@atomic-chat/core 0.5.1 → 0.13.2

package/dist/runtime/managed-engines/runtime.js

@@ -0,0 +1,402 @@
+/**
+ * The one `LocalRuntime` every managed engine is (change `add-vllm-runtime`, design D3; before it, the
+ * `tensorrt-llm` provider of change `add-tensorrt-llm-linux`, task 2.14): a provider on the
+ * engine-neutral managed-text lifecycle (`../managed-text/`), parameterized by the engine's
+ * `ManagedEngineSpec` — its adapter, settings, memory rule and route policy. What this file adds on
+ * top of the lifecycle is the provider's own part of a load:
+ *
+ *  - refusing it before any container exists — no Docker, no `ready` installation
+ *    (`MANAGED_ADAPTER_UNAVAILABLE`), settings the spec rejects (`INVALID_ARGUMENT`), a model that
+ *    is not installed, a host with no NVIDIA card, an embedding request;
+ *  - the card: the saved `gpu_id` when the probe still finds it, otherwise the one with the most free
+ *    memory, ties by the most total memory (`selectLaunchGpu`, the same rule `POST /check` uses),
+ *    with the replacement reported on every load event;
+ *  - one session at a time, and one resident model per card: the lifecycle's `stopping-previous`
+ *    stage asks core's GPU residency (`claimGpu`) to stop every other model of this provider and
+ *    every other engine on the chosen card, each with a confirmed exit, before this one's container
+ *    is created — or, with no residency wired, stops this provider's other models itself;
+ *  - what a session can do (`routePolicy`, `capabilities`), read off the pinned descriptor's
+ *    `model_families` entry for the model's architecture — never guessed;
+ *  - never growing a context or recreating a session: a restart of a multi-minute container in the
+ *    middle of a conversation is worse than an honest `context_length_exceeded`;
+ *  - the pre-launch check, split in two: before `stopPrevious` runs, `verifyModelFilesAndCompatibility`
+ *    (`prelaunch.ts`) re-verifies every file `model.yml` recorded present on disk at its size, and
+ *    recomputes architecture/format/compute-capability against `config.json`/`hf_quant_config.json`
+ *    as they sit in the model's directory right now — never memory, which the model this load is
+ *    about to replace may still be holding on the very same card. Memory is checked after
+ *    `stopPrevious`, through the lifecycle's own `beforeCreate` hook (`checkMemoryBeforeCreate`
+ *    below), by the engine's own memory rule, against a re-probed card;
+ *  - `family` (tools/structured output/route policy) is read off the descriptor's `model_families`
+ *    entry for the *verified* `config.json` architecture — `model.yml`'s own copy of the
+ *    architecture is never consulted for this, since the pre-launch check may have found it stale.
+ *
+ * Everything that touches the machine arrives injected: the lifecycle over the core's one Docker
+ * executor, the ready installation, the host facts, the model lookup and the stored settings.
+ */
+import { AtomicCoreError } from '../../contracts/index.js';
+import { throwIfLoadCancelled } from '../shared/index.js';
+import { checkCheckpointMemory, selectLaunchGpu } from '../managed-models/index.js';
+import { verifyModelFilesAndCompatibility } from '../tensorrt-llm/prelaunch.js';
+const NONE = {
+    tools: false,
+    reasoning: false,
+    structured_output: false,
+    vision: false,
+    embeddings: false,
+    responses: false,
+};
+/** `capabilities()`'s own lookup, off `model.yml`'s architecture: no disk verification runs for it. */
+function familyOf(ready, model) {
+    if (model.architecture === null)
+        return null;
+    return ready.descriptor.model_families[model.architecture] ?? null;
+}
+/**
+ * The load path's own lookup (finding 5): off the *verified* `config.json` architecture
+ * (`ResolvedCheckpoint.architectures[0]`, from the pre-launch check), never `model.yml`'s — the two
+ * can disagree (a stale `model.yml`, or none at all), and only the one the compatibility check just
+ * confirmed against the descriptor's `supported_architectures` is trustworthy enough to gate what a
+ * session is allowed to do.
+ */
+function familyFromResolved(ready, resolved) {
+    const architecture = resolved.architectures[0];
+    return architecture === undefined ? null : (ready.descriptor.model_families[architecture] ?? null);
+}
+export class ManagedTextRuntime {
+    engine;
+    deps;
+    current = null;
+    closed = false;
+    /** Removals of the engine in progress: while any holds loads off, every load is refused (M-1). */
+    loadHolds = 0;
+    /** Deletions of one model in progress (task 2.24): each refuses loads of that model only, counted. */
+    modelHolds = new Map();
+    /** Generations whose GPU claim succeeded: a load only holds its card from then on. */
+    claimed = new Set();
+    /** What each loaded session can do, keyed by model and pinned to the generation it was computed for. */
+    sessionCapabilities = new Map();
+    constructor(engine, deps) {
+        this.engine = engine;
+        this.deps = deps;
+    }
+    list() {
+        return this.current?.list() ?? [];
+    }
+    findSession(modelId) {
+        return this.current?.findSession(modelId);
+    }
+    getLoadedModels() {
+        return this.list().map((session) => session.model_id);
+    }
+    isLoading(modelId) {
+        return this.current?.isLoading(modelId) ?? false;
+    }
+    async load(modelId, opts = {}) {
+        const { signal } = opts;
+        this.assertOpen();
+        this.assertNotHeldOff(modelId);
+        throwIfLoadCancelled(signal);
+        if (opts.isEmbedding) {
+            throw new AtomicCoreError('INVALID_ARGUMENT', `${this.engine.provider} models do not serve embeddings.`, modelId);
+        }
+        // Validated first, before anything is asked of the machine (spec: schema validation before start).
+        const settings = this.engine.settings(this.deps.settings(), opts.overrides);
+        const lifecycle = await this.deps.lifecycle().catch((cause) => {
+            throw new AtomicCoreError('MANAGED_ADAPTER_UNAVAILABLE', `The managed container runtime failed to initialise, so ${this.engine.provider} cannot run models.`, cause instanceof Error ? cause.message : String(cause));
+        });
+        if (lifecycle === null) {
+            throw new AtomicCoreError('MANAGED_ADAPTER_UNAVAILABLE', `Docker is not installed on this machine, so ${this.engine.provider} cannot run models.`);
+        }
+        // Kept from the first load on: what `list`, `unload` and `shutdown` act on. Nothing is loaded
+        // before that, so there is nothing for them to find earlier either.
+        this.current = lifecycle;
+        const ready = await this.deps.readyInstallation();
+        throwIfLoadCancelled(signal);
+        const model = await this.deps.model(modelId);
+        const facts = await this.deps.hostFacts();
+        throwIfLoadCancelled(signal);
+        this.assertOpen();
+        const gpu = selectLaunchGpu(facts.gpus, facts.memory, settings.gpu_id ?? undefined);
+        if (gpu === null) {
+            throw new AtomicCoreError('MANAGED_PREREQUISITE_BLOCKED', `No NVIDIA GPU was found on this machine, so ${this.engine.provider} cannot load a model.`);
+        }
+        const substituted = settings.gpu_id !== null && gpu.gpu_id !== settings.gpu_id
+            ? { requested_gpu_id: settings.gpu_id, gpu_id: gpu.gpu_id }
+            : undefined;
+        const check = this.engine.check(settings);
+        // Pre-launch check, phase 1 (task 2.16, spec "Проверка файлов при загрузке"; round 1, finding 1
+        // (Critical)): every file model.yml recorded is still on disk at its size, and architecture/
+        // format/compute-capability are recomputed against config.json/hf_quant_config.json as they sit
+        // on disk right now — never model.yml's word for it, and never memory, which the model this load
+        // is about to replace may still be holding on this very card. No container exists yet; a failure
+        // here never creates one.
+        const resolved = await verifyModelFilesAndCompatibility(model, ready.descriptor, facts.gpus, facts.memory, { gpuId: gpu.gpu_id, engine: check });
+        throwIfLoadCancelled(signal);
+        this.assertOpen();
+        // family from the *verified* config.json architecture, never model.yml's own (finding 5).
+        const family = familyFromResolved(ready, resolved);
+        const { descriptor } = ready;
+        // Again right before the lifecycle registers this load (synchronously, inside `lifecycle.load`):
+        // a removal that began while this load was probing the host must not find it past the check.
+        this.assertNotHeldOff(modelId);
+        const session = await lifecycle.load({
+            modelId,
+            modelPath: model.dir,
+            weightBytes: model.weightBytes,
+            installation: {
+                descriptor_id: descriptor.descriptor_id,
+                engine_id: descriptor.engine_id,
+                adapter_id: descriptor.adapter_id,
+                adapter_contract_version: descriptor.adapter_contract_version,
+                image: ready.image,
+            },
+            family,
+            gpuUuid: gpu.gpu_id,
+            // No VRAM of its own (design D13): the launch bounds the KV cache by tokens, as the memory
+            // check (`checkModelMemory`, same rule) reserved for it.
+            unifiedMemory: gpu.total_vram_bytes === null,
+            gpuTotalVramBytes: gpu.total_vram_bytes,
+            gpuComputeCapability: gpu.compute_capability,
+            selinux: facts.selinux,
+            settings,
+            // Always passed, evaluated when the stage runs: a second model that arrived a moment earlier
+            // is still found, so two loads racing each other can never both end up running.
+            stopPrevious: (stageSignal, generation) => this.stopOthers(lifecycle, { modelId, generation, gpuId: gpu.gpu_id }, stageSignal),
+            // Pre-launch check, phase 2 (finding 1): the memory line alone, re-probed fresh once
+            // stopPrevious (above) has actually freed the card — never the snapshot `facts` took before
+            // eviction, which is why this is a lifecycle hook and not just more code in this function.
+            beforeCreate: () => this.checkMemoryBeforeCreate(gpu.gpu_id, descriptor, resolved, check, settings),
+            ...(opts.timeoutSecs !== undefined ? { timeoutMs: opts.timeoutSecs * 1000 } : {}),
+            ...(signal !== undefined ? { signal } : {}),
+            ...(substituted !== undefined ? { gpuSubstituted: substituted } : {}),
+        });
+        this.sessionCapabilities.set(modelId, {
+            generation: session.generation ?? '',
+            capabilities: this.engine.adapter.capabilities({ settings, family }),
+            limits: { contextLength: settings.context_length, maxOutputTokens: settings.max_output_tokens },
+        });
+        return session;
+    }
+    /**
+     * Throws `MANAGED_STOP_UNCONFIRMED` when Docker will not confirm the stop: the card stays reserved
+     * and the caller is told, rather than handed a success that is not one.
+     */
+    async unload(modelId) {
+        // Read before the stop: an id this provider never had answers `was_loaded: false` (spec
+        // `tensorrt-llm-runtime`), so a client's wrong id never reads as a stopped model (G-app-4).
+        const wasLoaded = this.residentModels().includes(modelId);
+        await this.current?.unload(modelId);
+        this.sessionCapabilities.delete(modelId);
+        return { success: true, was_loaded: wasLoaded };
+    }
+    /**
+     * Every model holding a card through this provider — loading, ready, stopping, or stopped without
+     * Docker's confirmation: what a removal of the engine must unload first (task 2.6), through core's
+     * own unload so each model's cross-process claim is released with it (final review M-1).
+     */
+    residentModels() {
+        return [...new Set((this.current?.reservations() ?? []).map((r) => r.model_id))];
+    }
+    /**
+     * Refuses every load with `MANAGED_OPERATION_CONFLICT` until the returned release is called (final
+     * review M-1): a removal of the engine holds loads off for as long as it runs, so no load creates
+     * a container on an installation being removed. Holds nest; each release counts once.
+     */
+    holdOffLoads() {
+        this.loadHolds += 1;
+        let released = false;
+        return () => {
+            if (released)
+                return;
+            released = true;
+            this.loadHolds -= 1;
+        };
+    }
+    /**
+     * Refuses loads of this one model with `MANAGED_OPERATION_CONFLICT` until the returned release is
+     * called (task 2.24): a deletion holds them off from before it stops the model until its files are
+     * gone, so no load mounts a folder being removed. Other models load as usual. Holds nest.
+     */
+    holdOffModel(modelId) {
+        this.modelHolds.set(modelId, (this.modelHolds.get(modelId) ?? 0) + 1);
+        let released = false;
+        return () => {
+            if (released)
+                return;
+            released = true;
+            const left = (this.modelHolds.get(modelId) ?? 1) - 1;
+            if (left > 0)
+                this.modelHolds.set(modelId, left);
+            else
+                this.modelHolds.delete(modelId);
+        };
+    }
+    /** Never: the context is fixed when the container starts (design D9, spec "без авто-роста"). */
+    autoIncreaseCtx(_modelId) {
+        return Promise.resolve({ ok: false, reason: 'unsupported' });
+    }
+    recreateSession(modelId) {
+        return Promise.reject(new AtomicCoreError('INVALID_ARGUMENT', `A ${this.engine.provider} session is not recreated in place; unload the model and load it again.`, modelId));
+    }
+    async shutdown() {
+        this.closed = true;
+        await this.current?.shutdown();
+        this.sessionCapabilities.clear();
+    }
+    routePolicy(modelId) {
+        const session = this.findSession(modelId);
+        if (session === undefined)
+            return undefined;
+        const known = this.sessionCapabilities.get(modelId);
+        // A session with no record of what it can do is treated as able to do nothing optional.
+        if (known === undefined || known.generation !== session.generation)
+            return this.engine.routePolicy(NONE);
+        return this.engine.routePolicy(known.capabilities, known.limits);
+    }
+    /** Answers rather than throws, like every provider's capabilities: all false while nothing resolves. */
+    async capabilities(modelId) {
+        const base = {
+            modelId,
+            mmprojExists: false,
+            isEmbedding: false,
+            isDecision: false,
+            audio: false,
+            gemmaMtp: false,
+            dflash: false,
+            dflashDrafts: [],
+            ...NONE,
+            architecture: null,
+        };
+        try {
+            const [ready, model] = await Promise.all([this.deps.readyInstallation(), this.deps.model(modelId)]);
+            const settings = this.engine.settings({});
+            return {
+                ...base,
+                ...this.engine.adapter.capabilities({ settings, family: familyOf(ready, model) }),
+                architecture: model.architecture,
+            };
+        }
+        catch {
+            return base;
+        }
+    }
+    async logs(modelId) {
+        const lifecycle = this.current;
+        const session = lifecycle?.findSession(modelId);
+        if (lifecycle && session) {
+            return {
+                model_id: modelId,
+                source: 'session',
+                generation: session.generation ?? '',
+                log_tail: (await lifecycle.logs(modelId)) ?? '',
+            };
+        }
+        const attempt = lifecycle?.lastAttempt(modelId);
+        if (attempt === undefined)
+            return { model_id: modelId, source: null, log_tail: '' };
+        return {
+            model_id: modelId,
+            source: 'last-attempt',
+            generation: attempt.generation,
+            log_tail: attempt.log_tail,
+            error: attempt.error,
+            at: attempt.at,
+        };
+    }
+    assertOpen() {
+        if (this.closed) {
+            throw new AtomicCoreError('CORE_NOT_RUNNING', `The ${this.engine.provider} provider is shutting down.`);
+        }
+    }
+    assertNotHeldOff(modelId) {
+        if (this.loadHolds > 0) {
+            throw new AtomicCoreError('MANAGED_OPERATION_CONFLICT', `The ${this.engine.label} engine is being removed; load the model once the removal has finished.`, modelId);
+        }
+        if (this.modelHolds.has(modelId)) {
+            throw new AtomicCoreError('MANAGED_OPERATION_CONFLICT', 'The model is being deleted.', modelId);
+        }
+    }
+    /**
+     * The card each container holds, for core's residency rule: every model past its claim — loading,
+     * ready, stopping, or stopped without Docker's confirmation, which still holds the card.
+     */
+    gpuOccupancy() {
+        return (this.current?.reservations() ?? [])
+            .filter((r) => r.state !== 'loading' || this.claimed.has(r.generation))
+            .map((r) => ({
+            model_id: r.model_id,
+            cards: [r.gpu_uuid],
+            auxiliary: false,
+            state: r.state,
+            ...(r.state === 'stop-unconfirmed' && r.container_id !== null
+                ? {
+                    remedy: `Loading again retries the stop; if Docker keeps failing, restart Docker or remove ` +
+                        `container ${r.container_id} (docker rm -f ${r.container_id}).`,
+                }
+                : {}),
+        }));
+    }
+    /**
+     * The `stopping-previous` stage. First, the same model in any other managed provider is unloaded
+     * (`releaseModelElsewhere`). Then, with core's residency: its claim on `gpuId`, which stops every other
+     * model of this provider (one session at a time) and every other engine on that card, each with a
+     * confirmed exit, and refuses with `GPU_BUSY` while one will not stop. Without it: every other
+     * model of this provider, loading or loaded, stopped with confirmation. Either way the load holds its
+     * card from then on — with residency, from the moment core grants the claim, inside core's turn.
+     */
+    async stopOthers(lifecycle, load, signal) {
+        // Generations whose reservation is gone are forgotten here, never while answering a read.
+        const live = new Set(lifecycle.reservations().map((r) => r.generation));
+        for (const generation of this.claimed)
+            if (!live.has(generation))
+                this.claimed.delete(generation);
+        const hold = () => {
+            this.claimed.add(load.generation);
+        };
+        // The same model in another managed provider goes first, before any card is claimed for it.
+        await this.deps.releaseModelElsewhere?.(load.modelId, signal);
+        if (this.deps.claimGpu) {
+            const claim = {
+                model_id: load.modelId,
+                cards: [load.gpuId],
+                auxiliary: false,
+                soleSessionOfProvider: true,
+            };
+            await this.deps.claimGpu(claim, signal, hold);
+        }
+        else {
+            const others = new Set(lifecycle.reservations().map((r) => r.model_id));
+            others.delete(load.modelId);
+            for (const other of others)
+                await this.unload(other);
+        }
+        hold();
+    }
+    /**
+     * The lifecycle's `beforeCreate` hook (task 2.16w round 1, finding 1 (Critical)): the pre-launch
+     * check's memory line, re-probed fresh right before the container is actually created — after
+     * `stopPrevious` has run, so a model switch on a single-GPU host sees the card the previous
+     * session just freed, never a snapshot `load()` took before eviction. `MANAGED_PREREQUISITE_
+     * BLOCKED` when the chosen card is no longer on the host at all (the same code a missing card gets
+     * earlier in `load()`); `MODEL_INCOMPATIBLE` when it is still there but does not have room.
+     */
+    async checkMemoryBeforeCreate(gpuId, descriptor, resolved, check, settings) {
+        const facts = await this.deps.hostFacts();
+        const card = facts.gpus.find((gpu) => gpu.gpu_id === gpuId);
+        if (card === undefined) {
+            throw new AtomicCoreError('MANAGED_PREREQUISITE_BLOCKED', 'The selected GPU disappeared before the container could start.', gpuId);
+        }
+        const verdict = checkCheckpointMemory(resolved, descriptor, facts.gpus, facts.memory, check);
+        if (!verdict.verdict.ok) {
+            const { code, message, details } = verdict.verdict.error;
+            throw new AtomicCoreError(code, message, details);
+        }
+        // The engine's launch plan from this same fresh look at the card (design D9 of add-vllm-runtime).
+        return this.engine.launchPlan?.(settings, {
+            weightBytesTotal: resolved.weightBytesTotal,
+            configJson: resolved.configJson,
+            hfQuantConfigJson: resolved.hfQuantConfigJson,
+        }, card, facts.memory);
+    }
+}
+//# sourceMappingURL=runtime.js.map

package/dist/runtime/managed-engines/runtime.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"runtime.js","sourceRoot":"","sources":["../../../src/runtime/managed-engines/runtime.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAkCG;AACH,OAAO,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAA;AAa1D,OAAO,EAAE,oBAAoB,EAAE,MAAM,oBAAoB,CAAA;AAUzD,OAAO,EAAE,qBAAqB,EAAE,eAAe,EAAE,MAAM,4BAA4B,CAAA;AAKnF,OAAO,EAAE,gCAAgC,EAAE,MAAM,8BAA8B,CAAA;AAyD/E,MAAM,IAAI,GAA4B;IACpC,KAAK,EAAE,KAAK;IACZ,SAAS,EAAE,KAAK;IAChB,iBAAiB,EAAE,KAAK;IACxB,MAAM,EAAE,KAAK;IACb,UAAU,EAAE,KAAK;IACjB,SAAS,EAAE,KAAK;CACjB,CAAA;AAED,uGAAuG;AACvG,SAAS,QAAQ,CAAC,KAAwB,EAAE,KAAmB;IAC7D,IAAI,KAAK,CAAC,YAAY,KAAK,IAAI;QAAE,OAAO,IAAI,CAAA;IAC5C,OAAO,KAAK,CAAC,UAAU,CAAC,cAAc,CAAC,KAAK,CAAC,YAAY,CAAC,IAAI,IAAI,CAAA;AACpE,CAAC;AAED;;;;;;GAMG;AACH,SAAS,kBAAkB,CACzB,KAAwB,EACxB,QAA4B;IAE5B,MAAM,YAAY,GAAG,QAAQ,CAAC,aAAa,CAAC,CAAC,CAAC,CAAA;IAC9C,OAAO,YAAY,KAAK,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,UAAU,CAAC,cAAc,CAAC,YAAY,CAAC,IAAI,IAAI,CAAC,CAAA;AACpG,CAAC;AAED,MAAM,OAAO,kBAAkB;IAsBlB;IACQ;IApBX,OAAO,GAAgC,IAAI,CAAA;IAC3C,MAAM,GAAG,KAAK,CAAA;IACtB,kGAAkG;IAC1F,SAAS,GAAG,CAAC,CAAA;IACrB,sGAAsG;IACrF,UAAU,GAAG,IAAI,GAAG,EAAkB,CAAA;IACvD,sFAAsF;IACrE,OAAO,GAAG,IAAI,GAAG,EAAU,CAAA;IAC5C,wGAAwG;IACvF,mBAAmB,GAAG,IAAI,GAAG,EAO3C,CAAA;IAEH,YACW,MAA4B,EACpB,IAA4B;QADpC,WAAM,GAAN,MAAM,CAAsB;QACpB,SAAI,GAAJ,IAAI,CAAwB;IAC5C,CAAC;IAEJ,IAAI;QACF,OAAO,IAAI,CAAC,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,CAAA;IACnC,CAAC;IAED,WAAW,CAAC,OAAe;QACzB,OAAO,IAAI,CAAC,OAAO,EAAE,WAAW,CAAC,OAAO,CAAC,CAAA;IAC3C,CAAC;IAED,eAAe;QACb,OAAO,IAAI,CAAC,IAAI,EAAE,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,QAAQ,CAAC,CAAA;IACvD,CAAC;IAED,SAAS,CAAC,OAAe;QACvB,OAAO,IAAI,CAAC,OAAO,EAAE,SAAS,CAAC,OAAO,CAAC,IAAI,KAAK,CAAA;IAClD,CAAC;IAED,KAAK,CAAC,IAAI,CAAC,OAAe,EAAE,OAAyB,EAAE;QACrD,MAAM,EAAE,MAAM,EAAE,GAAG,IAAI,CAAA;QACvB,IAAI,CAAC,UAAU,EAAE,CAAA;QACjB,IAAI,CAAC,gBAAgB,CAAC,OAAO,CAAC,CAAA;QAC9B,oBAAoB,CAAC,MAAM,CAAC,CAAA;QAC5B,IAAI,IAAI,CAAC,WAAW,EAAE,CAAC;YACrB,MAAM,IAAI,eAAe,CACvB,kBAAkB,EAClB,GAAG,IAAI,CAAC,MAAM,CAAC,QAAQ,kCAAkC,EACzD,OAAO,CACR,CAAA;QACH,CAAC;QACD,mGAAmG;QACnG,MAAM,QAAQ,GAAG,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,QAAQ,EAAE,EAAE,IAAI,CAAC,SAAS,CAAC,CAAA;QAC3E,MAAM,SAAS,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,CAAC,KAAK,CAAC,CAAC,KAAc,EAAE,EAAE;YACrE,MAAM,IAAI,eAAe,CACvB,6BAA6B,EAC7B,0DAA0D,IAAI,CAAC,MAAM,CAAC,QAAQ,qBAAqB,EACnG,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CACvD,CAAA;QACH,CAAC,CAAC,CAAA;QACF,IAAI,SAAS,KAAK,IAAI,EAAE,CAAC;YACvB,MAAM,IAAI,eAAe,CACvB,6BAA6B,EAC7B,+CAA+C,IAAI,CAAC,MAAM,CAAC,QAAQ,qBAAqB,CACzF,CAAA;QACH,CAAC;QACD,8FAA8F;QAC9F,oEAAoE;QACpE,IAAI,CAAC,OAAO,GAAG,SAAS,CAAA;QACxB,MAAM,KAAK,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,iBAAiB,EAAE,CAAA;QACjD,oBAAoB,CAAC,MAAM,CAAC,CAAA;QAC5B,MAAM,KAAK,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAA;QAC5C,MAAM,KAAK,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,CAAA;QACzC,oBAAoB,CAAC,MAAM,CAAC,CAAA;QAC5B,IAAI,CAAC,UAAU,EAAE,CAAA;QAEjB,MAAM,GAAG,GAAG,eAAe,CAAC,KAAK,CAAC,IAAI,EAAE,KAAK,CAAC,MAAM,EAAE,QAAQ,CAAC,MAAM,IAAI,SAAS,CAAC,CAAA;QACnF,IAAI,GAAG,KAAK,IAAI,EAAE,CAAC;YACjB,MAAM,IAAI,eAAe,CACvB,8BAA8B,EAC9B,+CAA+C,IAAI,CAAC,MAAM,CAAC,QAAQ,uBAAuB,CAC3F,CAAA;QACH,CAAC;QACD,MAAM,WAAW,GACf,QAAQ,CAAC,MAAM,KAAK,IAAI,IAAI,GAAG,CAAC,MAAM,KAAK,QAAQ,CAAC,MAAM;YACxD,CAAC,CAAC,EAAE,gBAAgB,EAAE,QAAQ,CAAC,MAAM,EAAE,MAAM,EAAE,GAAG,CAAC,MAAM,EAAE;YAC3D,CAAC,CAAC,SAAS,CAAA;QACf,MAAM,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAA;QAEzC,gGAAgG;QAChG,6FAA6F;QAC7F,gGAAgG;QAChG,iGAAiG;QACjG,iGAAiG;QACjG,0BAA0B;QAC1B,MAAM,QAAQ,GAAG,MAAM,gCAAgC,CACrD,KAAK,EACL,KAAK,CAAC,UAAU,EAChB,KAAK,CAAC,IAAI,EACV,KAAK,CAAC,MAAM,EACZ,EAAE,KAAK,EAAE,GAAG,CAAC,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,CACrC,CAAA;QACD,oBAAoB,CAAC,MAAM,CAAC,CAAA;QAC5B,IAAI,CAAC,UAAU,EAAE,CAAA;QAEjB,0FAA0F;QAC1F,MAAM,MAAM,GAAG,kBAAkB,CAAC,KAAK,EAAE,QAAQ,CAAC,CAAA;QAElD,MAAM,EAAE,UAAU,EAAE,GAAG,KAAK,CAAA;QAC5B,iGAAiG;QACjG,6FAA6F;QAC7F,IAAI,CAAC,gBAAgB,CAAC,OAAO,CAAC,CAAA;QAC9B,MAAM,OAAO,GAAG,MAAM,SAAS,CAAC,IAAI,CAAC;YACnC,OAAO;YACP,SAAS,EAAE,KAAK,CAAC,GAAG;YACpB,WAAW,EAAE,KAAK,CAAC,WAAW;YAC9B,YAAY,EAAE;gBACZ,aAAa,EAAE,UAAU,CAAC,aAAa;gBACvC,SAAS,EAAE,UAAU,CAAC,SAAS;gBAC/B,UAAU,EAAE,UAAU,CAAC,UAAU;gBACjC,wBAAwB,EAAE,UAAU,CAAC,wBAAwB;gBAC7D,KAAK,EAAE,KAAK,CAAC,KAAK;aACnB;YACD,MAAM;YACN,OAAO,EAAE,GAAG,CAAC,MAAM;YACnB,2FAA2F;YAC3F,yDAAyD;YACzD,aAAa,EAAE,GAAG,CAAC,gBAAgB,KAAK,IAAI;YAC5C,iBAAiB,EAAE,GAAG,CAAC,gBAAgB;YACvC,oBAAoB,EAAE,GAAG,CAAC,kBAAkB;YAC5C,OAAO,EAAE,KAAK,CAAC,OAAO;YACtB,QAAQ;YACR,6FAA6F;YAC7F,gFAAgF;YAChF,YAAY,EAAE,CAAC,WAAW,EAAE,UAAU,EAAE,EAAE,CACxC,IAAI,CAAC,UAAU,CAAC,SAAS,EAAE,EAAE,OAAO,EAAE,UAAU,EAAE,KAAK,EAAE,GAAG,CAAC,MAAM,EAAE,EAAE,WAAW,CAAC;YACrF,qFAAqF;YACrF,4FAA4F;YAC5F,2FAA2F;YAC3F,YAAY,EAAE,GAAG,EAAE,CAAC,IAAI,CAAC,uBAAuB,CAAC,GAAG,CAAC,MAAM,EAAE,UAAU,EAAE,QAAQ,EAAE,KAAK,EAAE,QAAQ,CAAC;YACnG,GAAG,CAAC,IAAI,CAAC,WAAW,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,SAAS,EAAE,IAAI,CAAC,WAAW,GAAG,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YACjF,GAAG,CAAC,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YAC3C,GAAG,CAAC,WAAW,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,cAAc,EAAE,WAAW,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;SACtE,CAAC,CAAA;QACF,IAAI,CAAC,mBAAmB,CAAC,GAAG,CAAC,OAAO,EAAE;YACpC,UAAU,EAAE,OAAO,CAAC,UAAU,IAAI,EAAE;YACpC,YAAY,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,QAAQ,EAAE,MAAM,EAAE,CAAC;YACpE,MAAM,EAAE,EAAE,aAAa,EAAE,QAAQ,CAAC,cAAc,EAAE,eAAe,EAAE,QAAQ,CAAC,iBAAiB,EAAE;SAChG,CAAC,CAAA;QACF,OAAO,OAAO,CAAA;IAChB,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,MAAM,CAAC,OAAe;QAC1B,wFAAwF;QACxF,4FAA4F;QAC5F,MAAM,SAAS,GAAG,IAAI,CAAC,cAAc,EAAE,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAA;QACzD,MAAM,IAAI,CAAC,OAAO,EAAE,MAAM,CAAC,OAAO,CAAC,CAAA;QACnC,IAAI,CAAC,mBAAmB,CAAC,MAAM,CAAC,OAAO,CAAC,CAAA;QACxC,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,UAAU,EAAE,SAAS,EAAE,CAAA;IACjD,CAAC;IAED;;;;OAIG;IACH,cAAc;QACZ,OAAO,CAAC,GAAG,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC,OAAO,EAAE,YAAY,EAAE,IAAI,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAA;IAClF,CAAC;IAED;;;;OAIG;IACH,YAAY;QACV,IAAI,CAAC,SAAS,IAAI,CAAC,CAAA;QACnB,IAAI,QAAQ,GAAG,KAAK,CAAA;QACpB,OAAO,GAAG,EAAE;YACV,IAAI,QAAQ;gBAAE,OAAM;YACpB,QAAQ,GAAG,IAAI,CAAA;YACf,IAAI,CAAC,SAAS,IAAI,CAAC,CAAA;QACrB,CAAC,CAAA;IACH,CAAC;IAED;;;;OAIG;IACH,YAAY,CAAC,OAAe;QAC1B,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,CAAA;QACrE,IAAI,QAAQ,GAAG,KAAK,CAAA;QACpB,OAAO,GAAG,EAAE;YACV,IAAI,QAAQ;gBAAE,OAAM;YACpB,QAAQ,GAAG,IAAI,CAAA;YACf,MAAM,IAAI,GAAG,CAAC,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAA;YACpD,IAAI,IAAI,GAAG,CAAC;gBAAE,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,OAAO,EAAE,IAAI,CAAC,CAAA;;gBAC3C,IAAI,CAAC,UAAU,CAAC,MAAM,CAAC,OAAO,CAAC,CAAA;QACtC,CAAC,CAAA;IACH,CAAC;IAED,gGAAgG;IAChG,eAAe,CAAC,QAAgB;QAC9B,OAAO,OAAO,CAAC,OAAO,CAAC,EAAE,EAAE,EAAE,KAAK,EAAE,MAAM,EAAE,aAAa,EAAE,CAAC,CAAA;IAC9D,CAAC;IAED,eAAe,CAAC,OAAe;QAC7B,OAAO,OAAO,CAAC,MAAM,CACnB,IAAI,eAAe,CACjB,kBAAkB,EAClB,KAAK,IAAI,CAAC,MAAM,CAAC,QAAQ,yEAAyE,EAClG,OAAO,CACR,CACF,CAAA;IACH,CAAC;IAED,KAAK,CAAC,QAAQ;QACZ,IAAI,CAAC,MAAM,GAAG,IAAI,CAAA;QAClB,MAAM,IAAI,CAAC,OAAO,EAAE,QAAQ,EAAE,CAAA;QAC9B,IAAI,CAAC,mBAAmB,CAAC,KAAK,EAAE,CAAA;IAClC,CAAC;IAED,WAAW,CAAC,OAAe;QACzB,MAAM,OAAO,GAAG,IAAI,CAAC,WAAW,CAAC,OAAO,CAAC,CAAA;QACzC,IAAI,OAAO,KAAK,SAAS;YAAE,OAAO,SAAS,CAAA;QAC3C,MAAM,KAAK,GAAG,IAAI,CAAC,mBAAmB,CAAC,GAAG,CAAC,OAAO,CAAC,CAAA;QACnD,wFAAwF;QACxF,IAAI,KAAK,KAAK,SAAS,IAAI,KAAK,CAAC,UAAU,KAAK,OAAO,CAAC,UAAU;YAAE,OAAO,IAAI,CAAC,MAAM,CAAC,WAAW,CAAC,IAAI,CAAC,CAAA;QACxG,OAAO,IAAI,CAAC,MAAM,CAAC,WAAW,CAAC,KAAK,CAAC,YAAY,EAAE,KAAK,CAAC,MAAM,CAAC,CAAA;IAClE,CAAC;IAED,wGAAwG;IACxG,KAAK,CAAC,YAAY,CAAC,OAAe;QAChC,MAAM,IAAI,GAA6B;YACrC,OAAO;YACP,YAAY,EAAE,KAAK;YACnB,WAAW,EAAE,KAAK;YAClB,UAAU,EAAE,KAAK;YACjB,KAAK,EAAE,KAAK;YACZ,QAAQ,EAAE,KAAK;YACf,MAAM,EAAE,KAAK;YACb,YAAY,EAAE,EAAE;YAChB,GAAG,IAAI;YACP,YAAY,EAAE,IAAI;SACnB,CAAA;QACD,IAAI,CAAC;YACH,MAAM,CAAC,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,iBAAiB,EAAE,EAAE,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAA;YACnG,MAAM,QAAQ,GAAG,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAA;YACzC,OAAO;gBACL,GAAG,IAAI;gBACP,GAAG,IAAI,CAAC,MAAM,CAAC,OAAO,CAAC,YAAY,CAAC,EAAE,QAAQ,EAAE,MAAM,EAAE,QAAQ,CAAC,KAAK,EAAE,KAAK,CAAC,EAAE,CAAC;gBACjF,YAAY,EAAE,KAAK,CAAC,YAAY;aACjC,CAAA;QACH,CAAC;QAAC,MAAM,CAAC;YACP,OAAO,IAAI,CAAA;QACb,CAAC;IACH,CAAC;IAED,KAAK,CAAC,IAAI,CAAC,OAAe;QACxB,MAAM,SAAS,GAAG,IAAI,CAAC,OAAO,CAAA;QAC9B,MAAM,OAAO,GAAG,SAAS,EAAE,WAAW,CAAC,OAAO,CAAC,CAAA;QAC/C,IAAI,SAAS,IAAI,OAAO,EAAE,CAAC;YACzB,OAAO;gBACL,QAAQ,EAAE,OAAO;gBACjB,MAAM,EAAE,SAAS;gBACjB,UAAU,EAAE,OAAO,CAAC,UAAU,IAAI,EAAE;gBACpC,QAAQ,EAAE,CAAC,MAAM,SAAS,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE;aAChD,CAAA;QACH,CAAC;QACD,MAAM,OAAO,GAAG,SAAS,EAAE,WAAW,CAAC,OAAO,CAAC,CAAA;QAC/C,IAAI,OAAO,KAAK,SAAS;YAAE,OAAO,EAAE,QAAQ,EAAE,OAAO,EAAE,MAAM,EAAE,IAAI,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAA;QACnF,OAAO;YACL,QAAQ,EAAE,OAAO;YACjB,MAAM,EAAE,cAAc;YACtB,UAAU,EAAE,OAAO,CAAC,UAAU;YAC9B,QAAQ,EAAE,OAAO,CAAC,QAAQ;YAC1B,KAAK,EAAE,OAAO,CAAC,KAAK;YACpB,EAAE,EAAE,OAAO,CAAC,EAAE;SACf,CAAA;IACH,CAAC;IAEO,UAAU;QAChB,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC;YAChB,MAAM,IAAI,eAAe,CAAC,kBAAkB,EAAE,OAAO,IAAI,CAAC,MAAM,CAAC,QAAQ,6BAA6B,CAAC,CAAA;QACzG,CAAC;IACH,CAAC;IAEO,gBAAgB,CAAC,OAAe;QACtC,IAAI,IAAI,CAAC,SAAS,GAAG,CAAC,EAAE,CAAC;YACvB,MAAM,IAAI,eAAe,CACvB,4BAA4B,EAC5B,OAAO,IAAI,CAAC,MAAM,CAAC,KAAK,yEAAyE,EACjG,OAAO,CACR,CAAA;QACH,CAAC;QACD,IAAI,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,CAAC;YACjC,MAAM,IAAI,eAAe,CAAC,4BAA4B,EAAE,6BAA6B,EAAE,OAAO,CAAC,CAAA;QACjG,CAAC;IACH,CAAC;IAED;;;OAGG;IACH,YAAY;QACV,OAAO,CAAC,IAAI,CAAC,OAAO,EAAE,YAAY,EAAE,IAAI,EAAE,CAAC;aACxC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,KAAK,SAAS,IAAI,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC;aACtE,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;YACX,QAAQ,EAAE,CAAC,CAAC,QAAQ;YACpB,KAAK,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAC;YACnB,SAAS,EAAE,KAAK;YAChB,KAAK,EAAE,CAAC,CAAC,KAAK;YACd,GAAG,CAAC,CAAC,CAAC,KAAK,KAAK,kBAAkB,IAAI,CAAC,CAAC,YAAY,KAAK,IAAI;gBAC3D,CAAC,CAAC;oBACE,MAAM,EACJ,oFAAoF;wBACpF,aAAa,CAAC,CAAC,YAAY,kBAAkB,CAAC,CAAC,YAAY,IAAI;iBAClE;gBACH,CAAC,CAAC,EAAE,CAAC;SACR,CAAC,CAAC,CAAA;IACP,CAAC;IAED;;;;;;;OAOG;IACK,KAAK,CAAC,UAAU,CACtB,SAA+B,EAC/B,IAA4D,EAC5D,MAAmB;QAEnB,0FAA0F;QAC1F,MAAM,IAAI,GAAG,IAAI,GAAG,CAAC,SAAS,CAAC,YAAY,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAA;QACvE,KAAK,MAAM,UAAU,IAAI,IAAI,CAAC,OAAO;YAAE,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,UAAU,CAAC;gBAAE,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,UAAU,CAAC,CAAA;QACjG,MAAM,IAAI,GAAG,GAAG,EAAE;YAChB,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,IAAI,CAAC,UAAU,CAAC,CAAA;QACnC,CAAC,CAAA;QACD,4FAA4F;QAC5F,MAAM,IAAI,CAAC,IAAI,CAAC,qBAAqB,EAAE,CAAC,IAAI,CAAC,OAAO,EAAE,MAAM,CAAC,CAAA;QAC7D,IAAI,IAAI,CAAC,IAAI,CAAC,QAAQ,EAAE,CAAC;YACvB,MAAM,KAAK,GAAG;gBACZ,QAAQ,EAAE,IAAI,CAAC,OAAO;gBACtB,KAAK,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC;gBACnB,SAAS,EAAE,KAAK;gBAChB,qBAAqB,EAAE,IAAI;aAC5B,CAAA;YACD,MAAM,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,KAAK,EAAE,MAAM,EAAE,IAAI,CAAC,CAAA;QAC/C,CAAC;aAAM,CAAC;YACN,MAAM,MAAM,GAAG,IAAI,GAAG,CAAC,SAAS,CAAC,YAAY,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAA;YACvE,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,OAAO,CAAC,CAAA;YAC3B,KAAK,MAAM,KAAK,IAAI,MAAM;gBAAE,MAAM,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAA;QACtD,CAAC;QACD,IAAI,EAAE,CAAA;IACR,CAAC;IAED;;;;;;;OAOG;IACK,KAAK,CAAC,uBAAuB,CACnC,KAAa,EACb,UAA6B,EAC7B,QAA4B,EAC5B,KAAyB,EACzB,QAAW;QAEX,MAAM,KAAK,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,CAAA;QACzC,MAAM,IAAI,GAAG,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,MAAM,KAAK,KAAK,CAAC,CAAA;QAC3D,IAAI,IAAI,KAAK,SAAS,EAAE,CAAC;YACvB,MAAM,IAAI,eAAe,CACvB,8BAA8B,EAC9B,gEAAgE,EAChE,KAAK,CACN,CAAA;QACH,CAAC;QACD,MAAM,OAAO,GAAG,qBAAqB,CAAC,QAAQ,EAAE,UAAU,EAAE,KAAK,CAAC,IAAI,EAAE,KAAK,CAAC,MAAM,EAAE,KAAK,CAAC,CAAA;QAC5F,IAAI,CAAC,OAAO,CAAC,OAAO,CAAC,EAAE,EAAE,CAAC;YACxB,MAAM,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,GAAG,OAAO,CAAC,OAAO,CAAC,KAAK,CAAA;YACxD,MAAM,IAAI,eAAe,CAAC,IAAI,EAAE,OAAO,EAAE,OAAO,CAAC,CAAA;QACnD,CAAC;QACD,kGAAkG;QAClG,OAAO,IAAI,CAAC,MAAM,CAAC,UAAU,EAAE,CAC7B,QAAQ,EACR;YACE,gBAAgB,EAAE,QAAQ,CAAC,gBAAgB;YAC3C,UAAU,EAAE,QAAQ,CAAC,UAAU;YAC/B,iBAAiB,EAAE,QAAQ,CAAC,iBAAiB;SAC9C,EACD,IAAI,EACJ,KAAK,CAAC,MAAM,CACb,CAAA;IACH,CAAC;CACF"}

package/dist/runtime/managed-engines/spec.d.ts

@@ -0,0 +1,70 @@
+/**
+ * The managed engine registry (change `add-vllm-runtime`, design D3): what one managed engine is to
+ * core, beyond the `ManagedTextAdapter` the container lifecycle runs. A second engine is a new spec,
+ * an adapter and a descriptor — never a branch on an engine id in shared code.
+ *
+ * A spec says where the engine's descriptor comes from, which adapter launches its containers, how
+ * its stored settings are read and validated, how the shared compatibility check sizes its memory and
+ * which checkpoints its loader cannot read (`check`, design D7), and how the public server routes its
+ * sessions. The memory rule and the checkpoint quirks belong to the spec, not to the adapter:
+ * `adapter_contract_version` is a contract between a descriptor and a container launch, and the check
+ * runs with no Docker and no lifecycle at all.
+ *
+ * `engine_id` is the provider id. Core has relied on that in four places (residency leftovers, the
+ * models of an engine, the removal's unloader, the descriptor of a load); the registry makes it an
+ * invariant, checked when core starts.
+ */
+import type { SessionRoutePolicy } from '../shared/index.js';
+import type { DescriptorSource } from '../environment/index.js';
+import type { ManagedTextAdapter, ManagedTextCapabilities } from '../managed-text/index.js';
+import type { GpuFacts } from '../../contracts/index.js';
+import type { CheckpointMemoryInputs, HostMemory, ManagedCheckEngine, ManagedModelCheckEngine } from '../managed-models/index.js';
+/** What every managed engine's validated settings carry, whatever else they hold. */
+export interface ManagedEngineSettings {
+    /** `GPU-<uuid>`/`MIG-<uuid>`, or `null` to let the load pick the card with the most free memory. */
+    gpu_id: string | null;
+    context_length: number;
+    max_output_tokens: number;
+}
+export interface ManagedEngineSpec<S extends ManagedEngineSettings = ManagedEngineSettings> {
+    /** The engine's id in conf and in installations. */
+    engine_id: string;
+    /** The provider id it is registered under: always `engine_id` (the registry checks). */
+    provider: string;
+    /** "TensorRT-LLM", "vLLM": the engine in messages. */
+    label: string;
+    /** conf main's `runtimes/<engine_id>.json`. */
+    descriptor: DescriptorSource;
+    adapter: ManagedTextAdapter<S>;
+    /**
+     * Stored values overlaid by a load's own overrides, reduced to this provider's keys and validated:
+     * `INVALID_ARGUMENT` for anything the schema refuses, before a container exists.
+     */
+    settings(stored: Record<string, unknown>, overrides?: Record<string, unknown>): S;
+    /** The compatibility check's hooks (memory rule, checkpoint quirks), sized by these settings. */
+    check(settings: S): ManagedCheckEngine;
+    /**
+     * Optional: the engine's launch plan, made right before its container is created — after the
+     * previous sessions stopped and the memory check passed on the card as it stands then — and handed
+     * to its adapter as `ManagedLaunchContext.plan` (vLLM's KV cache bytes and memory share, design D9).
+     */
+    launchPlan?(settings: S, checkpoint: CheckpointMemoryInputs, gpu: GpuFacts, host: HostMemory): unknown;
+    /** What the public server must know about a session to route it honestly. */
+    routePolicy(capabilities: ManagedTextCapabilities | null, limits?: {
+        contextLength: number;
+        maxOutputTokens: number;
+    } | null): SessionRoutePolicy;
+}
+/** The check route's side of an engine: its saved card and its memory rule, from stored settings. */
+export declare function managedModelCheckOf<S extends ManagedEngineSettings>(spec: ManagedEngineSpec<S>): ManagedModelCheckEngine;
+/** The engines this core can run, in priority order of registration. */
+export declare class ManagedEngineRegistry {
+    private readonly specs;
+    /** Throws when the spec breaks the registry's invariants: a programming error, found at startup. */
+    register<S extends ManagedEngineSettings>(spec: ManagedEngineSpec<S>): void;
+    get(engineId: string): ManagedEngineSpec | undefined;
+    has(engineId: string): boolean;
+    list(): ManagedEngineSpec[];
+    /** Every engine's descriptor source, for the descriptor provider. */
+    descriptorSources(): DescriptorSource[];
+}

package/dist/runtime/managed-engines/spec.js

@@ -0,0 +1,39 @@
+/** The check route's side of an engine: its saved card and its memory rule, from stored settings. */
+export function managedModelCheckOf(spec) {
+    return {
+        engineId: spec.engine_id,
+        gpuIdOf: (settings) => spec.settings(settings).gpu_id ?? null,
+        checkEngineOf: (settings) => spec.check(spec.settings(settings)),
+    };
+}
+/** The engines this core can run, in priority order of registration. */
+export class ManagedEngineRegistry {
+    specs = new Map();
+    /** Throws when the spec breaks the registry's invariants: a programming error, found at startup. */
+    register(spec) {
+        if (spec.engine_id !== spec.provider) {
+            throw new Error(`Managed engine ${spec.label}: engine_id "${spec.engine_id}" must equal its provider id "${spec.provider}".`);
+        }
+        if (spec.descriptor.engine_id !== spec.engine_id) {
+            throw new Error(`Managed engine ${spec.engine_id}: its descriptor source is for "${spec.descriptor.engine_id}".`);
+        }
+        if (this.specs.has(spec.engine_id)) {
+            throw new Error(`Managed engine ${spec.engine_id} is already registered.`);
+        }
+        this.specs.set(spec.engine_id, spec);
+    }
+    get(engineId) {
+        return this.specs.get(engineId);
+    }
+    has(engineId) {
+        return this.specs.has(engineId);
+    }
+    list() {
+        return [...this.specs.values()];
+    }
+    /** Every engine's descriptor source, for the descriptor provider. */
+    descriptorSources() {
+        return this.list().map((spec) => spec.descriptor);
+    }
+}
+//# sourceMappingURL=spec.js.map

package/dist/runtime/managed-engines/spec.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"spec.js","sourceRoot":"","sources":["../../../src/runtime/managed-engines/spec.ts"],"names":[],"mappings":"AAiEA,qGAAqG;AACrG,MAAM,UAAU,mBAAmB,CACjC,IAA0B;IAE1B,OAAO;QACL,QAAQ,EAAE,IAAI,CAAC,SAAS;QACxB,OAAO,EAAE,CAAC,QAAQ,EAAE,EAAE,CAAC,IAAI,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC,MAAM,IAAI,IAAI;QAC7D,aAAa,EAAE,CAAC,QAAQ,EAAE,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC;KACjE,CAAA;AACH,CAAC;AAED,wEAAwE;AACxE,MAAM,OAAO,qBAAqB;IACf,KAAK,GAAG,IAAI,GAAG,EAA6B,CAAA;IAE7D,oGAAoG;IACpG,QAAQ,CAAkC,IAA0B;QAClE,IAAI,IAAI,CAAC,SAAS,KAAK,IAAI,CAAC,QAAQ,EAAE,CAAC;YACrC,MAAM,IAAI,KAAK,CACb,kBAAkB,IAAI,CAAC,KAAK,gBAAgB,IAAI,CAAC,SAAS,iCAAiC,IAAI,CAAC,QAAQ,IAAI,CAC7G,CAAA;QACH,CAAC;QACD,IAAI,IAAI,CAAC,UAAU,CAAC,SAAS,KAAK,IAAI,CAAC,SAAS,EAAE,CAAC;YACjD,MAAM,IAAI,KAAK,CACb,kBAAkB,IAAI,CAAC,SAAS,mCAAmC,IAAI,CAAC,UAAU,CAAC,SAAS,IAAI,CACjG,CAAA;QACH,CAAC;QACD,IAAI,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,SAAS,CAAC,EAAE,CAAC;YACnC,MAAM,IAAI,KAAK,CAAC,kBAAkB,IAAI,CAAC,SAAS,yBAAyB,CAAC,CAAA;QAC5E,CAAC;QACD,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,SAAS,EAAE,IAAoC,CAAC,CAAA;IACtE,CAAC;IAED,GAAG,CAAC,QAAgB;QAClB,OAAO,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAA;IACjC,CAAC;IAED,GAAG,CAAC,QAAgB;QAClB,OAAO,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAA;IACjC,CAAC;IAED,IAAI;QACF,OAAO,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC,CAAA;IACjC,CAAC;IAED,qEAAqE;IACrE,iBAAiB;QACf,OAAO,IAAI,CAAC,IAAI,EAAE,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,UAAU,CAAC,CAAA;IACnD,CAAC;CACF"}

package/dist/runtime/managed-models/check.d.ts

@@ -0,0 +1,50 @@
+import type { GpuFacts, ModelCompatibility, RuntimeDescriptor } from '../../contracts/index.js';
+import type { InstallationStore, RuntimeDescriptorProvider } from '../environment/index.js';
+import type { HostMemory, ManagedCheckEngine, ModelCheckInput } from './compatibility.js';
+export interface ModelCheckHostFacts {
+    gpus: GpuFacts[];
+    memory: HostMemory;
+}
+export interface ModelCheckDeps {
+    /** The setup operation's installation records; a torn or foreign file is skipped there. */
+    installations: Pick<InstallationStore, 'list'>;
+    descriptors: Pick<RuntimeDescriptorProvider, 'forInstallation' | 'cachedForNewSetup'>;
+    /** Never asks Docker anything (see the file banner); a card can disappear between two checks. */
+    hostFacts: () => Promise<ModelCheckHostFacts>;
+    /** The provider's stored settings (`settings.get(<provider>)`), for its card and its memory rule. */
+    settings: () => Record<string, unknown>;
+    /**
+     * Windows only (change `add-tensorrt-llm-windows`, design D11): `hostFacts.memory` is then the WSL
+     * VM's, and this says what `.wslconfig` sets it to, for the warning when the weights do not fit in it.
+     */
+    wslVm?: () => Promise<{
+        memory_setting: string | null;
+    }>;
+}
+/** `INVALID_ARGUMENT` for anything that does not shape up as `ModelCheckInput`; every field the spec names. */
+export declare function parseModelCheckInput(body: unknown): ModelCheckInput;
+/**
+ * The descriptor to check against: the `ready` installation of `engineId` with its pinned descriptor
+ * when there is one, else that engine's latest accepted cached descriptor — spec "по закреплённому
+ * дескриптору установки этого движка (или по актуальному дескриптору этого движка, если движок не
+ * установлен)". Throws the descriptor provider's own `MANAGED_METADATA_INVALID` when neither is
+ * available (nothing has ever been cached for this engine).
+ */
+export declare function descriptorForCheck(engineId: string, deps: Pick<ModelCheckDeps, 'installations' | 'descriptors'>): Promise<RuntimeDescriptor>;
+/** What one engine brings to the check route: its id, and from its stored settings its card and its rule. */
+export interface ManagedModelCheckEngine {
+    engineId: string;
+    /** The card saved in the provider's settings, if any. */
+    gpuIdOf(settings: Record<string, unknown>): string | null;
+    /** The skeleton's hooks, sized by the provider's settings. */
+    checkEngineOf(settings: Record<string, unknown>): ManagedCheckEngine;
+}
+/**
+ * The full route for one engine: validate, resolve the descriptor and host facts (in parallel), then
+ * the pure check. Without an explicit `gpu_id` in the request, checks against the card a real load
+ * would pick — the provider's own stored `gpu_id` setting, falling back to `selectLaunchGpu`'s "most
+ * free memory, then most total memory" rule only when nothing is saved or the saved card is gone
+ * (design D12b of change `add-tensorrt-llm-linux`) — so the verdict and the launch cannot be about
+ * different cards.
+ */
+export declare function checkManagedModel(engine: ManagedModelCheckEngine, body: unknown, deps: ModelCheckDeps): Promise<ModelCompatibility>;