, not to a pipe of the
+// agent's. A pipe has one reader, and that reader is the agent that
+// started the engine: when that agent restarts, the engine it left running
+// — and adopts a moment later — writes to a pipe nobody reads, and its next
+// write past the pipe's buffer fails with EPIPE. A Python engine's progress
+// bar then raises, and takes the request it was reporting on with it. A
+// file has no reader to lose: the engine appends to it, every agent reads
+// it, and the output from before a restart is there for the one after.
+
+// maxOutputBytes bounds one process's file: past it the file is emptied and
+// the process goes on appending to it. A progress bar rewriting itself for
+// an hour would otherwise fill the disk.
+const maxOutputBytes = 64 << 20
+
+// outputPath is where a shard's process writes.
+func (s *Supervisor) outputPath(shardID, name string) string {
+	if name != "" {
+		return filepath.Join(s.cfg.WorkDir, "shard-"+shardID+"."+name+".log")
+	}
+	return filepath.Join(s.cfg.WorkDir, "shard-"+shardID+".log")
+}
+
+// openOutput is the file a process is started with: new, for a process
+// that is new.
+func (s *Supervisor) openOutput(shardID, name string) (*os.File, error) {
+	return os.OpenFile(s.outputPath(shardID, name), os.O_CREATE|os.O_WRONLY|os.O_APPEND|os.O_TRUNC, 0o600) //nolint:gosec // the agent's own directory
+}
+
+// follow reads a process's file into the shard's log as it is written,
+// until the shard's output is stopped and the file has been read to its
+// end. A file an earlier agent began is read from the start, so the output
+// from before a restart is in the log too.
+func (s *Supervisor) follow(sh *Shard, path, name string) {
+	defer sh.outDone.Done()
+	f, err := os.Open(path) //nolint:gosec // the agent's own directory
+	if err != nil {
+		sh.logs.add("[" + name + "] output unavailable: " + err.Error())
+		return
+	}
+	defer f.Close()
+	var offset int64
+	var rest []byte
+	buf := make([]byte, 64<<10)
+	read := func() {
+		for {
+			st, err := f.Stat()
+			if err != nil {
+				return
+			}
+			if st.Size() < offset {
+				// Emptied, by this agent or another: begin again.
+				offset, rest = 0, nil
+			}
+			if st.Size() > maxOutputBytes {
+				_ = os.Truncate(path, 0)
+				offset, rest = 0, nil
+				continue
+			}
+			n, err := f.ReadAt(buf, offset)
+			if n > 0 {
+				offset += int64(n)
+				rest = append(rest, buf[:n]...)
+				for {
+					i := bytes.IndexByte(rest, '\n')
+					if i < 0 {
+						break
+					}
+					s.addLine(sh, name, rest[:i])
+					rest = rest[i+1:]
+				}
+				// A progress bar is a line rewritten with \r and never ended:
+				// its latest state is what is wanted, not a line a megabyte
+				// long. Kept, as llama.cpp and tqdm write them.
+				if len(rest) > 1<<20 {
+					s.addLine(sh, name, rest)
+					rest = nil
+				}
+			}
+			if err != nil || n == 0 {
+				return
+			}
+		}
+	}
+	t := time.NewTicker(200 * time.Millisecond)
+	defer t.Stop()
+	for {
+		select {
+		case <-sh.outStop:
+			read()
+			if len(rest) > 0 {
+				s.addLine(sh, name, rest)
+			}
+			return
+		case <-t.C:
+			read()
+		}
+	}
+}
+
+func (s *Supervisor) addLine(sh *Shard, name string, b []byte) {
+	// The last state of a line rewritten with \r, as a terminal would show.
+	if i := bytes.LastIndexByte(b, '\r'); i >= 0 {
+		b = b[i+1:]
+	}
+	line := string(b)
+	if name != "" {
+		// Prefixed so the two processes of a Ray shard stay distinguishable
+		// in a single log tail.
+		line = "[" + name + "] " + line
+	}
+	sh.logs.add(line)
+	s.log.Debug("engine", "shard", sh.ID, "line", line)
+}
+
+// removeOutput drops a shard's files once nothing writes to them.
+func (s *Supervisor) removeOutput(shardID string) {
+	matches, _ := filepath.Glob(filepath.Join(s.cfg.WorkDir, "shard-"+shardID+".*log"))
+	for _, m := range matches {
+		_ = os.Remove(m)
+	}
+	_ = os.Remove(s.outputPath(shardID, ""))
+}
+
+// stopOutput has every follower read what is left and stop. Called once
+// the process has ended, before the shard is done, so what it wrote last
+// is in the log by the time anyone reads the log for why it ended.
+func (s *Supervisor) stopOutput(sh *Shard) {
+	sh.outOnce.Do(func() { close(sh.outStop) })
+	sh.outDone.Wait()
+}

internal/agent/supervisor/process_unix.go

@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"os"
 	"strconv"
+	"strings"
 	"syscall"
 )
 
@@ -63,3 +64,71 @@ func isZombie(pid int) bool {
 	}
 	return b[i+2] == 'Z'
 }
+
+// strays are a process's descendants that left its group, as an engine that
+// starts its worker in a session of its own does (Strata): signalling the
+// group misses them, and one left behind holds the card and the RAM the next
+// replica needs. Found before the group is signalled, while they are still
+// its descendants: after, they belong to init. Linux only; elsewhere none.
+func strays(pid int) []int {
+	entries, err := os.ReadDir("/proc")
+	if err != nil {
+		return nil
+	}
+	parent := map[int]int{}
+	group := map[int]int{}
+	for _, e := range entries {
+		p, err := strconv.Atoi(e.Name())
+		if err != nil {
+			continue
+		}
+		if pp, pg, ok := parentAndGroup(p); ok {
+			parent[p], group[p] = pp, pg
+		}
+	}
+	var out []int
+	for p := range parent {
+		if group[p] == pid || p == pid {
+			continue
+		}
+		// Bounded: no engine's tree is this deep, and a pid reused while
+		// /proc was read must not loop.
+		for q, n := parent[p], 0; q > 1 && n < 32; q, n = parent[q], n+1 {
+			if q == pid {
+				out = append(out, p)
+				break
+			}
+		}
+	}
+	return out
+}
+
+// parentAndGroup reads a process's parent and group from /proc.
+func parentAndGroup(pid int) (int, int, bool) {
+	b, err := os.ReadFile("/proc/" + strconv.Itoa(pid) + "/stat")
+	if err != nil {
+		return 0, 0, false
+	}
+	i := bytes.LastIndexByte(b, ')')
+	if i < 0 {
+		return 0, 0, false
+	}
+	f := strings.Fields(string(b[i+1:]))
+	if len(f) < 3 {
+		return 0, 0, false
+	}
+	pp, err1 := strconv.Atoi(f[1])
+	pg, err2 := strconv.Atoi(f[2])
+	return pp, pg, err1 == nil && err2 == nil
+}
+
+// signalStrays signals each stray, and its own group when it leads one.
+func signalStrays(pids []int, sig syscall.Signal) {
+	for _, p := range pids {
+		if _, pg, ok := parentAndGroup(p); ok && pg == p {
+			_ = syscall.Kill(-p, sig)
+			continue
+		}
+		_ = syscall.Kill(p, sig)
+	}
+}

internal/agent/supervisor/runas.go

@@ -3,6 +3,7 @@ package supervisor
 import (
 	"fmt"
 	"os"
+	"os/exec"
 	"os/user"
 	"path/filepath"
 	"strconv"
@@ -90,3 +91,25 @@ func credentialFor(u *user.User, workDir string) (*runAs, error) {
 func (r *runAs) env() []string {
 	return []string{"HOME=" + r.home, "XDG_CACHE_HOME=" + filepath.Join(r.home, ".cache"), "USER=" + r.name, "LOGNAME=" + r.name}
 }
+
+// AsEngineUser has cmd run as the engine user, as an engine would, with its
+// home under workDir: for work the agent does that is not an engine but
+// reads what an engine reads — training an adapter. It answers that home
+// and a way to hand a path to the user, or why it runs as this process.
+func AsEngineUser(cmd *exec.Cmd, name, workDir string) (home string, give func(path string) error, why string) {
+	r, why := resolveRunAs(name, workDir)
+	if r == nil {
+		return "", func(string) error { return nil }, why
+	}
+	cmd.SysProcAttr = &syscall.SysProcAttr{Credential: r.cred, Setpgid: true}
+	cmd.Env = append(cmd.Env, r.env()...)
+	uid, gid := int(r.cred.Uid), int(r.cred.Gid)
+	return r.home, func(path string) error {
+		return filepath.WalkDir(path, func(p string, _ os.DirEntry, err error) error {
+			if err != nil {
+				return err
+			}
+			return os.Lchown(p, uid, gid)
+		})
+	}, ""
+}

internal/agent/supervisor/supervisor.go

@@ -1,12 +1,10 @@
 package supervisor
 
 import (
-	"bufio"
 	"context"
 	"encoding/json"
 	"errors"
 	"fmt"
-	"io"
 	"log/slog"
 	"net"
 	"os"
@@ -104,6 +102,12 @@ type Shard struct {
 	ID   string
 	Spec engines.ShardSpec
 
+	// outStop ends the followers of the processes' output files, and
+	// outDone is their having read to the end.
+	outStop chan struct{}
+	outOnce sync.Once
+	outDone sync.WaitGroup
+
 	mu       sync.Mutex
 	state    State
 	pid      int
@@ -248,6 +252,7 @@ func (s *Supervisor) Start(ctx context.Context, id string, spec engines.ShardSpe
 		logs:    newRingBuffer(s.cfg.LogLines),
 		started: time.Now(),
 		done:    make(chan struct{}),
+		outStop: make(chan struct{}),
 	}
 	s.shards[id] = sh
 	s.mu.Unlock()
@@ -315,14 +320,15 @@ func (s *Supervisor) launch(ctx context.Context, sh *Shard, cmds []engines.Comma
 	// CommandContext kills only the leader, so replace that with a group kill.
 	c.Cancel = func() error { return killGroup(c.Process.Pid, syscall.SIGKILL) }
 
-	stdout, err := c.StdoutPipe()
+	out, err := s.openOutput(sh.ID, "")
 	if err != nil {
 		cancel()
 		return err
 	}
-	c.Stderr = c.Stdout.(io.Writer) // interleave, as the engine writes them
-
-	if err := c.Start(); err != nil {
+	c.Stdout, c.Stderr = out, out // interleaved, as the engine writes them
+	err = c.Start()
+	out.Close() // the process holds its own
+	if err != nil {
 		cancel()
 		return fmt.Errorf("starting %s: %w", cmd.Path, err)
 	}
@@ -339,7 +345,8 @@ func (s *Supervisor) launch(ctx context.Context, sh *Shard, cmds []engines.Comma
 			"shard", sh.ID, "err", err)
 	}
 
-	go s.pump(sh, stdout, cmd.Name)
+	sh.outDone.Add(1)
+	go s.follow(sh, s.outputPath(sh.ID, ""), cmd.Name)
 	go s.wait(sh)
 	return nil
 }
@@ -353,16 +360,18 @@ func (s *Supervisor) spawn(ctx context.Context, sh *Shard, c engines.Command) (*
 	s.dropTo(proc)
 	proc.Cancel = func() error { return killGroup(proc.Process.Pid, syscall.SIGKILL) }
 
-	stdout, err := proc.StdoutPipe()
+	out, err := s.openOutput(sh.ID, c.Name)
 	if err != nil {
 		return nil, err
 	}
-	proc.Stderr = proc.Stdout.(io.Writer)
-
-	if err := proc.Start(); err != nil {
+	proc.Stdout, proc.Stderr = out, out
+	err = proc.Start()
+	out.Close()
+	if err != nil {
 		return nil, fmt.Errorf("starting %s: %w", c.Path, err)
 	}
-	go s.pump(sh, stdout, c.Name)
+	sh.outDone.Add(1)
+	go s.follow(sh, s.outputPath(sh.ID, c.Name), c.Name)
 	// Reaped so it does not become a zombie; its exit does not end the shard.
 	go func() { _ = proc.Wait() }()
 	return proc, nil
@@ -395,24 +404,6 @@ func waitForPort(ctx context.Context, addr string, timeout time.Duration) (err e
 }
 
 // pump copies a process's output into the ring buffer and the log.
-func (s *Supervisor) pump(sh *Shard, r io.Reader, name string) {
-	sc := bufio.NewScanner(r)
-	// Engine tracebacks and progress bars produce long lines; the default
-	// 64 KiB limit truncates them exactly when they matter.
-	sc.Buffer(make([]byte, 0, 64<<10), 1<<20)
-
-	for sc.Scan() {
-		line := sc.Text()
-		if name != "" {
-			// Prefixed so the two processes of a Ray shard stay distinguishable
-			// in a single log tail.
-			line = "[" + name + "] " + line
-		}
-		sh.logs.add(line)
-		s.log.Debug("engine", "shard", sh.ID, "line", line)
-	}
-}
-
 // wait reaps the process and records why it ended.
 func (s *Supervisor) wait(sh *Shard) {
 	defer close(sh.done)
@@ -440,6 +431,7 @@ func (s *Supervisor) wait(sh *Shard) {
 	final, code := sh.state, sh.exitCode
 	sh.mu.Unlock()
 
+	s.stopOutput(sh)
 	s.removeState(sh)
 	if final == Failed {
 		s.log.Error("shard failed", "shard", sh.ID, "exit_code", code, "err", err,
@@ -590,10 +582,24 @@ func (s *Supervisor) Stop(ctx context.Context, id string) (err error) {
 	}
 
 	// Signal the group, not the process: engines fork workers that hold GPU
-	// memory, and those outlive a signal sent only to the leader.
+	// memory, and those outlive a signal sent only to the leader. And what
+	// left the group, found while it is still the leader's.
+	left := strays(pid)
 	if err := killGroup(pid, syscall.SIGTERM); err != nil && !errors.Is(err, os.ErrProcessDone) {
 		s.log.Warn("SIGTERM failed", "shard", id, "err", err)
 	}
+	signalStrays(left, syscall.SIGTERM)
+	defer func() {
+		// Given as long as the group to go; then killed.
+		deadline := time.Now().Add(5 * time.Second)
+		for time.Now().Before(deadline) && slices.ContainsFunc(left, processAlive) {
+			time.Sleep(100 * time.Millisecond)
+		}
+		if slices.ContainsFunc(left, processAlive) {
+			s.log.Warn("killing an engine's processes that left its group", "shard", id, "pids", left)
+			signalStrays(left, syscall.SIGKILL)
+		}
+	}()
 
 	// force kills the group outright. exec.Cmd's cancel does this for shards we
 	// started; an adopted shard has none, so signal its group directly.
@@ -606,6 +612,7 @@ func (s *Supervisor) Stop(ctx context.Context, id string) (err error) {
 		if err := killGroup(pid, syscall.SIGKILL); err != nil && !errors.Is(err, os.ErrProcessDone) {
 			s.log.Warn("SIGKILL failed", "shard", id, "err", err)
 		}
+		signalStrays(left, syscall.SIGKILL)
 	}
 
 	select {
@@ -760,6 +767,7 @@ func (s *Supervisor) writeState(sh *Shard) error {
 
 func (s *Supervisor) removeState(sh *Shard) {
 	_ = os.Remove(s.statePath(sh.ID))
+	s.removeOutput(sh.ID)
 }
 
 // Orphan is an engine process left behind by a previous agent.
@@ -838,15 +846,18 @@ func (s *Supervisor) Adopt(o Orphan) (*Shard, error) {
 		logs:    newRingBuffer(s.cfg.LogLines),
 		started: o.StartedAt,
 		done:    make(chan struct{}),
+		outStop: make(chan struct{}),
 		adopted: true,
 		cmdline: o.Cmdline,
 	}
 	s.shards[o.ShardID] = sh
 	s.mu.Unlock()
 
-	// Output is gone with the previous agent; say so rather than presenting an
-	// empty log as if the engine had been silent.
-	sh.logs.add("[adopted from a previous agent; earlier output is not available]")
+	// Its output file is read from the start: what it wrote under the
+	// previous agent is there, and what it writes now goes on arriving.
+	sh.logs.add("[adopted from a previous agent]")
+	sh.outDone.Add(1)
+	go s.follow(sh, s.outputPath(sh.ID, ""), "")
 
 	go s.poll(sh)
 	return sh, nil
@@ -874,6 +885,7 @@ func (s *Supervisor) poll(sh *Shard) {
 		}
 		sh.mu.Unlock()
 
+		s.stopOutput(sh)
 		s.removeState(sh)
 		s.log.Info("adopted shard exited", "shard", sh.ID, "expected", expected)
 		return
@@ -900,6 +912,7 @@ func (s *Supervisor) KillOrphan(o Orphan) error {
 	}
 	// An already-missing state file is success: the previous agent may have
 	// reaped the process itself between the scan and this call.
+	s.removeOutput(o.ShardID)
 	if err := os.Remove(s.statePath(o.ShardID)); err != nil && !errors.Is(err, os.ErrNotExist) {
 		return err
 	}

internal/agent/supervisor/supervisor_test.go

@@ -1,12 +1,14 @@
 package supervisor
 
 import (
+	"bytes"
 	"context"
 	"errors"
 	"fmt"
 	"log/slog"
 	"net"
 	"os"
+	"os/exec"
 	"path/filepath"
 	"slices"
 	"strings"
@@ -905,3 +907,113 @@ func TestAnAdoptedShardsComputeIsKeptWithoutCrashing(t *testing.T) {
 		t.Fatalf("lost: %+v", again)
 	}
 }
+
+func TestAnAdoptedEngineGoesOnWritingAndItsEarlierOutputIsKept(t *testing.T) {
+	// An engine's output went to a pipe of the agent's. The agent that
+	// restarted and adopted the engine left that pipe with no reader, and
+	// the engine's next write past the pipe's buffer failed with EPIPE —
+	// which, in a Python engine, raised inside whatever was being reported
+	// on and failed the request. Output goes to a file now: a reader is not
+	// something a file can lose, and the lines from before the restart are
+	// there for the agent after it.
+	s, _ := newSupervisor(t)
+	// 256 KiB at once outgrows a pipe's buffer; a reader gone means EPIPE.
+	sh, err := s.Start(t.Context(), "a", engines.ShardSpec{ReplicaID: "r"},
+		shellCommand(`echo before-restart; sleep 1; head -c 262144 /dev/zero | tr '\0' x; echo; echo after-restart-done; sleep 30`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	eventually(t, 3*time.Second, "the first line was never captured", func() bool {
+		return strings.Contains(strings.Join(sh.Logs(10), "\n"), "before-restart")
+	})
+	// The agent goes away without stopping the engine: nothing reads the
+	// output between here and the adoption.
+	pid := sh.Status().PID
+	sh.outOnce.Do(func() { close(sh.outStop) })
+	sh.outDone.Wait()
+
+	s2, err := New(Config{WorkDir: s.cfg.WorkDir, StopTimeout: 3 * time.Second, LogLines: 64},
+		&scriptedHealth{}, slog.New(slog.DiscardHandler))
+	if err != nil {
+		t.Fatal(err)
+	}
+	orphans, _ := s2.FindOrphans()
+	adopted, err := s2.Adopt(orphans[0])
+	if err != nil {
+		t.Fatal(err)
+	}
+	eventually(t, 5*time.Second, "the engine's output after the restart never arrived: "+strings.Join(adopted.Logs(5), " | "), func() bool {
+		return strings.Contains(strings.Join(adopted.Logs(64), "\n"), "after-restart-done")
+	})
+	if !strings.Contains(strings.Join(adopted.Logs(64), "\n"), "before-restart") {
+		t.Error("the output from before the restart is not in the adopted log")
+	}
+	if !processAlive(pid) {
+		t.Fatal("the engine died writing output nobody was reading")
+	}
+	if err := s2.Stop(t.Context(), "a"); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := os.Stat(s2.outputPath("a", "")); !errors.Is(err, os.ErrNotExist) {
+		t.Errorf("the output file was not removed with the shard: %v", err)
+	}
+}
+
+func TestAnOutputFileIsEmptiedPastItsBound(t *testing.T) {
+	s, _ := newSupervisor(t)
+	path := s.outputPath("a", "")
+	// A file already past the bound, as a long-running engine leaves one.
+	big, _ := os.OpenFile(path, os.O_CREATE|os.O_WRONLY, 0o644)
+	_, _ = big.Write(bytes.Repeat([]byte("x"), maxOutputBytes+1))
+	big.Close()
+	sh := &Shard{ID: "a", logs: newRingBuffer(8), done: make(chan struct{}), outStop: make(chan struct{})}
+	sh.outDone.Add(1)
+	go s.follow(sh, path, "")
+	time.Sleep(300 * time.Millisecond)
+	f, _ := os.OpenFile(path, os.O_WRONLY|os.O_APPEND, 0o644)
+	_, _ = f.WriteString("still here\n")
+	f.Close()
+	s.stopOutput(sh)
+	st, _ := os.Stat(path)
+	if st.Size() > 1024 {
+		t.Errorf("the file was not emptied: %d bytes", st.Size())
+	}
+	if !strings.Contains(strings.Join(sh.Logs(8), "\n"), "still here") {
+		t.Errorf("a line written after the file was emptied was lost: %v", sh.Logs(8))
+	}
+}
+
+// An engine's worker in a session of its own, as Strata's is, is not in the
+// group a stop signals: it is found and stopped as well.
+func TestStopKillsAWorkerThatLeftTheGroup(t *testing.T) {
+	if _, err := os.Stat("/proc/self/stat"); err != nil {
+		t.Skip("needs /proc")
+	}
+	if _, err := exec.LookPath("setsid"); err != nil {
+		t.Skip("needs setsid")
+	}
+	s, _ := newSupervisor(t)
+	marker := filepath.Join(t.TempDir(), "worker.pid")
+	script := fmt.Sprintf(`setsid sh -c 'echo $$ > %s; sleep 30' & sleep 30`, marker)
+	if _, err := s.Start(t.Context(), "a", engines.ShardSpec{}, shellCommand(script)); err != nil {
+		t.Fatal(err)
+	}
+	var worker int
+	for deadline := time.Now().Add(3 * time.Second); time.Now().Before(deadline) && worker == 0; time.Sleep(10 * time.Millisecond) {
+		if b, err := os.ReadFile(marker); err == nil {
+			_, _ = fmt.Sscanf(strings.TrimSpace(string(b)), "%d", &worker)
+		}
+	}
+	if worker == 0 {
+		t.Fatal("the worker never started")
+	}
+	if err := s.Stop(t.Context(), "a"); err != nil {
+		t.Fatal(err)
+	}
+	for deadline := time.Now().Add(8 * time.Second); time.Now().Before(deadline); time.Sleep(50 * time.Millisecond) {
+		if !processAlive(worker) {
+			return
+		}
+	}
+	t.Errorf("the worker %d that left the group survived the stop", worker)
+}

internal/agent/train/runner.go

@@ -0,0 +1,574 @@
+// Package train runs LoRA training jobs on a node (BACKLOG F4): with an
+// installed engine's Python — its torch and transformers, already matched to
+// the node's CUDA — and a small overlay of what training needs besides, so a
+// node that serves can train without a second multi-gigabyte environment.
+package train
+
+import (
+	"archive/tar"
+	"bufio"
+	"bytes"
+	"compress/gzip"
+	"context"
+	"crypto/sha256"
+	_ "embed"
+	"encoding/hex"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"log/slog"
+	"net/http"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"sort"
+	"strconv"
+	"strings"
+	"sync"
+	"syscall"
+	"time"
+
+	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/supervisor"
+)
+
+//go:embed trainer.py
+var trainerPy []byte
+
+// overlayPackages are what training needs besides an engine's Python, pinned
+// and installed without their dependencies: those are the engine's own, and
+// letting an installer resolve them pulled a second torch.
+var overlayPackages = []string{"peft==0.21.2", "accelerate==1.15.0"}
+
+// Job is one training job, as the control plane sent it.
+type Job struct {
+	ID           string
+	BaseRepo     string
+	BaseRevision string
+	// Data is NDJSON of {"q", "a"}.
+	Data []byte
+	// Params is the trainer's settings as JSON.
+	Params       string
+	GPUMemoryMiB int
+	GGUF         bool
+}
+
+// Status is how a job is going.
+type Status struct {
+	State   string
+	Message string
+	Step    int
+	Steps   int
+	Loss    float64
+	PeakMiB int
+}
+
+// Env is where a node keeps what training needs.
+type Env struct {
+	// EnginesDir is where engines are installed (the agent's engines_dir).
+	EnginesDir string
+	// WorkDir is the trainer's own directory: its overlay, the llama.cpp
+	// conversion code, and each job's files while it runs.
+	WorkDir string
+	// RunAs is the user jobs run as: the engines' user.
+	RunAs string
+	Log   *slog.Logger
+	// Python, UV, ConvDir and GPUs replace what would be found, for tests.
+	Python, UV, ConvDir string
+	GPUs                func(ctx context.Context) ([]GPU, error)
+}
+
+// GPU is one card and how much of it is free.
+type GPU struct {
+	Index   int
+	FreeMiB int
+}
+
+// Runner runs one job at a time.
+type Runner struct {
+	env    Env
+	mu     sync.Mutex
+	cancel context.CancelFunc
+	job    string
+}
+
+// New is a runner.
+func New(env Env) *Runner {
+	if env.Log == nil {
+		env.Log = slog.New(slog.DiscardHandler)
+	}
+	return &Runner{env: env}
+}
+
+// Cancel stops a job if it is the one running.
+func (r *Runner) Cancel(id string) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	if r.job == id && r.cancel != nil {
+		r.cancel()
+	}
+}
+
+// ErrBusy is a node already training.
+var ErrBusy = errors.New("this node is already training")
+
+// Run runs a job to its end, telling status as it goes and artifact each
+// piece of what it made. It answers the error it failed with.
+func (r *Runner) Run(ctx context.Context, job Job, status func(Status), artifact func(kind string, seq int, data []byte, last bool) error) error {
+	r.mu.Lock()
+	if r.job != "" {
+		r.mu.Unlock()
+		return ErrBusy
+	}
+	ctx, cancel := context.WithCancel(ctx)
+	r.job, r.cancel = job.ID, cancel
+	r.mu.Unlock()
+	defer func() {
+		cancel()
+		r.mu.Lock()
+		r.job, r.cancel = "", nil
+		r.mu.Unlock()
+	}()
+
+	status(Status{State: "preparing", Message: "preparing the trainer"})
+	py, err := r.python()
+	if err != nil {
+		return err
+	}
+	overlay, err := r.overlay(ctx, py)
+	if err != nil {
+		return fmt.Errorf("installing %s: %w", strings.Join(overlayPackages, ", "), err)
+	}
+	conv := ""
+	if job.GGUF {
+		if conv, err = r.conversion(ctx); err != nil {
+			return fmt.Errorf("fetching llama.cpp's conversion code: %w", err)
+		}
+	}
+	gpu, err := r.pickGPU(ctx, job.GPUMemoryMiB)
+	if err != nil {
+		return err
+	}
+
+	dir := filepath.Join(r.env.WorkDir, "jobs", job.ID)
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		return err
+	}
+	defer os.RemoveAll(dir)
+	out := filepath.Join(dir, "adapter")
+	var params map[string]any
+	if job.Params != "" {
+		if err := json.Unmarshal([]byte(job.Params), &params); err != nil {
+			return fmt.Errorf("the job's params: %w", err)
+		}
+	}
+	spec, _ := json.Marshal(map[string]any{"base_repo": job.BaseRepo, "base_revision": job.BaseRevision,
+		"data": filepath.Join(dir, "data.jsonl"), "out": out, "params": params})
+	for name, b := range map[string][]byte{"data.jsonl": job.Data, "job.json": spec, "trainer.py": trainerPy} {
+		if err := os.WriteFile(filepath.Join(dir, name), b, 0o644); err != nil {
+			return err
+		}
+	}
+
+	var baseDir string
+	err = r.python3(ctx, py, dir, []string{"PYTHONPATH=" + overlay, "CUDA_VISIBLE_DEVICES=" + strconv.Itoa(gpu)},
+		[]string{filepath.Join(dir, "trainer.py"), filepath.Join(dir, "job.json")},
+		func(line map[string]any) {
+			s := Status{State: str(line["state"]), Message: str(line["message"]), Step: num(line["step"]),
+				Steps: num(line["steps"]), PeakMiB: num(line["peak_mib"])}
+			if v, ok := line["loss"].(float64); ok {
+				s.Loss = v
+			}
+			if b := str(line["base_dir"]); b != "" {
+				baseDir = b
+			}
+			if s.State == "trained" {
+				s.State = "running"
+				s.Message = "trained; packing the adapter"
+			}
+			status(s)
+		})
+	if err != nil {
+		return err
+	}
+
+	if job.GGUF {
+		status(Status{State: "running", Message: "converting the adapter to GGUF for llama.cpp"})
+		gguf := filepath.Join(dir, "adapter.gguf")
+		err := r.python3(ctx, py, dir, []string{"PYTHONPATH=" + strings.Join([]string{filepath.Join(conv, "gguf-py"), conv, overlay}, ":")},
+			[]string{filepath.Join(conv, "convert_lora_to_gguf.py"), "--base", baseDir, "--outtype", "f16", "--outfile", gguf, out}, nil)
+		if err != nil {
+			return fmt.Errorf("converting to GGUF: %w", err)
+		}
+		if err := sendFile(gguf, "gguf", artifact); err != nil {
+			return err
+		}
+	}
+	var tgz bytes.Buffer
+	if err := tarGz(&tgz, out); err != nil {
+		return err
+	}
+	return sendBytes(tgz.Bytes(), "peft", artifact)
+}
+
+func str(v any) string {
+	s, _ := v.(string)
+	return s
+}
+
+func num(v any) int {
+	f, _ := v.(float64)
+	return int(f)
+}
+
+// python3 runs a Python script as the engines' user, passing each JSON line
+// it prints to line, and answers the last of what it wrote to stderr when it
+// fails.
+func (r *Runner) python3(ctx context.Context, py, dir string, env, args []string, line func(map[string]any)) error {
+	cmd := exec.CommandContext(ctx, py, args...)
+	cmd.Dir = dir
+	cmd.Env = append(os.Environ(), "PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True", "TOKENIZERS_PARALLELISM=false",
+		"HF_HUB_DISABLE_PROGRESS_BARS=1", "PYTHONUNBUFFERED=1")
+	cmd.Env = append(cmd.Env, env...)
+	home, give, why := supervisor.AsEngineUser(cmd, r.env.RunAs, r.env.WorkDir)
+	if why != "" {
+		r.env.Log.Warn(why)
+	}
+	if home != "" {
+		cmd.Env = append(cmd.Env, "HF_HOME="+filepath.Join(home, ".cache", "huggingface"))
+	} else {
+		cmd.Env = append(cmd.Env, "HF_HOME="+filepath.Join(r.env.WorkDir, "hf"))
+	}
+	if err := give(dir); err != nil {
+		return fmt.Errorf("handing %s to %s: %w", dir, r.env.RunAs, err)
+	}
+	// The whole process group goes when the job is cancelled: torch starts
+	// workers of its own.
+	if cmd.SysProcAttr == nil {
+		cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true}
+	}
+	cmd.Cancel = func() error { return syscall.Kill(-cmd.Process.Pid, syscall.SIGKILL) }
+	stdout, err := cmd.StdoutPipe()
+	if err != nil {
+		return err
+	}
+	var stderr tailBuffer
+	cmd.Stderr = &stderr
+	if err := cmd.Start(); err != nil {
+		return err
+	}
+	sc := bufio.NewScanner(stdout)
+	sc.Buffer(make([]byte, 64<<10), 1<<20)
+	for sc.Scan() {
+		var m map[string]any
+		if line != nil && json.Unmarshal(sc.Bytes(), &m) == nil {
+			line(m)
+		}
+	}
+	err = cmd.Wait()
+	if ctx.Err() != nil {
+		return errors.New("cancelled")
+	}
+	if err != nil {
+		return fmt.Errorf("%v: %s", err, stderr.last())
+	}
+	return nil
+}
+
+// tailBuffer keeps the end of what a process wrote: the reason it failed is
+// at the end, and the rest is progress bars.
+type tailBuffer struct{ b []byte }
+
+func (t *tailBuffer) Write(p []byte) (int, error) {
+	t.b = append(t.b, p...)
+	if len(t.b) > 16<<10 {
+		t.b = t.b[len(t.b)-16<<10:]
+	}
+	return len(p), nil
+}
+
+func (t *tailBuffer) last() string {
+	lines := strings.Split(strings.TrimSpace(string(t.b)), "\n")
+	if len(lines) > 3 {
+		lines = lines[len(lines)-3:]
+	}
+	return strings.Join(lines, " | ")
+}
+
+// python is the newest installed vLLM's interpreter: its torch is built for
+// this node's CUDA.
+func (r *Runner) python() (string, error) {
+	if r.env.Python != "" {
+		return r.env.Python, nil
+	}
+	found, _ := filepath.Glob(filepath.Join(r.env.EnginesDir, "vllm", "*", "bin", "python3"))
+	if len(found) == 0 {
+		return "", errors.New("training needs vLLM installed on this node (its torch is what trains); " +
+			"llmgw ctl engine-build, or serving any vLLM model here, installs it")
+	}
+	sort.Strings(found)
+	return found[len(found)-1], nil
+}
+
+func (r *Runner) uv() (string, error) {
+	if r.env.UV != "" {
+		return r.env.UV, nil
+	}
+	var found []string
+	_ = filepath.WalkDir(filepath.Join(r.env.EnginesDir, "uv"), func(p string, d os.DirEntry, err error) error {
+		if err == nil && !d.IsDir() && d.Name() == "uv" {
+			found = append(found, p)
+		}
+		return nil
+	})
+	if len(found) == 0 {
+		return "", errors.New("no uv under " + filepath.Join(r.env.EnginesDir, "uv"))
+	}
+	sort.Strings(found)
+	return found[len(found)-1], nil
+}
+
+// overlay installs the pinned packages once per Python and set of pins.
+func (r *Runner) overlay(ctx context.Context, py string) (string, error) {
+	sum := sha256.Sum256([]byte(py + "\x00" + strings.Join(overlayPackages, " ")))
+	dir := filepath.Join(r.env.WorkDir, "overlay-"+hex.EncodeToString(sum[:6]))
+	if _, err := os.Stat(filepath.Join(dir, ".done")); err == nil {
+		return dir, nil
+	}
+	uv, err := r.uv()
+	if err != nil {
+		return "", err
+	}
+	tmp := dir + ".new"
+	_ = os.RemoveAll(tmp)
+	args := append([]string{"pip", "install", "--quiet", "--python", py, "--no-deps", "--target", tmp}, overlayPackages...)
+	cmd := exec.CommandContext(ctx, uv, args...)
+	if out, err := cmd.CombinedOutput(); err != nil {
+		return "", fmt.Errorf("%v: %s", err, strings.TrimSpace(string(out)))
+	}
+	_ = os.WriteFile(filepath.Join(tmp, ".done"), nil, 0o644)
+	_ = os.RemoveAll(dir)
+	if err := os.Rename(tmp, dir); err != nil {
+		return "", err
+	}
+	// Readable by the engines' user, which runs the trainer.
+	_ = filepath.WalkDir(dir, func(p string, d os.DirEntry, err error) error {
+		if err == nil {
+			mode := os.FileMode(0o644)
+			if d.IsDir() {
+				mode = 0o755
+			}
+			_ = os.Chmod(p, mode)
+		}
+		return nil
+	})
+	return dir, nil
+}
+
+// llamaCppTag is the release of the newest llama.cpp installed: its
+// conversion code is the one whose GGUF that llama.cpp reads.
+func (r *Runner) llamaCppTag() (string, error) {
+	found, _ := filepath.Glob(filepath.Join(r.env.EnginesDir, "llamacpp", "b*"))
+	if len(found) == 0 {
+		return "", errors.New("a GGUF adapter needs llama.cpp installed on this node, to match its release")
+	}
+	sort.Strings(found)
+	tag, _, _ := strings.Cut(filepath.Base(found[len(found)-1]), "-")
+	return tag, nil
+}
+
+// conversionSource is where llama.cpp's source is fetched from, by tag.
+var conversionSource = "https://codeload.github.com/ggml-org/llama.cpp/tar.gz/refs/tags/"
+
+// conversion fetches llama.cpp's LoRA conversion code for the installed
+// release, once.
+func (r *Runner) conversion(ctx context.Context) (string, error) {
+	if r.env.ConvDir != "" {
+		return r.env.ConvDir, nil
+	}
+	tag, err := r.llamaCppTag()
+	if err != nil {
+		return "", err
+	}
+	dir := filepath.Join(r.env.WorkDir, "llama.cpp-"+tag)
+	if _, err := os.Stat(filepath.Join(dir, "convert_lora_to_gguf.py")); err == nil {
+		return dir, nil
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, conversionSource+tag, nil)
+	if err != nil {
+		return "", err
+	}
+	res, err := http.DefaultClient.Do(req)
+	if err != nil {
+		return "", err
+	}
+	defer res.Body.Close()
+	if res.StatusCode != http.StatusOK {
+		return "", fmt.Errorf("%s answered %s", conversionSource+tag, res.Status)
+	}
+	tmp := dir + ".new"
+	_ = os.RemoveAll(tmp)
+	if err := untarConversion(res.Body, tmp); err != nil {
+		return "", err
+	}
+	_ = os.RemoveAll(dir)
+	return dir, os.Rename(tmp, dir)
+}
+
+// untarConversion keeps, of llama.cpp's source, the conversion scripts and
+// the Python packages they import.
+func untarConversion(r io.Reader, dir string) error {
+	gz, err := gzip.NewReader(r)
+	if err != nil {
+		return err
+	}
+	tr := tar.NewReader(gz)
+	kept := 0
+	for {
+		h, err := tr.Next()
+		if errors.Is(err, io.EOF) {
+			break
+		}
+		if err != nil {
+			return err
+		}
+		_, rel, ok := strings.Cut(h.Name, "/")
+		if !ok || h.Typeflag != tar.TypeReg || strings.Contains(rel, "..") {
+			continue
+		}
+		if rel != "convert_lora_to_gguf.py" && rel != "convert_hf_to_gguf.py" &&
+			!strings.HasPrefix(rel, "conversion/") && !strings.HasPrefix(rel, "gguf-py/") {
+			continue
+		}
+		p := filepath.Join(dir, filepath.FromSlash(rel))
+		if err := os.MkdirAll(filepath.Dir(p), 0o755); err != nil {
+			return err
+		}
+		f, err := os.OpenFile(p, os.O_CREATE|os.O_WRONLY|os.O_TRUNC, 0o644)
+		if err != nil {
+			return err
+		}
+		_, err = io.Copy(f, io.LimitReader(tr, 64<<20))
+		f.Close()
+		if err != nil {
+			return err
+		}
+		kept++
+	}
+	if kept == 0 {
+		return errors.New("no conversion code in llama.cpp's source")
+	}
+	return nil
+}
+
+// pickGPU is the card with the most memory free, if that is enough.
+func (r *Runner) pickGPU(ctx context.Context, need int) (int, error) {
+	list := r.env.GPUs
+	if list == nil {
+		list = nvidiaSMI
+	}
+	gpus, err := list(ctx)
+	if err != nil {
+		return 0, fmt.Errorf("reading the cards' free memory: %w", err)
+	}
+	if len(gpus) == 0 {
+		return 0, errors.New("this node has no GPU to train on")
+	}
+	best := gpus[0]
+	for _, g := range gpus[1:] {
+		if g.FreeMiB > best.FreeMiB {
+			best = g
+		}
+	}
+	if best.FreeMiB < need {
+		return 0, fmt.Errorf("no card here has %d MiB free (the most free is GPU %d with %d MiB); "+
+			"a smaller base, or a card freed for it, would do", need, best.Index, best.FreeMiB)
+	}
+	return best.Index, nil
+}
+
+func nvidiaSMI(ctx context.Context) ([]GPU, error) {
+	ctx, cancel := context.WithTimeout(ctx, 15*time.Second)
+	defer cancel()
+	out, err := exec.CommandContext(ctx, "nvidia-smi", "--query-gpu=index,memory.free", "--format=csv,noheader,nounits").Output()
+	if err != nil {
+		return nil, err
+	}
+	return parseGPUs(string(out)), nil
+}
+
+func parseGPUs(s string) []GPU {
+	var out []GPU
+	for _, line := range strings.Split(strings.TrimSpace(s), "\n") {
+		a, b, ok := strings.Cut(line, ",")
+		if !ok {
+			continue
+		}
+		i, err1 := strconv.Atoi(strings.TrimSpace(a))
+		f, err2 := strconv.Atoi(strings.TrimSpace(b))
+		if err1 == nil && err2 == nil {
+			out = append(out, GPU{Index: i, FreeMiB: f})
+		}
+	}
+	return out
+}
+
+// tarGz packs a directory.
+func tarGz(w io.Writer, dir string) error {
+	gz := gzip.NewWriter(w)
+	tw := tar.NewWriter(gz)
+	err := filepath.WalkDir(dir, func(p string, d os.DirEntry, err error) error {
+		if err != nil || d.IsDir() {
+			return err
+		}
+		rel, _ := filepath.Rel(dir, p)
+		info, err := d.Info()
+		if err != nil {
+			return err
+		}
+		if err := tw.WriteHeader(&tar.Header{Name: filepath.ToSlash(rel), Mode: 0o644, Size: info.Size(), ModTime: info.ModTime()}); err != nil {
+			return err
+		}
+		f, err := os.Open(p)
+		if err != nil {
+			return err
+		}
+		defer f.Close()
+		_, err = io.Copy(tw, f)
+		return err
+	})
+	if err != nil {
+		return err
+	}
+	if err := tw.Close(); err != nil {
+		return err
+	}
+	return gz.Close()
+}
+
+// artifactChunk is how much of an artifact goes in one message.
+const artifactChunk = 1 << 20
+
+func sendFile(path, kind string, send func(string, int, []byte, bool) error) error {
+	b, err := os.ReadFile(path)
+	if err != nil {
+		return err
+	}
+	return sendBytes(b, kind, send)
+}
+
+func sendBytes(b []byte, kind string, send func(string, int, []byte, bool) error) error {
+	seq := 0
+	for {
+		n := min(artifactChunk, len(b))
+		last := n == len(b)
+		if err := send(kind, seq, b[:n], last); err != nil {
+			return err
+		}
+		if last {
+			return nil
+		}
+		b = b[n:]
+		seq++
+	}
+}

internal/agent/train/runner_test.go

@@ -0,0 +1,164 @@
+package train
+
+import (
+	"archive/tar"
+	"bytes"
+	"compress/gzip"
+	"context"
+	"io"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+)
+
+// fakePython stands in for an engine's interpreter: as the trainer it
+// prints progress and writes an adapter; as the converter it writes a GGUF;
+// told to fail, it fails.
+const fakePython = `#!/bin/sh
+case "$1" in
+*trainer.py)
+  out=$(sed -n 's/.*"out": *"\([^"]*\)".*/\1/p' "$2")
+  grep -q FAIL "$(dirname "$2")/data.jsonl" && { echo "CUDA out of memory" >&2; exit 1; }
+  echo '{"state":"preparing","message":"fetching base"}'
+  echo '{"state":"running","step":5,"steps":10,"loss":0.9,"peak_mib":1900}'
+  mkdir -p "$out"; echo '{"r":16}' > "$out/adapter_config.json"; echo weights > "$out/adapter_model.safetensors"
+  echo '{"state":"trained","step":10,"steps":10,"peak_mib":2100,"base_dir":"/base"}'
+  ;;
+*convert_lora_to_gguf.py)
+  while [ $# -gt 0 ]; do [ "$1" = "--outfile" ] && echo GGUF > "$2"; shift; done
+  ;;
+esac
+`
+
+func runner(t *testing.T, free int) *Runner {
+	t.Helper()
+	dir := t.TempDir()
+	py := filepath.Join(dir, "python3")
+	_ = os.WriteFile(py, []byte(fakePython), 0o755)
+	uv := filepath.Join(dir, "uv")
+	_ = os.WriteFile(uv, []byte("#!/bin/sh\nwhile [ $# -gt 0 ]; do [ \"$1\" = \"--target\" ] && mkdir -p \"$2\"; shift; done\n"), 0o755)
+	return New(Env{WorkDir: filepath.Join(dir, "work"), Python: py, UV: uv, ConvDir: dir,
+		GPUs: func(context.Context) ([]GPU, error) { return []GPU{{0, 500}, {1, free}}, nil }})
+}
+
+type sink struct {
+	mu       sync.Mutex
+	statuses []Status
+	parts    map[string][]byte
+	lasts    int
+}
+
+func (s *sink) status(st Status) { s.mu.Lock(); s.statuses = append(s.statuses, st); s.mu.Unlock() }
+func (s *sink) artifact(kind string, _ int, b []byte, last bool) error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.parts == nil {
+		s.parts = map[string][]byte{}
+	}
+	s.parts[kind] = append(s.parts[kind], b...)
+	if last {
+		s.lasts++
+	}
+	return nil
+}
+
+func TestAJobTrainsOnTheFreestCardAndSendsWhatItMade(t *testing.T) {
+	r := runner(t, 3000)
+	s := &sink{}
+	err := r.Run(context.Background(), Job{ID: "j1", BaseRepo: "Qwen/x", Data: []byte(`{"q":"a","a":"b"}`), GPUMemoryMiB: 2500, GGUF: true},
+		s.status, s.artifact)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var running bool
+	for _, st := range s.statuses {
+		if st.State == "running" && st.Step == 5 && st.Steps == 10 && st.PeakMiB == 1900 {
+			running = true
+		}
+	}
+	if !running {
+		t.Errorf("statuses %+v", s.statuses)
+	}
+	if string(s.parts["gguf"]) != "GGUF\n" || s.lasts != 2 {
+		t.Errorf("gguf %q, lasts %d", s.parts["gguf"], s.lasts)
+	}
+	files := map[string]string{}
+	gz, err := gzip.NewReader(bytes.NewReader(s.parts["peft"]))
+	if err != nil {
+		t.Fatal(err)
+	}
+	tr := tar.NewReader(gz)
+	for {
+		h, err := tr.Next()
+		if err != nil {
+			break
+		}
+		b, _ := io.ReadAll(tr)
+		files[h.Name] = string(b)
+	}
+	if files["adapter_model.safetensors"] != "weights\n" || files["adapter_config.json"] == "" {
+		t.Errorf("peft archive %v", files)
+	}
+	// Nothing of the job is left on the node.
+	if entries, _ := os.ReadDir(filepath.Join(r.env.WorkDir, "jobs")); len(entries) != 0 {
+		t.Errorf("left behind: %v", entries)
+	}
+}
+
+func TestAJobWaitsForNoCardItDoesNotFit(t *testing.T) {
+	r := runner(t, 1000)
+	err := r.Run(context.Background(), Job{ID: "j1", Data: []byte(`{"q":"a","a":"b"}`), GPUMemoryMiB: 2500}, (&sink{}).status, (&sink{}).artifact)
+	if err == nil || !strings.Contains(err.Error(), "no card here has 2500 MiB free") {
+		t.Errorf("got %v", err)
+	}
+}
+
+func TestATrainerThatFailsSaysWhy(t *testing.T) {
+	r := runner(t, 3000)
+	err := r.Run(context.Background(), Job{ID: "j1", Data: []byte(`{"q":"FAIL","a":"b"}`)}, (&sink{}).status, (&sink{}).artifact)
+	if err == nil || !strings.Contains(err.Error(), "CUDA out of memory") {
+		t.Errorf("got %v", err)
+	}
+}
+
+func TestOneJobAtATime(t *testing.T) {
+	r := runner(t, 3000)
+	r.job = "other"
+	if err := r.Run(context.Background(), Job{ID: "j2"}, (&sink{}).status, (&sink{}).artifact); err != ErrBusy {
+		t.Errorf("got %v", err)
+	}
+}
+
+func TestCardsAreReadFromNvidiaSMI(t *testing.T) {
+	got := parseGPUs("0, 2310\n1, 80000\nbad line\n")
+	if len(got) != 2 || got[1].FreeMiB != 80000 || got[0].Index != 0 {
+		t.Errorf("parsed %+v", got)
+	}
+}
+
+func TestOnlyTheConversionCodeIsKeptOfLlamaCpp(t *testing.T) {
+	var buf bytes.Buffer
+	gz := gzip.NewWriter(&buf)
+	tw := tar.NewWriter(gz)
+	for _, name := range []string{"llama.cpp-b1/convert_lora_to_gguf.py", "llama.cpp-b1/conversion/qwen.py",
+		"llama.cpp-b1/gguf-py/gguf/__init__.py", "llama.cpp-b1/src/llama.cpp", "llama.cpp-b1/../evil"} {
+		_ = tw.WriteHeader(&tar.Header{Name: name, Mode: 0o644, Size: 1, Typeflag: tar.TypeReg})
+		_, _ = tw.Write([]byte("x"))
+	}
+	_ = tw.Close()
+	_ = gz.Close()
+	dir := filepath.Join(t.TempDir(), "conv")
+	if err := untarConversion(&buf, dir); err != nil {
+		t.Fatal(err)
+	}
+	for _, want := range []string{"convert_lora_to_gguf.py", "conversion/qwen.py", "gguf-py/gguf/__init__.py"} {
+		if _, err := os.Stat(filepath.Join(dir, want)); err != nil {
+			t.Errorf("%s not kept", want)
+		}
+	}
+	if _, err := os.Stat(filepath.Join(dir, "src")); err == nil {
+		t.Error("the rest of llama.cpp was kept")
+	}
+}

internal/agent/train/trainer.py

@@ -0,0 +1,101 @@
+# llmgw's LoRA trainer (BACKLOG F4), run by the node agent with an installed
+# engine's Python (its torch and transformers) and a small overlay (peft,
+# accelerate). Reads a job, trains, writes the PEFT adapter, and says how it
+# goes as one JSON object a line on standard output.
+import json, math, os, random, sys, threading, subprocess, time
+
+def say(**kw):
+    print(json.dumps(kw), flush=True)
+
+job = json.load(open(sys.argv[1], encoding="utf-8"))
+p = {"rank": 16, "alpha": 32, "dropout": 0.05, "epochs": 6, "lr": 2e-4, "batch": 4, "max_len": 512, "seed": 0}
+p.update(job.get("params") or {})
+
+say(state="preparing", message="loading libraries")
+import torch
+from huggingface_hub import snapshot_download
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import LoraConfig, get_peft_model
+
+random.seed(p["seed"]); torch.manual_seed(p["seed"])
+if not torch.cuda.is_available():
+    raise SystemExit("no CUDA device is visible to the trainer")
+
+say(state="preparing", message="fetching " + job["base_repo"])
+base_dir = snapshot_download(job["base_repo"], revision=job.get("base_revision") or None,
+                             allow_patterns=["*.json", "*.safetensors", "*.txt", "*.model", "*.tiktoken"])
+say(state="preparing", message="loading the base model", base_dir=base_dir)
+
+# bf16 where the card has it; fp16 with a loss scaler where it does not
+# (Turing and older).
+bf16 = torch.cuda.is_bf16_supported()
+dtype = torch.bfloat16 if bf16 else torch.float16
+tok = AutoTokenizer.from_pretrained(base_dir)
+if tok.pad_token_id is None:
+    tok.pad_token = tok.eos_token
+model = AutoModelForCausalLM.from_pretrained(base_dir, dtype=dtype).to("cuda")
+
+rows = [json.loads(l) for l in open(job["data"], encoding="utf-8") if l.strip()]
+rows = [r for r in rows if r.get("q") and r.get("a")]
+if not rows:
+    raise SystemExit("no {q, a} pairs in the data")
+
+def encode(r):
+    prompt = tok.apply_chat_template([{"role": "user", "content": r["q"]}], add_generation_prompt=True, tokenize=False)
+    full = tok.apply_chat_template([{"role": "user", "content": r["q"]}, {"role": "assistant", "content": r["a"]}], tokenize=False)
+    pi = tok(prompt, add_special_tokens=False)["input_ids"]
+    fi = tok(full, add_special_tokens=False)["input_ids"][: p["max_len"]]
+    return fi, ([-100] * len(pi) + fi[len(pi):])[: len(fi)]
+
+data = [encode(r) for r in rows]
+
+model.gradient_checkpointing_enable()
+model.enable_input_require_grads()
+model = get_peft_model(model, LoraConfig(r=int(p["rank"]), lora_alpha=int(p["alpha"]), lora_dropout=float(p["dropout"]),
+    task_type="CAUSAL_LM", target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]))
+trainable = [x for x in model.parameters() if x.requires_grad]
+for x in trainable:  # fp32 trainable weights, so a loss scaler can unscale them
+    x.data = x.data.float()
+
+peak = {"mib": 0}
+stop = threading.Event()
+def watch():
+    while not stop.is_set():
+        try:
+            out = subprocess.run(["nvidia-smi", "--query-compute-apps=pid,used_memory", "--format=csv,noheader,nounits"],
+                                 capture_output=True, text=True, timeout=5).stdout
+            for line in out.splitlines():
+                pid, mib = [v.strip() for v in line.split(",")]
+                if int(pid) == os.getpid():
+                    peak["mib"] = max(peak["mib"], int(mib))
+        except Exception:
+            pass
+        time.sleep(0.5)
+threading.Thread(target=watch, daemon=True).start()
+
+opt = torch.optim.AdamW(trainable, lr=float(p["lr"]), weight_decay=0.0)
+scaler = torch.amp.GradScaler("cuda", enabled=not bf16)
+bs, epochs = int(p["batch"]), int(p["epochs"])
+steps = epochs * math.ceil(len(data) / bs)
+warm = max(1, steps // 10)
+sched = torch.optim.lr_scheduler.LambdaLR(opt, lambda s: min(1.0, (s + 1) / warm) * max(0.0, 1 - s / steps))
+model.train(); step = 0; last = time.time()
+for ep in range(epochs):
+    random.shuffle(data)
+    for i in range(0, len(data), bs):
+        batch = data[i:i + bs]; n = max(len(x[0]) for x in batch)
+        ids = torch.tensor([x[0] + [tok.pad_token_id] * (n - len(x[0])) for x in batch], device="cuda")
+        lab = torch.tensor([x[1] + [-100] * (n - len(x[1])) for x in batch], device="cuda")
+        att = torch.tensor([[1] * len(x[0]) + [0] * (n - len(x[0])) for x in batch], device="cuda")
+        with torch.autocast("cuda", dtype=dtype):
+            loss = model(input_ids=ids, attention_mask=att, labels=lab).loss
+        scaler.scale(loss).backward(); scaler.unscale_(opt)
+        torch.nn.utils.clip_grad_norm_(trainable, 1.0)
+        scaler.step(opt); scaler.update(); opt.zero_grad(set_to_none=True); sched.step(); step += 1
+        if time.time() - last > 2 or step == steps:
+            last = time.time()
+            say(state="running", message="training", step=step, steps=steps, loss=round(loss.item(), 4), peak_mib=peak["mib"])
+
+model.save_pretrained(job["out"])
+stop.set()
+say(state="trained", step=step, steps=steps, peak_mib=peak["mib"], base_dir=base_dir)

internal/agent/trainjob.go

@@ -0,0 +1,46 @@
+package agent
+
+import (
+	"context"
+
+	llmgwv1 "gitea.mynodes.xyz/ilker/llmgw/api/gen/llmgw/v1"
+	"gitea.mynodes.xyz/ilker/llmgw/internal/agent/train"
+)
+
+// trainJob runs a training job the control plane sent, telling it how the
+// job goes and sending what it made. A stream that ends ends the job: the
+// control plane, which sees the node go, marks it failed, and it is started
+// again from the beginning rather than resumed.
+func (a *Agent) trainJob(ctx context.Context, m *llmgwv1.TrainStart, out chan<- *llmgwv1.AgentMessage) {
+	id := m.GetJobId()
+	send := func(msg *llmgwv1.AgentMessage) bool {
+		select {
+		case out <- msg:
+			return true
+		case <-ctx.Done():
+			return false
+		}
+	}
+	status := func(s train.Status) {
+		send(&llmgwv1.AgentMessage{Msg: &llmgwv1.AgentMessage_TrainStatus{TrainStatus: &llmgwv1.TrainStatus{
+			JobId: id, State: s.State, Message: s.Message, Step: int32(s.Step), Steps: int32(s.Steps), //nolint:gosec // counts
+			Loss: s.Loss, PeakMib: int32(s.PeakMiB)}}}) //nolint:gosec // MiB
+	}
+	artifact := func(kind string, seq int, data []byte, last bool) error {
+		if !send(&llmgwv1.AgentMessage{Msg: &llmgwv1.AgentMessage_TrainArtifact{TrainArtifact: &llmgwv1.TrainArtifact{
+			JobId: id, Kind: kind, Seq: int32(seq), Data: data, Last: last}}}) { //nolint:gosec // chunk count
+			return ctx.Err()
+		}
+		return nil
+	}
+	a.log.Info("training an adapter", "job", id, "base", m.GetBaseRepo())
+	err := a.trainer.Run(ctx, train.Job{ID: id, BaseRepo: m.GetBaseRepo(), BaseRevision: m.GetBaseRevision(),
+		Data: m.GetData(), Params: m.GetParams(), GPUMemoryMiB: int(m.GetGpuMemoryMib()), GGUF: m.GetGguf()}, status, artifact)
+	if err != nil {
+		a.log.Warn("training failed", "job", id, "err", err)
+		status(train.Status{State: "failed", Message: err.Error()})
+		return
+	}
+	a.log.Info("trained an adapter", "job", id)
+	status(train.Status{State: "done"})
+}

internal/agent/upgrade.go

@@ -35,12 +35,25 @@ func (a *Agent) upgrade(ctx context.Context, m *llmgwv1.UpgradeAgent, out chan<-
 		select {
 		case out <- &llmgwv1.AgentMessage{Msg: &llmgwv1.AgentMessage_Upgrade{Upgrade: reply}}:
 		case <-ctx.Done():
+		case <-time.After(5 * time.Second):
 		}
 	}
+	// One at a time: a second request while one downloads would race it to
+	// the same file.
+	if !a.upgrading.CompareAndSwap(false, true) {
+		reply.Error = "an upgrade is already under way on this node"
+		send()
+		return
+	}
+	defer a.upgrading.Store(false)
 	a.log.Info("upgrading", "to", m.GetVersion(), "from", a.cfg.Version)
 	self, err := upgradeTarget()
 	if err == nil {
-		dctx, cancel := context.WithTimeout(ctx, 10*time.Minute)
+		// Not the stream's context: a stream that breaks while the release
+		// downloads — a control plane restarting, leadership moving — cancelled
+		// the upgrade with nobody told. The node registering again on the
+		// version is what the control plane watches for.
+		dctx, cancel := context.WithTimeout(context.WithoutCancel(ctx), upgradeDownload)
 		_, err = release.Install(dctx, nil, m.GetReleases(), m.GetVersion(), runtime.GOOS, runtime.GOARCH, self)
 		cancel()
 	}
@@ -53,6 +66,13 @@ func (a *Agent) upgrade(ctx context.Context, m *llmgwv1.UpgradeAgent, out chan<-
 	send()
 	a.log.Info("upgraded; restarting into the new version", "to", m.GetVersion())
 	// A moment for the reply to leave before the stream goes.
-	time.Sleep(500 * time.Millisecond)
+	time.Sleep(upgradeLinger)
 	exitForUpgrade()
 }
+
+// upgradeDownload bounds an upgrade's download; the control plane waits
+// longer (agentsvc.UpgradeWait). upgradeLinger is the moment its answer has
+// to leave before the agent exits; a variable so tests need not wait it.
+const upgradeDownload = 10 * time.Minute
+
+var upgradeLinger = 2 * time.Second