@atomic-chat/core 0.5.1 → 0.13.2

package/dist/runtime/tensorrt-llm/index.js

@@ -0,0 +1,31 @@
+/**
+ * The `tensorrt-llm` provider (openspec change `add-tensorrt-llm-linux`).
+ *
+ * `compatibility.ts` is TensorRT-LLM's memory rule and checkpoint quirks plugged into the shared
+ * check skeleton (`../managed-models/`, change `add-vllm-runtime`), behind
+ * `POST /atomic/v1/models/tensorrt-llm/check` (spec `tensorrt-llm-models`); the shared naming rule
+ * and the GPU-selection rule the load path reuses are re-exported from there. `adapter.ts` (task 2.13) is the `ManagedTextAdapter`
+ * the engine-neutral load lifecycle (`../managed-text/`) plugs into. Task 2.14 adds the provider
+ * itself: `runtime.ts` (the `LocalRuntime`), `settings.ts` (stored settings → adapter settings),
+ * `route-policy.ts` (what the public server must refuse or map for a session), `installation.ts`
+ * (the ready installation and its pinned descriptor), `host-facts.ts` (cards, SELinux and
+ * `MemAvailable`/`MemTotal`, per load) and `model-dir.ts` (the single-model lookup `runtime.ts`'s
+ * load path uses). Task 2.16 wires the pure check in: `registry.ts` (the full model listing scan),
+ * `check.ts` (the route handler behind `compatibility.ts`'s pure verdict) and `prelaunch.ts` (the
+ * file-and-compatibility re-check `runtime.ts`'s load path runs before a container is created).
+ * Task 2.24 adds `delete.ts`: a deleted model's engine caches and folder, with the bytes freed.
+ *
+ * Public API of this module is exported from this file only.
+ */
+export * from './compatibility.js';
+export * from '../managed-models/quant-format.js';
+export * from './adapter.js';
+export * from './kv-cache.js';
+export * from './settings.js';
+export * from './route-policy.js';
+export * from './installation.js';
+export * from './host-facts.js';
+export * from './check.js';
+export * from './prelaunch.js';
+export * from './engine.js';
+//# sourceMappingURL=index.js.map

package/dist/runtime/tensorrt-llm/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../src/runtime/tensorrt-llm/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;GAkBG;AACH,cAAc,oBAAoB,CAAA;AAClC,cAAc,mCAAmC,CAAA;AACjD,cAAc,cAAc,CAAA;AAC5B,cAAc,eAAe,CAAA;AAC7B,cAAc,eAAe,CAAA;AAC7B,cAAc,mBAAmB,CAAA;AACjC,cAAc,mBAAmB,CAAA;AACjC,cAAc,iBAAiB,CAAA;AAC/B,cAAc,YAAY,CAAA;AAC1B,cAAc,gBAAgB,CAAA;AAC9B,cAAc,aAAa,CAAA"}

package/dist/runtime/tensorrt-llm/installation.d.ts

@@ -0,0 +1,23 @@
+import type { PlatformImage, RuntimeDescriptor, RuntimeInstallation } from '../../contracts/index.js';
+import type { InstallationStore, RuntimeDescriptorProvider } from '../environment/index.js';
+export type ContainerPlatform = 'linux/amd64' | 'linux/arm64';
+/** `process.arch` to the descriptor's image key; null for a CPU no image is published for. */
+export declare function containerPlatformFor(arch: string): ContainerPlatform | null;
+export interface ReadyInstallation {
+    installation: RuntimeInstallation;
+    descriptor: RuntimeDescriptor;
+    /** The descriptor's image for this host's platform, pinned by digest. */
+    image: PlatformImage;
+}
+export interface ResolveReadyInstallationDeps {
+    /** The setup operation's installation records (`InstallationStore`); a torn or foreign file is skipped there. */
+    installations: Pick<InstallationStore, 'list'>;
+    descriptors: Pick<RuntimeDescriptorProvider, 'forInstallation'>;
+    /** This host's container platform; null when the CPU has no published image. */
+    platform: ContainerPlatform | null;
+    /** The managed engine the load is for (change `add-vllm-runtime`); `tensorrt-llm` when omitted. */
+    engineId?: string;
+    /** That engine in messages; defaults to its known name, else its id. */
+    label?: string;
+}
+export declare function resolveReadyInstallation(deps: ResolveReadyInstallationDeps): Promise<ReadyInstallation>;

package/dist/runtime/tensorrt-llm/installation.js

@@ -0,0 +1,53 @@
+/**
+ * Which TensorRT-LLM installation a load runs (spec `tensorrt-llm-runtime`, "Загрузка без готовой
+ * установки"; design D7): only a `ready` one, only with the descriptor it was pinned to
+ * (`active_descriptor_id`, read from the descriptor cache — never fetched, so a release published
+ * since cannot change what runs), and only with that descriptor's image for this host's CPU. Anything
+ * short of that refuses the load with `MANAGED_ADAPTER_UNAVAILABLE` before a container exists.
+ *
+ * The installations themselves are read by the setup operation's own store (`InstallationStore`,
+ * `../environment/installations.ts`): the records under the shared per-user root that both scopes
+ * read, so a setup finished by the app is `ready` for the CLI core too. There is one reader of that
+ * format; this file only decides which of its installations a load may run.
+ */
+import { AtomicCoreError } from '../../contracts/index.js';
+/** The engine a load needs an installation of when the caller names none (the descriptor's own `engine_id`). */
+const ENGINE_ID = 'tensorrt-llm';
+const LABELS = { 'tensorrt-llm': 'TensorRT-LLM' };
+/** `process.arch` to the descriptor's image key; null for a CPU no image is published for. */
+export function containerPlatformFor(arch) {
+    if (arch === 'x64')
+        return 'linux/amd64';
+    if (arch === 'arm64')
+        return 'linux/arm64';
+    return null;
+}
+function unavailable(message, details) {
+    return new AtomicCoreError('MANAGED_ADAPTER_UNAVAILABLE', message, details);
+}
+export async function resolveReadyInstallation(deps) {
+    const engineId = deps.engineId ?? ENGINE_ID;
+    const label = deps.label ?? LABELS[engineId] ?? engineId;
+    const ours = (await deps.installations.list())
+        .map((record) => record.installation)
+        .filter((i) => i.engine_id === engineId);
+    const ready = ours.find((i) => i.status === 'ready' && i.active_descriptor_id !== null);
+    if (ready === undefined) {
+        const status = ours.map((i) => `${i.installation_id}: ${i.status}`).join(', ');
+        throw unavailable(ours.length === 0
+            ? `The ${label} engine is not installed; set it up before loading a model.`
+            : `The ${label} engine is not ready; finish or repair its setup before loading a model.`, status === '' ? undefined : status);
+    }
+    const resolved = await deps.descriptors.forInstallation(ready.active_descriptor_id);
+    if (resolved.kind !== 'available')
+        throw resolved.error;
+    const { descriptor } = resolved;
+    if (descriptor.engine_id !== engineId) {
+        throw new AtomicCoreError('MANAGED_METADATA_INVALID', 'The installation is pinned to a descriptor of another engine.', `${descriptor.descriptor_id}: ${descriptor.engine_id}`);
+    }
+    if (deps.platform === null) {
+        throw unavailable(`${label} publishes no image for this CPU architecture.`);
+    }
+    return { installation: ready, descriptor, image: descriptor.image[deps.platform] };
+}
+//# sourceMappingURL=installation.js.map

package/dist/runtime/tensorrt-llm/installation.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"installation.js","sourceRoot":"","sources":["../../../src/runtime/tensorrt-llm/installation.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AACH,OAAO,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAA;AAI1D,gHAAgH;AAChH,MAAM,SAAS,GAAG,cAAc,CAAA;AAChC,MAAM,MAAM,GAA2B,EAAE,cAAc,EAAE,cAAc,EAAE,CAAA;AAIzE,8FAA8F;AAC9F,MAAM,UAAU,oBAAoB,CAAC,IAAY;IAC/C,IAAI,IAAI,KAAK,KAAK;QAAE,OAAO,aAAa,CAAA;IACxC,IAAI,IAAI,KAAK,OAAO;QAAE,OAAO,aAAa,CAAA;IAC1C,OAAO,IAAI,CAAA;AACb,CAAC;AAqBD,SAAS,WAAW,CAAC,OAAe,EAAE,OAAgB;IACpD,OAAO,IAAI,eAAe,CAAC,6BAA6B,EAAE,OAAO,EAAE,OAAO,CAAC,CAAA;AAC7E,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,wBAAwB,CAC5C,IAAkC;IAElC,MAAM,QAAQ,GAAG,IAAI,CAAC,QAAQ,IAAI,SAAS,CAAA;IAC3C,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,IAAI,MAAM,CAAC,QAAQ,CAAC,IAAI,QAAQ,CAAA;IACxD,MAAM,IAAI,GAAG,CAAC,MAAM,IAAI,CAAC,aAAa,CAAC,IAAI,EAAE,CAAC;SAC3C,GAAG,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,MAAM,CAAC,YAAY,CAAC;SACpC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,SAAS,KAAK,QAAQ,CAAC,CAAA;IAC1C,MAAM,KAAK,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,MAAM,KAAK,OAAO,IAAI,CAAC,CAAC,oBAAoB,KAAK,IAAI,CAAC,CAAA;IACvF,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;QACxB,MAAM,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,CAAC,eAAe,KAAK,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;QAC9E,MAAM,WAAW,CACf,IAAI,CAAC,MAAM,KAAK,CAAC;YACf,CAAC,CAAC,OAAO,KAAK,6DAA6D;YAC3E,CAAC,CAAC,OAAO,KAAK,0EAA0E,EAC1F,MAAM,KAAK,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,MAAM,CACnC,CAAA;IACH,CAAC;IACD,MAAM,QAAQ,GAAG,MAAM,IAAI,CAAC,WAAW,CAAC,eAAe,CAAC,KAAK,CAAC,oBAA8B,CAAC,CAAA;IAC7F,IAAI,QAAQ,CAAC,IAAI,KAAK,WAAW;QAAE,MAAM,QAAQ,CAAC,KAAK,CAAA;IACvD,MAAM,EAAE,UAAU,EAAE,GAAG,QAAQ,CAAA;IAC/B,IAAI,UAAU,CAAC,SAAS,KAAK,QAAQ,EAAE,CAAC;QACtC,MAAM,IAAI,eAAe,CACvB,0BAA0B,EAC1B,+DAA+D,EAC/D,GAAG,UAAU,CAAC,aAAa,KAAK,UAAU,CAAC,SAAS,EAAE,CACvD,CAAA;IACH,CAAC;IACD,IAAI,IAAI,CAAC,QAAQ,KAAK,IAAI,EAAE,CAAC;QAC3B,MAAM,WAAW,CAAC,GAAG,KAAK,gDAAgD,CAAC,CAAA;IAC7E,CAAC;IACD,OAAO,EAAE,YAAY,EAAE,KAAK,EAAE,UAAU,EAAE,KAAK,EAAE,UAAU,CAAC,KAAK,CAAC,IAAI,CAAC,QAAQ,CAAC,EAAE,CAAA;AACpF,CAAC"}

package/dist/runtime/tensorrt-llm/kv-cache.d.ts

@@ -0,0 +1,25 @@
+/**
+ * The KV-cache numbers the launch (`adapter.ts`) and the memory check (`compatibility.ts`) must agree
+ * on, in one module both import: the pure policy never reaches into the adapter (and through it the
+ * managed-text lifecycle) for a constant.
+ */
+/** The one default KV-cache fraction: the launch passes it (`adapter.ts`'s `buildTensorrtLlmLaunch`, through the validated settings), the model check
+ *  and the pre-launch check size their reserve with it (`compatibility.ts`'s `kvCacheReserveBytes`,
+ *  through the same validated settings), and `src/settings/schema/tensorrt-llm.json` stores it
+ *  (pinned equal by `settings.test.ts`). Below `trtllm-serve`'s own 0.9 on purpose: at 0.9 an 8 GB
+ *  card ran out of memory on the engine's non-KV allocations after the KV cache took its share
+ *  (budget table `_no_capture_init_kv_cache: 3.50 / 0.35`); 0.85 failed too, 0.8 loaded (docs/decisions/2026-09-29-tensorrt-llm-kv-cache-fraction-0-8-and-oom-read-from-the-whole-log.md). */
+export declare const TENSORRT_LLM_DEFAULT_KV_CACHE_FREE_FRACTION = 0.8;
+/**
+ * How many full contexts the KV cache holds on a unified-memory card: one full-context request plus
+ * a concurrent one. A starting value, not a measurement (docs/decisions/2026-09-29-tensorrt-llm-
+ * unified-memory-kv-cache-bounded-by-tokens.md).
+ */
+export declare const TENSORRT_LLM_UNIFIED_KV_CONTEXTS = 2;
+/**
+ * The `kv_cache_config.max_tokens` a unified-memory launch writes, and the token count its memory
+ * check reserves KV for. On a unified-memory card (GB10/DGX Spark) "free GPU memory" is the system's
+ * free RAM, so `--kv_cache_free_gpu_memory_fraction` alone would hand most of the machine's memory to
+ * the KV cache whatever the context length; this bounds it to what the configured context needs.
+ */
+export declare function tensorrtLlmUnifiedKvMaxTokens(contextLength: number): number;

package/dist/runtime/tensorrt-llm/kv-cache.js

@@ -0,0 +1,28 @@
+/**
+ * The KV-cache numbers the launch (`adapter.ts`) and the memory check (`compatibility.ts`) must agree
+ * on, in one module both import: the pure policy never reaches into the adapter (and through it the
+ * managed-text lifecycle) for a constant.
+ */
+/** The one default KV-cache fraction: the launch passes it (`adapter.ts`'s `buildTensorrtLlmLaunch`, through the validated settings), the model check
+ *  and the pre-launch check size their reserve with it (`compatibility.ts`'s `kvCacheReserveBytes`,
+ *  through the same validated settings), and `src/settings/schema/tensorrt-llm.json` stores it
+ *  (pinned equal by `settings.test.ts`). Below `trtllm-serve`'s own 0.9 on purpose: at 0.9 an 8 GB
+ *  card ran out of memory on the engine's non-KV allocations after the KV cache took its share
+ *  (budget table `_no_capture_init_kv_cache: 3.50 / 0.35`); 0.85 failed too, 0.8 loaded (docs/decisions/2026-09-29-tensorrt-llm-kv-cache-fraction-0-8-and-oom-read-from-the-whole-log.md). */
+export const TENSORRT_LLM_DEFAULT_KV_CACHE_FREE_FRACTION = 0.8;
+/**
+ * How many full contexts the KV cache holds on a unified-memory card: one full-context request plus
+ * a concurrent one. A starting value, not a measurement (docs/decisions/2026-09-29-tensorrt-llm-
+ * unified-memory-kv-cache-bounded-by-tokens.md).
+ */
+export const TENSORRT_LLM_UNIFIED_KV_CONTEXTS = 2;
+/**
+ * The `kv_cache_config.max_tokens` a unified-memory launch writes, and the token count its memory
+ * check reserves KV for. On a unified-memory card (GB10/DGX Spark) "free GPU memory" is the system's
+ * free RAM, so `--kv_cache_free_gpu_memory_fraction` alone would hand most of the machine's memory to
+ * the KV cache whatever the context length; this bounds it to what the configured context needs.
+ */
+export function tensorrtLlmUnifiedKvMaxTokens(contextLength) {
+    return contextLength * TENSORRT_LLM_UNIFIED_KV_CONTEXTS;
+}
+//# sourceMappingURL=kv-cache.js.map

package/dist/runtime/tensorrt-llm/kv-cache.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"kv-cache.js","sourceRoot":"","sources":["../../../src/runtime/tensorrt-llm/kv-cache.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH;;;;;8LAK8L;AAC9L,MAAM,CAAC,MAAM,2CAA2C,GAAG,GAAG,CAAA;AAE9D;;;;GAIG;AACH,MAAM,CAAC,MAAM,gCAAgC,GAAG,CAAC,CAAA;AAEjD;;;;;GAKG;AACH,MAAM,UAAU,6BAA6B,CAAC,aAAqB;IACjE,OAAO,aAAa,GAAG,gCAAgC,CAAA;AACzD,CAAC"}

package/dist/runtime/tensorrt-llm/prelaunch.d.ts

@@ -0,0 +1,20 @@
+import type { GpuFacts, RuntimeDescriptor } from '../../contracts/index.js';
+import type { HostMemory, ManagedCheckEngine, ResolvedCheckpoint } from '../managed-models/compatibility.js';
+import type { ManagedModel } from '../managed-models/model-dir.js';
+export interface VerifyModelFilesOptions {
+    /** The card this load is about to use; already resolved (and possibly substituted) by the caller. */
+    gpuId: string;
+    /** The engine's hooks for the shared check (its memory rule, its checkpoint quirks). */
+    engine: ManagedCheckEngine;
+}
+/**
+ * The pre-`stopPrevious` half of the pre-launch check: files, architecture, format and compute
+ * capability — never memory (see the file banner). Throws `MODEL_FILE_NOT_FOUND`/`MODEL_FILE_CORRUPT`
+ * for a missing or resized checkpoint file, or the pure check's own error
+ * (`MODEL_INCOMPATIBLE`/`MANAGED_METADATA_INVALID`/`MANAGED_PREREQUISITE_BLOCKED`) for a checkpoint
+ * that no longer checks out. On success, returns the `ResolvedCheckpoint` `checkModelMemory` needs
+ * for the memory half — including `config.json`'s own verified `architectures`, which is what
+ * `runtime.ts` now looks the descriptor's `model_families` entry up by (finding 5), never
+ * `model.yml`'s possibly-stale copy.
+ */
+export declare function verifyModelFilesAndCompatibility(model: Pick<ManagedModel, 'dir' | 'repository' | 'revision' | 'files'>, descriptor: RuntimeDescriptor, gpus: readonly GpuFacts[], hostMemory: HostMemory, options: VerifyModelFilesOptions): Promise<ResolvedCheckpoint>;

package/dist/runtime/tensorrt-llm/prelaunch.js

@@ -0,0 +1,168 @@
+/**
+ * The pre-launch check `TensorrtLlmRuntime.load` runs once `deps.model` has resolved an installed
+ * model (task 2.16, spec `tensorrt-llm-models`, "Проверка файлов при загрузке"): every file
+ * `model.yml` recorded is still on disk at its declared size — a shard deleted after download is
+ * refused by name, the way a corrupt GGUF already is for llama.cpp (`MODEL_FILE_NOT_FOUND`/
+ * `MODEL_FILE_CORRUPT`, `../llamacpp/load-plan.ts`'s own convention, reused here) — and the
+ * compatibility verdict is recomputed against `config.json` and `hf_quant_config.json` exactly as
+ * they sit in the model's directory right now, not as `model.yml` last recorded them, and against
+ * the card this load is about to use (its `gpu_id` is passed in, already resolved and possibly
+ * substituted by the caller — this never re-picks a card).
+ *
+ * `verifyModelFilesAndCompatibility` below is everything above the memory line only — files,
+ * architecture, format, compute capability — and returns a `ResolvedCheckpoint`
+ * (`compatibility.ts`) rather than a final verdict, on purpose: memory is checked separately, later,
+ * by `checkModelMemory` (re-exported from `compatibility.ts`) called from `runtime.ts`'s
+ * `beforeCreate` hook, once `stopPrevious` has actually freed the card (task 2.16w round 1, finding
+ * 1 (Critical) — reading free memory here, before eviction, would refuse a same-card model switch
+ * outright on a single-GPU host). No container exists before either call; a failure in either never
+ * creates one.
+ *
+ * `hf_quant_config.json` is read whenever it exists on disk (finding 4), independent of whether
+ * `model.yml`'s own `files` list happens to mention it — `model.yml` is a record the app/CLI wrote
+ * when the download finished, and can be stale or incomplete in ways the file itself on disk is not.
+ *
+ * Direct `node:fs/promises`, no injected `exec`/`readFile`: reading the model's own checkpoint
+ * directory is local disk I/O, the same convention `model-dir.ts` already uses for `model.yml`
+ * itself — nothing here is a shell command or a network call that would need faking on a non-Linux
+ * test host.
+ */
+import { open, readFile, stat } from 'node:fs/promises';
+import { join } from 'node:path';
+import { AtomicCoreError } from '../../contracts/index.js';
+import { checkCheckpointFiles, normalizeWeightNames, safetensorsHeaderLength, safetensorsTensorNames, weightSafetensorsFiles, } from '../managed-models/compatibility.js';
+const CONFIG_FILE = 'config.json';
+const HF_QUANT_CONFIG_FILE = 'hf_quant_config.json';
+/** A `model.yml` file path, joined onto the model directory; refuses one that would climb out of it. */
+function safeJoin(dir, relativePath) {
+    const segments = relativePath.split('/');
+    if (segments.some((segment) => segment === '' || segment === '.' || segment === '..')) {
+        throw new AtomicCoreError('MANAGED_METADATA_INVALID', `model.yml names an unsafe file path: ${relativePath}`, relativePath);
+    }
+    return join(dir, ...segments);
+}
+async function fileSizeOnDisk(path) {
+    try {
+        return (await stat(path)).size;
+    }
+    catch {
+        return undefined;
+    }
+}
+/** Every file `model.yml` lists, present at its recorded size; a violation refuses by name. */
+async function assertFilesOnDisk(dir, files) {
+    for (const file of files) {
+        const path = safeJoin(dir, file.path);
+        const size = await fileSizeOnDisk(path);
+        if (size === undefined) {
+            throw new AtomicCoreError('MODEL_FILE_NOT_FOUND', `The model is missing a file recorded in model.yml: ${file.path}`, path);
+        }
+        if (size !== file.size) {
+            throw new AtomicCoreError('MODEL_FILE_CORRUPT', `${file.path} is ${size} bytes on disk, but model.yml recorded ${file.size}.`, path);
+        }
+    }
+}
+/** `null` when the file does not exist; throws `MANAGED_METADATA_INVALID` for one that exists but is not a JSON object. */
+async function readJsonObjectFile(path) {
+    let text;
+    try {
+        text = await readFile(path, 'utf8');
+    }
+    catch {
+        return null;
+    }
+    let parsed;
+    try {
+        parsed = JSON.parse(text);
+    }
+    catch (error) {
+        throw new AtomicCoreError('MANAGED_METADATA_INVALID', `${path} is not valid JSON.`, String(error));
+    }
+    if (parsed === null || typeof parsed !== 'object' || Array.isArray(parsed)) {
+        throw new AtomicCoreError('MANAGED_METADATA_INVALID', `${path} is not a JSON object.`, path);
+    }
+    return parsed;
+}
+/**
+ * The tensor names in one `.safetensors` file's own header (8-byte little-endian length, then JSON):
+ * what the file really holds, unlike `model.safetensors.index.json`, which a repository can ship
+ * stale or copied from another model (PrimeIntellect/GLM-0.5B: an 18k-name index over a 429-tensor
+ * file). `undefined` for a file that does not shape up as safetensors — the shape checks are then
+ * skipped, never failed on a header this core cannot read.
+ */
+async function tensorNamesOnDisk(path) {
+    const handle = await open(path, 'r').catch(() => undefined);
+    if (handle === undefined)
+        return undefined;
+    try {
+        const prefix = Buffer.alloc(8);
+        if ((await handle.read(prefix, 0, 8, 0)).bytesRead !== 8)
+            return undefined;
+        const length = safetensorsHeaderLength(prefix);
+        if (length === undefined)
+            return undefined;
+        const header = Buffer.alloc(length);
+        if ((await handle.read(header, 0, length, 8)).bytesRead !== length)
+            return undefined;
+        return safetensorsTensorNames(JSON.parse(header.toString('utf8')));
+    }
+    catch {
+        return undefined;
+    }
+    finally {
+        await handle.close();
+    }
+}
+/** Every weight file's tensor names, folded; `undefined` when there is none or any header is unreadable. */
+async function weightNamesOnDisk(dir, files) {
+    const weights = weightSafetensorsFiles(files);
+    if (weights.length === 0)
+        return undefined;
+    const names = [];
+    for (const file of weights) {
+        const own = await tensorNamesOnDisk(safeJoin(dir, file.path));
+        if (own === undefined)
+            return undefined;
+        names.push(...own);
+    }
+    return normalizeWeightNames(names);
+}
+/**
+ * The pre-`stopPrevious` half of the pre-launch check: files, architecture, format and compute
+ * capability — never memory (see the file banner). Throws `MODEL_FILE_NOT_FOUND`/`MODEL_FILE_CORRUPT`
+ * for a missing or resized checkpoint file, or the pure check's own error
+ * (`MODEL_INCOMPATIBLE`/`MANAGED_METADATA_INVALID`/`MANAGED_PREREQUISITE_BLOCKED`) for a checkpoint
+ * that no longer checks out. On success, returns the `ResolvedCheckpoint` `checkModelMemory` needs
+ * for the memory half — including `config.json`'s own verified `architectures`, which is what
+ * `runtime.ts` now looks the descriptor's `model_families` entry up by (finding 5), never
+ * `model.yml`'s possibly-stale copy.
+ */
+export async function verifyModelFilesAndCompatibility(model, descriptor, gpus, hostMemory, options) {
+    await assertFilesOnDisk(model.dir, model.files);
+    const configJson = await readJsonObjectFile(join(model.dir, CONFIG_FILE));
+    if (configJson === null) {
+        throw new AtomicCoreError('MODEL_FILE_NOT_FOUND', 'The model has no config.json in its directory.', join(model.dir, CONFIG_FILE));
+    }
+    // Read whenever the file is actually there, independent of model.yml's own files list (finding 4).
+    const hfQuantConfigJson = await readJsonObjectFile(join(model.dir, HF_QUANT_CONFIG_FILE));
+    // The same tensor names the download-time check gets from the app, so a checkpoint downloaded
+    // before that check knew to look (or by an older app) is refused here, before any container.
+    const weightNames = await weightNamesOnDisk(model.dir, model.files);
+    const input = {
+        repository: model.repository ?? '',
+        revision: model.revision ?? '',
+        config_json: configJson,
+        hf_quant_config_json: hfQuantConfigJson,
+        files: model.files,
+        gpu_id: options.gpuId,
+        ...(weightNames === undefined ? {} : { weight_names: weightNames }),
+    };
+    const result = checkCheckpointFiles(input, descriptor, gpus, hostMemory, options.engine);
+    if (!result.ok) {
+        // `FilesCheckResult`'s own type guarantees `result.verdict.verdict` is the failed branch here.
+        const { code, message, details } = result.verdict.verdict.error;
+        throw new AtomicCoreError(code, message, details);
+    }
+    return result.resolved;
+}
+//# sourceMappingURL=prelaunch.js.map

package/dist/runtime/tensorrt-llm/prelaunch.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"prelaunch.js","sourceRoot":"","sources":["../../../src/runtime/tensorrt-llm/prelaunch.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4BG;AACH,OAAO,EAAE,IAAI,EAAE,QAAQ,EAAE,IAAI,EAAE,MAAM,kBAAkB,CAAA;AACvD,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAA;AAChC,OAAO,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAA;AAE1D,OAAO,EACL,oBAAoB,EACpB,oBAAoB,EACpB,uBAAuB,EACvB,sBAAsB,EACtB,sBAAsB,GACvB,MAAM,oCAAoC,CAAA;AAW3C,MAAM,WAAW,GAAG,aAAa,CAAA;AACjC,MAAM,oBAAoB,GAAG,sBAAsB,CAAA;AAEnD,wGAAwG;AACxG,SAAS,QAAQ,CAAC,GAAW,EAAE,YAAoB;IACjD,MAAM,QAAQ,GAAG,YAAY,CAAC,KAAK,CAAC,GAAG,CAAC,CAAA;IACxC,IAAI,QAAQ,CAAC,IAAI,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,KAAK,EAAE,IAAI,OAAO,KAAK,GAAG,IAAI,OAAO,KAAK,IAAI,CAAC,EAAE,CAAC;QACtF,MAAM,IAAI,eAAe,CACvB,0BAA0B,EAC1B,wCAAwC,YAAY,EAAE,EACtD,YAAY,CACb,CAAA;IACH,CAAC;IACD,OAAO,IAAI,CAAC,GAAG,EAAE,GAAG,QAAQ,CAAC,CAAA;AAC/B,CAAC;AAED,KAAK,UAAU,cAAc,CAAC,IAAY;IACxC,IAAI,CAAC;QACH,OAAO,CAAC,MAAM,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,CAAA;IAChC,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,SAAS,CAAA;IAClB,CAAC;AACH,CAAC;AAED,+FAA+F;AAC/F,KAAK,UAAU,iBAAiB,CAAC,GAAW,EAAE,KAAgC;IAC5E,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QACzB,MAAM,IAAI,GAAG,QAAQ,CAAC,GAAG,EAAE,IAAI,CAAC,IAAI,CAAC,CAAA;QACrC,MAAM,IAAI,GAAG,MAAM,cAAc,CAAC,IAAI,CAAC,CAAA;QACvC,IAAI,IAAI,KAAK,SAAS,EAAE,CAAC;YACvB,MAAM,IAAI,eAAe,CACvB,sBAAsB,EACtB,sDAAsD,IAAI,CAAC,IAAI,EAAE,EACjE,IAAI,CACL,CAAA;QACH,CAAC;QACD,IAAI,IAAI,KAAK,IAAI,CAAC,IAAI,EAAE,CAAC;YACvB,MAAM,IAAI,eAAe,CACvB,oBAAoB,EACpB,GAAG,IAAI,CAAC,IAAI,OAAO,IAAI,0CAA0C,IAAI,CAAC,IAAI,GAAG,EAC7E,IAAI,CACL,CAAA;QACH,CAAC;IACH,CAAC;AACH,CAAC;AAED,2HAA2H;AAC3H,KAAK,UAAU,kBAAkB,CAAC,IAAY;IAC5C,IAAI,IAAY,CAAA;IAChB,IAAI,CAAC;QACH,IAAI,GAAG,MAAM,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC,CAAA;IACrC,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,IAAI,CAAA;IACb,CAAC;IACD,IAAI,MAAe,CAAA;IACnB,IAAI,CAAC;QACH,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAA;IAC3B,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,MAAM,IAAI,eAAe,CAAC,0BAA0B,EAAE,GAAG,IAAI,qBAAqB,EAAE,MAAM,CAAC,KAAK,CAAC,CAAC,CAAA;IACpG,CAAC;IACD,IAAI,MAAM,KAAK,IAAI,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,EAAE,CAAC;QAC3E,MAAM,IAAI,eAAe,CAAC,0BAA0B,EAAE,GAAG,IAAI,wBAAwB,EAAE,IAAI,CAAC,CAAA;IAC9F,CAAC;IACD,OAAO,MAAoB,CAAA;AAC7B,CAAC;AAED;;;;;;GAMG;AACH,KAAK,UAAU,iBAAiB,CAAC,IAAY;IAC3C,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,IAAI,EAAE,GAAG,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAA;IAC3D,IAAI,MAAM,KAAK,SAAS;QAAE,OAAO,SAAS,CAAA;IAC1C,IAAI,CAAC;QACH,MAAM,MAAM,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAA;QAC9B,IAAI,CAAC,MAAM,MAAM,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,SAAS,KAAK,CAAC;YAAE,OAAO,SAAS,CAAA;QAC1E,MAAM,MAAM,GAAG,uBAAuB,CAAC,MAAM,CAAC,CAAA;QAC9C,IAAI,MAAM,KAAK,SAAS;YAAE,OAAO,SAAS,CAAA;QAC1C,MAAM,MAAM,GAAG,MAAM,CAAC,KAAK,CAAC,MAAM,CAAC,CAAA;QACnC,IAAI,CAAC,MAAM,MAAM,CAAC,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,CAAC,SAAS,KAAK,MAAM;YAAE,OAAO,SAAS,CAAA;QACpF,OAAO,sBAAsB,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,CAAA;IACpE,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,SAAS,CAAA;IAClB,CAAC;YAAS,CAAC;QACT,MAAM,MAAM,CAAC,KAAK,EAAE,CAAA;IACtB,CAAC;AACH,CAAC;AAED,4GAA4G;AAC5G,KAAK,UAAU,iBAAiB,CAC9B,GAAW,EACX,KAAgC;IAEhC,MAAM,OAAO,GAAG,sBAAsB,CAAC,KAAK,CAAC,CAAA;IAC7C,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,SAAS,CAAA;IAC1C,MAAM,KAAK,GAAa,EAAE,CAAA;IAC1B,KAAK,MAAM,IAAI,IAAI,OAAO,EAAE,CAAC;QAC3B,MAAM,GAAG,GAAG,MAAM,iBAAiB,CAAC,QAAQ,CAAC,GAAG,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC,CAAA;QAC7D,IAAI,GAAG,KAAK,SAAS;YAAE,OAAO,SAAS,CAAA;QACvC,KAAK,CAAC,IAAI,CAAC,GAAG,GAAG,CAAC,CAAA;IACpB,CAAC;IACD,OAAO,oBAAoB,CAAC,KAAK,CAAC,CAAA;AACpC,CAAC;AASD;;;;;;;;;GASG;AACH,MAAM,CAAC,KAAK,UAAU,gCAAgC,CACpD,KAAsE,EACtE,UAA6B,EAC7B,IAAyB,EACzB,UAAsB,EACtB,OAAgC;IAEhC,MAAM,iBAAiB,CAAC,KAAK,CAAC,GAAG,EAAE,KAAK,CAAC,KAAK,CAAC,CAAA;IAE/C,MAAM,UAAU,GAAG,MAAM,kBAAkB,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,EAAE,WAAW,CAAC,CAAC,CAAA;IACzE,IAAI,UAAU,KAAK,IAAI,EAAE,CAAC;QACxB,MAAM,IAAI,eAAe,CACvB,sBAAsB,EACtB,gDAAgD,EAChD,IAAI,CAAC,KAAK,CAAC,GAAG,EAAE,WAAW,CAAC,CAC7B,CAAA;IACH,CAAC;IACD,mGAAmG;IACnG,MAAM,iBAAiB,GAAG,MAAM,kBAAkB,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,EAAE,oBAAoB,CAAC,CAAC,CAAA;IACzF,8FAA8F;IAC9F,6FAA6F;IAC7F,MAAM,WAAW,GAAG,MAAM,iBAAiB,CAAC,KAAK,CAAC,GAAG,EAAE,KAAK,CAAC,KAAK,CAAC,CAAA;IAEnE,MAAM,KAAK,GAAoB;QAC7B,UAAU,EAAE,KAAK,CAAC,UAAU,IAAI,EAAE;QAClC,QAAQ,EAAE,KAAK,CAAC,QAAQ,IAAI,EAAE;QAC9B,WAAW,EAAE,UAAU;QACvB,oBAAoB,EAAE,iBAAiB;QACvC,KAAK,EAAE,KAAK,CAAC,KAAK;QAClB,MAAM,EAAE,OAAO,CAAC,KAAK;QACrB,GAAG,CAAC,WAAW,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,YAAY,EAAE,WAAW,EAAE,CAAC;KACpE,CAAA;IACD,MAAM,MAAM,GAAG,oBAAoB,CAAC,KAAK,EAAE,UAAU,EAAE,IAAI,EAAE,UAAU,EAAE,OAAO,CAAC,MAAM,CAAC,CAAA;IACxF,IAAI,CAAC,MAAM,CAAC,EAAE,EAAE,CAAC;QACf,+FAA+F;QAC/F,MAAM,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,GAAI,MAAM,CAAC,OAAO,CAAC,OAA2C,CAAC,KAAK,CAAA;QACpG,MAAM,IAAI,eAAe,CAAC,IAAI,EAAE,OAAO,EAAE,OAAO,CAAC,CAAA;IACnD,CAAC;IACD,OAAO,MAAM,CAAC,QAAQ,CAAA;AACxB,CAAC"}

package/dist/runtime/tensorrt-llm/route-policy.d.ts

@@ -0,0 +1,19 @@
+/**
+ * What the public server (`:1337`) must know about a `tensorrt-llm` session to route it honestly
+ * (spec `tensorrt-llm-runtime`, "Возможности модели объявляются, а не угадываются" and "Переполнение
+ * контекста без авто-роста"; design D9): only the routes the adapter declares, tool calls only when
+ * the pinned descriptor names a parser for the model's family, and `trtllm-serve`'s context-overflow
+ * text mapped to OpenAI's `context_length_exceeded` — never a context increase and a replay.
+ */
+import type { SessionRoutePolicy } from '../shared/index.js';
+import type { ManagedTextCapabilities } from '../managed-text/index.js';
+/**
+ * `capabilities` are the loaded session's own, `limits` the context and output cap it was loaded
+ * with. `null` — a session the core did not start and so cannot describe (one the app registered as
+ * external) — keeps the route declaration and the error mapping but leaves tool calls and JSON
+ * output to the engine: refusing them would guess just as much as allowing.
+ */
+export declare function tensorrtLlmRoutePolicy(capabilities: ManagedTextCapabilities | null, limits?: {
+    contextLength: number;
+    maxOutputTokens: number;
+} | null): SessionRoutePolicy;

package/dist/runtime/tensorrt-llm/route-policy.js

@@ -0,0 +1,17 @@
+import { TENSORRT_LLM_ROUTES, mapTensorrtLlmContextLengthError } from './adapter.js';
+/**
+ * `capabilities` are the loaded session's own, `limits` the context and output cap it was loaded
+ * with. `null` — a session the core did not start and so cannot describe (one the app registered as
+ * external) — keeps the route declaration and the error mapping but leaves tool calls and JSON
+ * output to the engine: refusing them would guess just as much as allowing.
+ */
+export function tensorrtLlmRoutePolicy(capabilities, limits = null) {
+    return {
+        routes: TENSORRT_LLM_ROUTES,
+        tools: capabilities?.tools ?? true,
+        structuredOutput: capabilities?.structured_output ?? true,
+        mapError: mapTensorrtLlmContextLengthError,
+        ...(limits === null ? {} : limits),
+    };
+}
+//# sourceMappingURL=route-policy.js.map

package/dist/runtime/tensorrt-llm/route-policy.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"route-policy.js","sourceRoot":"","sources":["../../../src/runtime/tensorrt-llm/route-policy.ts"],"names":[],"mappings":"AASA,OAAO,EAAE,mBAAmB,EAAE,gCAAgC,EAAE,MAAM,cAAc,CAAA;AAEpF;;;;;GAKG;AACH,MAAM,UAAU,sBAAsB,CACpC,YAA4C,EAC5C,SAAoE,IAAI;IAExE,OAAO;QACL,MAAM,EAAE,mBAAmB;QAC3B,KAAK,EAAE,YAAY,EAAE,KAAK,IAAI,IAAI;QAClC,gBAAgB,EAAE,YAAY,EAAE,iBAAiB,IAAI,IAAI;QACzD,QAAQ,EAAE,gCAAgC;QAC1C,GAAG,CAAC,MAAM,KAAK,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,MAAM,CAAC;KACnC,CAAA;AACH,CAAC"}

package/dist/runtime/tensorrt-llm/settings.d.ts

@@ -0,0 +1,7 @@
+import type { TensorrtLlmSettings } from './adapter.js';
+/**
+ * Stored values, overlaid by a load's own overrides (the control API's `overrides`), reduced to this
+ * provider's keys and validated. Anything else in either object — another provider's key, a stray
+ * field a client sent — is ignored rather than passed to the engine.
+ */
+export declare function tensorrtLlmSettings(stored: Record<string, unknown>, overrides?: Record<string, unknown>): TensorrtLlmSettings;

package/dist/runtime/tensorrt-llm/settings.js

@@ -0,0 +1,48 @@
+/**
+ * The `tensorrt-llm` provider settings as the settings store holds them (`src/settings/schema/
+ * tensorrt-llm.json`, task 2.14) turned into what the adapter validates (`TensorrtLlmSettings`). The
+ * store keeps "not set" as a value a settings form can show — `gpu_id: ''`, `load_timeout_seconds:
+ * 0` — while the adapter spells it `null`; this is the one place that translates between the two.
+ * Validation itself is the adapter's (`validateTensorrtLlmSettings`), so a load is refused with
+ * `INVALID_ARGUMENT` before any container exists (spec "Настройки MUST валидироваться схемой до
+ * запуска контейнера").
+ */
+import { validateTensorrtLlmSettings } from './adapter.js';
+import { GENERATION_DEFAULT_KEYS } from '../managed-text/generation-defaults.js';
+const KEYS = [
+    'gpu_id',
+    'context_length',
+    'max_output_tokens',
+    'kv_cache_free_gpu_memory_fraction',
+    'max_batch_size',
+    'kv_cache_max_tokens',
+    'cuda_graphs',
+    'kv_cache_dtype',
+    'load_timeout_seconds',
+    'enable_prefix_caching',
+    'overlap_scheduler',
+    'capacity_scheduler_policy',
+    'dtype',
+    ...Object.values(GENERATION_DEFAULT_KEYS),
+];
+/**
+ * Stored values, overlaid by a load's own overrides (the control API's `overrides`), reduced to this
+ * provider's keys and validated. Anything else in either object — another provider's key, a stray
+ * field a client sent — is ignored rather than passed to the engine.
+ */
+export function tensorrtLlmSettings(stored, overrides = {}) {
+    const merged = {};
+    for (const key of KEYS) {
+        const value = key in overrides ? overrides[key] : stored[key];
+        if (value !== undefined)
+            merged[key] = value;
+    }
+    if (typeof merged['gpu_id'] === 'string' && merged['gpu_id'].trim() === '')
+        merged['gpu_id'] = null;
+    if (merged['load_timeout_seconds'] === 0)
+        merged['load_timeout_seconds'] = null;
+    if (merged['kv_cache_max_tokens'] === 0)
+        merged['kv_cache_max_tokens'] = null;
+    return validateTensorrtLlmSettings(merged);
+}
+//# sourceMappingURL=settings.js.map

package/dist/runtime/tensorrt-llm/settings.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"settings.js","sourceRoot":"","sources":["../../../src/runtime/tensorrt-llm/settings.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AACH,OAAO,EAAE,2BAA2B,EAAE,MAAM,cAAc,CAAA;AAE1D,OAAO,EAAE,uBAAuB,EAAE,MAAM,wCAAwC,CAAA;AAEhF,MAAM,IAAI,GAAG;IACX,QAAQ;IACR,gBAAgB;IAChB,mBAAmB;IACnB,mCAAmC;IACnC,gBAAgB;IAChB,qBAAqB;IACrB,aAAa;IACb,gBAAgB;IAChB,sBAAsB;IACtB,uBAAuB;IACvB,mBAAmB;IACnB,2BAA2B;IAC3B,OAAO;IACP,GAAG,MAAM,CAAC,MAAM,CAAC,uBAAuB,CAAC;CACjC,CAAA;AAEV;;;;GAIG;AACH,MAAM,UAAU,mBAAmB,CACjC,MAA+B,EAC/B,YAAqC,EAAE;IAEvC,MAAM,MAAM,GAA4B,EAAE,CAAA;IAC1C,KAAK,MAAM,GAAG,IAAI,IAAI,EAAE,CAAC;QACvB,MAAM,KAAK,GAAG,GAAG,IAAI,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,CAAA;QAC7D,IAAI,KAAK,KAAK,SAAS;YAAE,MAAM,CAAC,GAAG,CAAC,GAAG,KAAK,CAAA;IAC9C,CAAC;IACD,IAAI,OAAO,MAAM,CAAC,QAAQ,CAAC,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE;QAAE,MAAM,CAAC,QAAQ,CAAC,GAAG,IAAI,CAAA;IACnG,IAAI,MAAM,CAAC,sBAAsB,CAAC,KAAK,CAAC;QAAE,MAAM,CAAC,sBAAsB,CAAC,GAAG,IAAI,CAAA;IAC/E,IAAI,MAAM,CAAC,qBAAqB,CAAC,KAAK,CAAC;QAAE,MAAM,CAAC,qBAAqB,CAAC,GAAG,IAAI,CAAA;IAC7E,OAAO,2BAA2B,CAAC,MAAM,CAAC,CAAA;AAC5C,CAAC"}

package/dist/runtime/vllm/adapter.d.ts

@@ -0,0 +1,46 @@
+import type { ModelFamilySupport } from '../../contracts/index.js';
+import { type ManagedExitClassification, type ManagedLaunchContext, type ManagedEngineLaunch, type ManagedRoute, type ManagedTextAdapter, type ManagedTextCapabilities } from '../managed-text/index.js';
+import { type VllmSettings } from './settings.js';
+/** The port `vllm serve` listens on inside its container. */
+export declare const VLLM_CONTAINER_PORT = 8000;
+/** Below this much card memory, `cuda_graphs: auto` runs eager: graphs cost memory a small card needs. */
+export declare const VLLM_CUDA_GRAPHS_MIN_VRAM_BYTES: number;
+/** What core decides for a launch from the card as it stands right before the container (design D9). */
+export interface VllmLaunchPlan {
+    /** `--kv-cache-memory-bytes` from the KV cache size setting; `null` lets vLLM size it within its share. */
+    kvCacheMemoryBytes: number | null;
+    /** `--gpu-memory-utilization`: what vLLM's start-up check compares the card's free memory against. */
+    gpuMemoryUtilization: number;
+    /**
+     * The checkpoint has a vision or audio part: `--limit-mm-per-prompt` then reserves nothing for it.
+     * A text-only model gets no multimodal flag at all.
+     */
+    multimodal: boolean;
+}
+export declare function buildVllmLaunch(context: ManagedLaunchContext<VllmSettings>): ManagedEngineLaunch;
+/** A base for engine start-up, imports and the first graph compilation; more per GiB of weights. */
+export declare const VLLM_READINESS_BASE_MS: number;
+export declare const VLLM_READINESS_PER_GIB_MS = 20000;
+export declare function vllmReadinessTimeoutMs(weightBytes: number, settings: VllmSettings): number;
+export declare function classifyVllmExit(log: string, exitCode: number | null): ManagedExitClassification;
+export declare function vllmCapabilities(context: {
+    settings: VllmSettings;
+    family: ModelFamilySupport | null;
+}): ManagedTextCapabilities;
+export declare const VLLM_ROUTES: readonly ManagedRoute[];
+export declare const VLLM_REWRITABLE_ROUTES: readonly ManagedRoute[];
+/**
+ * Caps the output a request asks for at `max_output_tokens` (one key, as the client sent it); a
+ * request that asks for none is capped by vLLM's launch default instead. Refuses what the session cannot do with `:1337`'s own wording
+ * and code: tools without a parser, structured output the family does not declare, images.
+ */
+export declare function vllmRewriteRequestBody(route: string, body: unknown, settings: VllmSettings, capabilities?: ManagedTextCapabilities): unknown;
+/** vLLM's context overflow as OpenAI's `context_length_exceeded`; anything else relayed as it is. */
+export declare function mapVllmContextLengthError(status: number, body: string): object | null;
+/**
+ * Reasoning parsers that may file a no-thinking reply under `reasoning_content` (design D8: the same
+ * hook as TensorRT-LLM's; the live run of task 6.1 confirms which of vLLM 0.31's do).
+ */
+export declare const VLLM_REASONING_AT_START_PARSERS: ReadonlySet<string>;
+export declare function vllmRewriteResponseFor(route: string, requestBody: unknown, family: ModelFamilySupport | null): ((json: Record<string, unknown>) => Record<string, unknown>) | null;
+export declare const vllmAdapter: ManagedTextAdapter<VllmSettings>;