@atomic-chat/core 0.5.1 → 0.13.2

package/dist/runtime/vllm/adapter.js

@@ -0,0 +1,425 @@
+/**
+ * The `vllm` engine's `ManagedTextAdapter` (change `add-vllm-runtime`, task 3.2, design D8; spec
+ * `vllm-runtime`): how core runs `vllm serve` from the `vllm/vllm-openai` image on the shared
+ * managed-text lifecycle.
+ *
+ * - **argv**: `vllm serve <model> --served-model-name <id> --host 0.0.0.0 --port 8000
+ *   --max-model-len <ctx> --max-num-seqs <n> --gpu-memory-utilization <U>
+ *   --override-generation-config {"max_new_tokens":<cap>, …sampling defaults}`,
+ *   `--limit-mm-per-prompt {"image":0,"video":0}` for a multimodal checkpoint, the further options the
+ *   settings set (`--max-num-batched-tokens`, `--no-enable-prefix-caching`, `--cpu-offload-gb`,
+ *   `--dtype`, `--seed`, `--async-scheduling`, `--kv-cache-memory-bytes` when the KV cache size is
+ *   set), plus `--enforce-eager` when CUDA graphs are off,
+ *   `--kv-cache-dtype fp8` on compute capability 8.9 and newer, and the family's tool-call and
+ *   reasoning parsers. `U` is core's (`plan`, from the card as it stands right before the container
+ *   is created, design D9): vLLM's default share of the card's *total* memory would refuse to start on
+ *   any desktop whose card is partly taken. Within `U`, vLLM sizes the KV cache itself — it knows every
+ *   model's KV shape, hybrid ones included — unless the KV cache size setting fixes it. Never `--trust-remote-code` (no code from a model repository runs), `--enable-log-requests`
+ *   (prompts would land in the logs core serves) or `--api-key` (the session gateway checks the key).
+ * - **env**: usage stats off (`VLLM_NO_USAGE_STATS`, `DO_NOT_TRACK`), Hugging Face offline (every file
+ *   is already in the store), and every compile cache — vLLM's own, Triton's, inductor's, FlashInfer's
+ *   — under the engine cache, so a second start of the same model reuses the first one's work.
+ * - **readiness** `GET /health`; **routes** chat completions, completions and the model list;
+ *   requests capped at the output setting, refused for tools without a parser, structured output the
+ *   family does not declare, and images (a cap the client did not ask for comes from the launch's
+ *   default, so a long prompt is not refused for the room a written cap would take); vLLM's context overflow answered as OpenAI's
+ *   `context_length_exceeded` with both numbers.
+ *
+ * Log lines, exit wording and error bodies follow vLLM 0.31's source (`test/helpers/vllm-log-fixtures.ts`);
+ * the live run of task 6.1 confirms or replaces them.
+ */
+import { AtomicCoreError } from '../../contracts/index.js';
+import { MANAGED_TEXT_ADAPTER_CONTRACT_VERSION, ManagedRequestRefusal, asksForStructuredOutput, asksForTools, reasoningIntoContent, thinkingRequested, } from '../managed-text/index.js';
+import { validateVllmSettings } from './settings.js';
+const GiB = 1024 ** 3;
+/** The port `vllm serve` listens on inside its container. */
+export const VLLM_CONTAINER_PORT = 8000;
+/** Below this much card memory, `cuda_graphs: auto` runs eager: graphs cost memory a small card needs. */
+export const VLLM_CUDA_GRAPHS_MIN_VRAM_BYTES = 12 * GiB;
+function isPlan(value) {
+    if (value === null || typeof value !== 'object')
+        return false;
+    const plan = value;
+    return ((plan['kvCacheMemoryBytes'] === null ||
+        (typeof plan['kvCacheMemoryBytes'] === 'number' &&
+            Number.isInteger(plan['kvCacheMemoryBytes']) &&
+            plan['kvCacheMemoryBytes'] > 0)) &&
+        typeof plan['gpuMemoryUtilization'] === 'number' &&
+        plan['gpuMemoryUtilization'] > 0 &&
+        plan['gpuMemoryUtilization'] <= 1 &&
+        typeof plan['multimodal'] === 'boolean');
+}
+/** Compute capability 8.9 (Ada) and newer: FP8 KV cache. Unknown or older: the model's own precision. */
+function supportsFp8Kv(computeCapability) {
+    if (computeCapability === null || computeCapability === undefined)
+        return false;
+    const [major, minor] = computeCapability.split('.').map((part) => Number.parseInt(part, 10));
+    if (major === undefined || Number.isNaN(major))
+        return false;
+    return major > 8 || (major === 8 && (minor ?? 0) >= 9);
+}
+function eager(context) {
+    const mode = context.settings.cuda_graphs;
+    if (mode === 'on')
+        return false;
+    if (mode === 'off')
+        return true;
+    // A unified-memory card has the host's memory: never the small-card case.
+    if (context.unifiedMemory)
+        return false;
+    const total = context.gpuTotalVramBytes ?? null;
+    return total === null || total < VLLM_CUDA_GRAPHS_MIN_VRAM_BYTES;
+}
+/** Every compile cache of the engine, under the engine cache (`ManagedLaunchContext.engineCachePath`). */
+function cacheEnv(cache) {
+    return {
+        VLLM_CACHE_ROOT: `${cache}/vllm`,
+        TRITON_CACHE_DIR: `${cache}/triton`,
+        TORCHINDUCTOR_CACHE_DIR: `${cache}/inductor`,
+        FLASHINFER_WORKSPACE_BASE: `${cache}/flashinfer`,
+    };
+}
+export function buildVllmLaunch(context) {
+    if (!isPlan(context.plan)) {
+        throw new AtomicCoreError('INVALID_ARGUMENT', 'A vLLM launch needs core’s memory plan (KV cache bytes and memory share), computed right before the container is created.');
+    }
+    const { settings, family } = context;
+    const argv = [
+        'vllm',
+        'serve',
+        context.modelPath,
+        '--served-model-name',
+        context.modelId,
+        '--host',
+        '0.0.0.0',
+        '--port',
+        String(VLLM_CONTAINER_PORT),
+        '--max-model-len',
+        String(settings.context_length),
+        '--max-num-seqs',
+        String(settings.max_num_seqs),
+        '--gpu-memory-utilization',
+        String(Number(context.plan.gpuMemoryUtilization.toFixed(4))),
+        // The output cap for a request that names none: vLLM's own default, so a long prompt keeps the
+        // room it has (a cap written into every request would make vLLM refuse input + cap > context).
+        '--override-generation-config',
+        JSON.stringify({ max_new_tokens: settings.max_output_tokens, ...settings.generation }),
+    ];
+    if (context.plan.kvCacheMemoryBytes !== null) {
+        argv.push('--kv-cache-memory-bytes', String(context.plan.kvCacheMemoryBytes));
+    }
+    if (settings.max_num_batched_tokens !== null) {
+        argv.push('--max-num-batched-tokens', String(settings.max_num_batched_tokens));
+    }
+    if (!settings.enable_prefix_caching)
+        argv.push('--no-enable-prefix-caching');
+    if (settings.cpu_offload_gb > 0)
+        argv.push('--cpu-offload-gb', String(settings.cpu_offload_gb));
+    if (settings.dtype !== 'auto')
+        argv.push('--dtype', settings.dtype);
+    if (settings.seed !== null)
+        argv.push('--seed', String(settings.seed));
+    if (settings.async_scheduling)
+        argv.push('--async-scheduling');
+    if (context.plan.multimodal)
+        argv.push('--limit-mm-per-prompt', JSON.stringify({ image: 0, video: 0 }));
+    if (eager(context))
+        argv.push('--enforce-eager');
+    if (settings.kv_cache_dtype === 'fp8' && supportsFp8Kv(context.gpuComputeCapability)) {
+        argv.push('--kv-cache-dtype', 'fp8');
+    }
+    if (family?.tool_parser)
+        argv.push('--enable-auto-tool-choice', '--tool-call-parser', family.tool_parser);
+    if (family?.reasoning_parser)
+        argv.push('--reasoning-parser', family.reasoning_parser);
+    return {
+        engine: { container_port: VLLM_CONTAINER_PORT },
+        argv,
+        env: {
+            VLLM_NO_USAGE_STATS: '1',
+            DO_NOT_TRACK: '1',
+            HF_HUB_OFFLINE: '1',
+            TRANSFORMERS_OFFLINE: '1',
+            ...cacheEnv(context.engineCachePath),
+        },
+    };
+}
+/** A base for engine start-up, imports and the first graph compilation; more per GiB of weights. */
+export const VLLM_READINESS_BASE_MS = 10 * 60_000;
+export const VLLM_READINESS_PER_GIB_MS = 20_000;
+export function vllmReadinessTimeoutMs(weightBytes, settings) {
+    if (settings.load_timeout_seconds !== null)
+        return settings.load_timeout_seconds * 1000;
+    return Math.ceil(VLLM_READINESS_BASE_MS + (weightBytes / GiB) * VLLM_READINESS_PER_GIB_MS);
+}
+// ---------------------------------------------------------------------------------------------
+// Exit classification
+// ---------------------------------------------------------------------------------------------
+const OOM = /CUDA out of memory\. Tried to allocate ([\d.]+) ?(GiB|MiB|KiB).*?of which ([\d.]+) ?(GiB|MiB|KiB|bytes) is free/;
+// vLLM 0.31 names the device (`on device cuda:0 (6.89/8.0 GiB)`); earlier releases did not.
+const FREE_BELOW_SHARE = /Free memory on device(?: [^\s(]+)? \(([\d.]+)\/([\d.]+) GiB\) on startup is less than desired GPU memory utilization \(([\d.]+), ([\d.]+) GiB\)/;
+const KV_TOO_SMALL = /max seq len \((\d+)\).*?\(([\d.]+) GiB KV cache is needed, which is larger than the available KV cache memory \(([\d.]+) GiB\)\.(?: Based on the available memory, the estimated maximum model length is (\d+)\.)?/;
+const UNSUPPORTED_ARCHITECTURE = /Model architectures \[([^\]]*)\] are not supported for now/;
+const ANY_OOM = /out of memory/i;
+const toGiB = (amount, unit) => unit === 'GiB'
+    ? amount
+    : unit === 'MiB'
+        ? amount / 1024
+        : unit === 'KiB'
+            ? amount / 1024 ** 2
+            : amount / GiB;
+const round = (value) => Math.round(value * 100) / 100;
+export function classifyVllmExit(log, exitCode) {
+    const kv = KV_TOO_SMALL.exec(log);
+    if (kv !== null) {
+        const maxModelLen = Number(kv[1]);
+        const fits = kv[4] === undefined ? null : Number(kv[4]);
+        return {
+            kind: 'out-of-memory',
+            message: `The context length ${maxModelLen} does not fit the KV cache: it needs ${kv[2]} GiB, ${kv[3]} GiB is available` +
+                (fits === null ? '.' : `, which holds ${fits} tokens. Lower the context length or free GPU memory.`),
+            numbers: {
+                max_model_len: maxModelLen,
+                kv_needed_gib: Number(kv[2]),
+                kv_available_gib: Number(kv[3]),
+                ...(fits === null ? {} : { kv_cache_tokens: fits }),
+            },
+            excerpt: kv[0],
+        };
+    }
+    const free = FREE_BELOW_SHARE.exec(log);
+    if (free !== null) {
+        return {
+            kind: 'out-of-memory',
+            message: `The GPU had ${free[1]} GiB of ${free[2]} GiB free once vLLM had started, less than the ${free[4]} GiB ` +
+                `the model's weights and KV cache need. Free GPU memory, or lower the context length or concurrent requests.`,
+            numbers: {
+                free_gib: Number(free[1]),
+                total_gib: Number(free[2]),
+                utilization: Number(free[3]),
+                requested_gib: Number(free[4]),
+            },
+            excerpt: free[0],
+        };
+    }
+    const oom = OOM.exec(log);
+    if (oom !== null) {
+        const tried = round(toGiB(Number(oom[1]), oom[2]));
+        const left = round(toGiB(Number(oom[3]), oom[4]));
+        return {
+            kind: 'out-of-memory',
+            message: `The GPU ran out of memory: vLLM tried to allocate ${oom[1]} ${oom[2]} with ${oom[3]} ${oom[4]} free.`,
+            numbers: { requested_gib: tried, free_gib: left },
+            excerpt: oom[0],
+        };
+    }
+    const arch = UNSUPPORTED_ARCHITECTURE.exec(log);
+    if (arch !== null) {
+        return {
+            kind: 'unsupported-model',
+            message: `This vLLM release cannot run the model's architecture: ${arch[1]?.replace(/'/g, '')}.`,
+            excerpt: arch[0],
+        };
+    }
+    if (ANY_OOM.test(log)) {
+        return { kind: 'out-of-memory', message: 'The GPU ran out of memory while vLLM started.' };
+    }
+    return {
+        kind: 'other',
+        message: `vLLM exited${exitCode === null ? '' : ` with code ${exitCode}`} before it was ready.`,
+    };
+}
+// ---------------------------------------------------------------------------------------------
+// Capabilities, routes, request and response rewrites
+// ---------------------------------------------------------------------------------------------
+export function vllmCapabilities(context) {
+    const family = context.family;
+    return {
+        tools: Boolean(family?.tool_parser),
+        reasoning: Boolean(family?.reasoning_parser),
+        structured_output: family?.structured_output === true,
+        vision: false,
+        embeddings: false,
+        responses: false,
+    };
+}
+export const VLLM_ROUTES = [
+    { method: 'POST', path: '/v1/chat/completions' },
+    { method: 'POST', path: '/v1/completions' },
+    { method: 'GET', path: '/v1/models' },
+];
+export const VLLM_REWRITABLE_ROUTES = [
+    { method: 'POST', path: '/v1/chat/completions' },
+    { method: 'POST', path: '/v1/completions' },
+];
+const REWRITABLE = new Set(VLLM_REWRITABLE_ROUTES.map((route) => route.path));
+function positive(body, key) {
+    if (!(key in body) || body[key] === null)
+        return { present: false, value: 0 };
+    const value = body[key];
+    if (typeof value !== 'number' || !Number.isInteger(value) || value <= 0) {
+        throw new AtomicCoreError('INVALID_ARGUMENT', `${key} must be a positive integer.`);
+    }
+    return { present: true, value };
+}
+/** Any message part that is not text: an image, a video, audio — none of which this engine accepts here. */
+const MEDIA_PART = new Set([
+    'image_url',
+    'input_image',
+    'image',
+    'video_url',
+    'video',
+    'input_audio',
+    'audio_url',
+]);
+function carriesMedia(body) {
+    const messages = body['messages'];
+    if (!Array.isArray(messages))
+        return false;
+    return messages.some((message) => {
+        const content = message !== null && typeof message === 'object' ? message.content : null;
+        return (Array.isArray(content) &&
+            content.some((part) => part !== null &&
+                typeof part === 'object' &&
+                MEDIA_PART.has(String(part.type))));
+    });
+}
+/**
+ * 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 function vllmRewriteRequestBody(route, body, settings, capabilities) {
+    if (!REWRITABLE.has(route))
+        return body;
+    if (typeof body !== 'object' || body === null || Array.isArray(body))
+        return body;
+    const client = body;
+    const model = typeof client['model'] === 'string' ? `The model '${client['model']}'` : 'This model';
+    if (carriesMedia(client)) {
+        throw new ManagedRequestRefusal(`${model} does not support images.`, 'unsupported_capability');
+    }
+    if (capabilities !== undefined) {
+        if (!capabilities.tools && asksForTools(client)) {
+            throw new ManagedRequestRefusal(`${model} does not support tool calling.`, 'unsupported_capability');
+        }
+        if (!capabilities.structured_output && asksForStructuredOutput(client)) {
+            throw new ManagedRequestRefusal(`${model} does not support structured output.`, 'unsupported_capability');
+        }
+    }
+    const cap = settings.max_output_tokens;
+    if (route === '/v1/completions') {
+        const field = positive(client, 'max_tokens');
+        return field.present ? { ...client, max_tokens: Math.min(field.value, cap) } : client;
+    }
+    const legacy = positive(client, 'max_tokens');
+    const modern = positive(client, 'max_completion_tokens');
+    const asked = [legacy, modern].filter((field) => field.present).map((field) => field.value);
+    // Nothing asked: vLLM's launch default (`--override-generation-config`) caps it, within the room left.
+    if (asked.length === 0)
+        return client;
+    const rest = { ...client };
+    delete rest['max_tokens'];
+    delete rest['max_completion_tokens'];
+    rest[modern.present ? 'max_completion_tokens' : 'max_tokens'] = Math.min(cap, ...asked);
+    return rest;
+}
+/** vLLM's 400 wording for a request over the context, current and older, with the limit and the request. */
+const OVERFLOW = [
+    /maximum context length is (\d+) tokens\. However, your request has (\d+) input tokens/,
+    /maximum context length is (\d+) tokens\. However, you requested (\d+) tokens/,
+    /maximum context length is (\d+) tokens and your request has (\d+) input tokens/,
+];
+function errorMessage(body) {
+    try {
+        const parsed = JSON.parse(body);
+        const message = parsed.error?.message ?? parsed.message;
+        return typeof message === 'string' ? message : '';
+    }
+    catch {
+        return '';
+    }
+}
+/** vLLM's context overflow as OpenAI's `context_length_exceeded`; anything else relayed as it is. */
+export function mapVllmContextLengthError(status, body) {
+    if (status !== 400)
+        return null;
+    const message = errorMessage(body);
+    for (const pattern of OVERFLOW) {
+        const match = pattern.exec(message);
+        if (match === null)
+            continue;
+        const limit = Number(match[1]);
+        const requested = Number(match[2]);
+        return {
+            error: {
+                message: `This model's maximum context length is ${limit} tokens, but the request has ${requested} tokens. Shorten the conversation or the requested output.`,
+                type: 'invalid_request_error',
+                param: 'messages',
+                code: 'context_length_exceeded',
+            },
+        };
+    }
+    return 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 const VLLM_REASONING_AT_START_PARSERS = new Set(['qwen3']);
+export function vllmRewriteResponseFor(route, requestBody, family) {
+    if (route !== '/v1/chat/completions')
+        return null;
+    const parser = family?.reasoning_parser ?? null;
+    if (parser === null || !VLLM_REASONING_AT_START_PARSERS.has(parser))
+        return null;
+    return thinkingRequested(requestBody) ? null : reasoningIntoContent;
+}
+/**
+ * `initializing-engine` from the first line that shows vLLM working on the model: reading weights,
+ * profiling memory, compiling, capturing CUDA graphs. Before them the load stays in
+ * `starting-container` (Python, CUDA and the API server coming up).
+ */
+const STAGE_MARKERS = [
+    { stage: 'initializing-engine', pattern: /Starting to load model/ },
+    { stage: 'initializing-engine', pattern: /Loading safetensors checkpoint shards/ },
+    { stage: 'initializing-engine', pattern: /Model loading took/ },
+    { stage: 'initializing-engine', pattern: /Compiling a graph for dynamic shape/ },
+    { stage: 'initializing-engine', pattern: /Available KV cache memory/ },
+    { stage: 'initializing-engine', pattern: /Capturing CUDA graphs/ },
+];
+export const vllmAdapter = {
+    id: 'vllm',
+    contractVersion: MANAGED_TEXT_ADAPTER_CONTRACT_VERSION,
+    readiness: { path: '/health', expectedStatus: 200 },
+    routes: VLLM_ROUTES,
+    rewritableRoutes: VLLM_REWRITABLE_ROUTES,
+    stageMarkers: STAGE_MARKERS,
+    validateSettings: (raw) => validateVllmSettings((raw ?? {})),
+    buildLaunch: buildVllmLaunch,
+    readinessTimeoutMs: vllmReadinessTimeoutMs,
+    classifyExit: classifyVllmExit,
+    capabilities: vllmCapabilities,
+    rewriteRequestBody: vllmRewriteRequestBody,
+    rewriteResponseFor: vllmRewriteResponseFor,
+    mapErrorResponse: (_route, status, body) => mapVllmContextLengthError(status, body),
+    // Everything vLLM is started with — the output cap too, its default for a request that names none;
+    // only the load timeout is a load's alone.
+    restartKey: (settings) => [
+        settings.context_length,
+        settings.max_output_tokens,
+        settings.max_num_seqs,
+        settings.kv_cache_memory_gib,
+        settings.gpu_memory_utilization,
+        settings.cuda_graphs,
+        settings.kv_cache_dtype,
+        settings.max_num_batched_tokens,
+        settings.enable_prefix_caching,
+        settings.cpu_offload_gb,
+        settings.dtype,
+        settings.seed,
+        settings.async_scheduling,
+        JSON.stringify(settings.generation),
+    ],
+};
+//# sourceMappingURL=adapter.js.map

package/dist/runtime/vllm/adapter.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"adapter.js","sourceRoot":"","sources":["../../../src/runtime/vllm/adapter.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA6BG;AACH,OAAO,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAA;AAE1D,OAAO,EACL,qCAAqC,EACrC,qBAAqB,EACrB,uBAAuB,EACvB,YAAY,EACZ,oBAAoB,EACpB,iBAAiB,GAQlB,MAAM,0BAA0B,CAAA;AACjC,OAAO,EAAE,oBAAoB,EAAqB,MAAM,eAAe,CAAA;AAEvE,MAAM,GAAG,GAAG,IAAI,IAAI,CAAC,CAAA;AAErB,6DAA6D;AAC7D,MAAM,CAAC,MAAM,mBAAmB,GAAG,IAAI,CAAA;AAEvC,0GAA0G;AAC1G,MAAM,CAAC,MAAM,+BAA+B,GAAG,EAAE,GAAG,GAAG,CAAA;AAevD,SAAS,MAAM,CAAC,KAAc;IAC5B,IAAI,KAAK,KAAK,IAAI,IAAI,OAAO,KAAK,KAAK,QAAQ;QAAE,OAAO,KAAK,CAAA;IAC7D,MAAM,IAAI,GAAG,KAAgC,CAAA;IAC7C,OAAO,CACL,CAAC,IAAI,CAAC,oBAAoB,CAAC,KAAK,IAAI;QAClC,CAAC,OAAO,IAAI,CAAC,oBAAoB,CAAC,KAAK,QAAQ;YAC7C,MAAM,CAAC,SAAS,CAAC,IAAI,CAAC,oBAAoB,CAAC,CAAC;YAC5C,IAAI,CAAC,oBAAoB,CAAC,GAAG,CAAC,CAAC,CAAC;QACpC,OAAO,IAAI,CAAC,sBAAsB,CAAC,KAAK,QAAQ;QAChD,IAAI,CAAC,sBAAsB,CAAC,GAAG,CAAC;QAChC,IAAI,CAAC,sBAAsB,CAAC,IAAI,CAAC;QACjC,OAAO,IAAI,CAAC,YAAY,CAAC,KAAK,SAAS,CACxC,CAAA;AACH,CAAC;AAED,yGAAyG;AACzG,SAAS,aAAa,CAAC,iBAA4C;IACjE,IAAI,iBAAiB,KAAK,IAAI,IAAI,iBAAiB,KAAK,SAAS;QAAE,OAAO,KAAK,CAAA;IAC/E,MAAM,CAAC,KAAK,EAAE,KAAK,CAAC,GAAG,iBAAiB,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,MAAM,CAAC,QAAQ,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC,CAAA;IAC5F,IAAI,KAAK,KAAK,SAAS,IAAI,MAAM,CAAC,KAAK,CAAC,KAAK,CAAC;QAAE,OAAO,KAAK,CAAA;IAC5D,OAAO,KAAK,GAAG,CAAC,IAAI,CAAC,KAAK,KAAK,CAAC,IAAI,CAAC,KAAK,IAAI,CAAC,CAAC,IAAI,CAAC,CAAC,CAAA;AACxD,CAAC;AAED,SAAS,KAAK,CAAC,OAA2C;IACxD,MAAM,IAAI,GAAG,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAA;IACzC,IAAI,IAAI,KAAK,IAAI;QAAE,OAAO,KAAK,CAAA;IAC/B,IAAI,IAAI,KAAK,KAAK;QAAE,OAAO,IAAI,CAAA;IAC/B,0EAA0E;IAC1E,IAAI,OAAO,CAAC,aAAa;QAAE,OAAO,KAAK,CAAA;IACvC,MAAM,KAAK,GAAG,OAAO,CAAC,iBAAiB,IAAI,IAAI,CAAA;IAC/C,OAAO,KAAK,KAAK,IAAI,IAAI,KAAK,GAAG,+BAA+B,CAAA;AAClE,CAAC;AAED,0GAA0G;AAC1G,SAAS,QAAQ,CAAC,KAAa;IAC7B,OAAO;QACL,eAAe,EAAE,GAAG,KAAK,OAAO;QAChC,gBAAgB,EAAE,GAAG,KAAK,SAAS;QACnC,uBAAuB,EAAE,GAAG,KAAK,WAAW;QAC5C,yBAAyB,EAAE,GAAG,KAAK,aAAa;KACjD,CAAA;AACH,CAAC;AAED,MAAM,UAAU,eAAe,CAAC,OAA2C;IACzE,IAAI,CAAC,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC;QAC1B,MAAM,IAAI,eAAe,CACvB,kBAAkB,EAClB,2HAA2H,CAC5H,CAAA;IACH,CAAC;IACD,MAAM,EAAE,QAAQ,EAAE,MAAM,EAAE,GAAG,OAAO,CAAA;IACpC,MAAM,IAAI,GAAG;QACX,MAAM;QACN,OAAO;QACP,OAAO,CAAC,SAAS;QACjB,qBAAqB;QACrB,OAAO,CAAC,OAAO;QACf,QAAQ;QACR,SAAS;QACT,QAAQ;QACR,MAAM,CAAC,mBAAmB,CAAC;QAC3B,iBAAiB;QACjB,MAAM,CAAC,QAAQ,CAAC,cAAc,CAAC;QAC/B,gBAAgB;QAChB,MAAM,CAAC,QAAQ,CAAC,YAAY,CAAC;QAC7B,0BAA0B;QAC1B,MAAM,CAAC,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,oBAAoB,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC;QAC5D,+FAA+F;QAC/F,+FAA+F;QAC/F,8BAA8B;QAC9B,IAAI,CAAC,SAAS,CAAC,EAAE,cAAc,EAAE,QAAQ,CAAC,iBAAiB,EAAE,GAAG,QAAQ,CAAC,UAAU,EAAE,CAAC;KACvF,CAAA;IACD,IAAI,OAAO,CAAC,IAAI,CAAC,kBAAkB,KAAK,IAAI,EAAE,CAAC;QAC7C,IAAI,CAAC,IAAI,CAAC,yBAAyB,EAAE,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,kBAAkB,CAAC,CAAC,CAAA;IAC/E,CAAC;IACD,IAAI,QAAQ,CAAC,sBAAsB,KAAK,IAAI,EAAE,CAAC;QAC7C,IAAI,CAAC,IAAI,CAAC,0BAA0B,EAAE,MAAM,CAAC,QAAQ,CAAC,sBAAsB,CAAC,CAAC,CAAA;IAChF,CAAC;IACD,IAAI,CAAC,QAAQ,CAAC,qBAAqB;QAAE,IAAI,CAAC,IAAI,CAAC,4BAA4B,CAAC,CAAA;IAC5E,IAAI,QAAQ,CAAC,cAAc,GAAG,CAAC;QAAE,IAAI,CAAC,IAAI,CAAC,kBAAkB,EAAE,MAAM,CAAC,QAAQ,CAAC,cAAc,CAAC,CAAC,CAAA;IAC/F,IAAI,QAAQ,CAAC,KAAK,KAAK,MAAM;QAAE,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,QAAQ,CAAC,KAAK,CAAC,CAAA;IACnE,IAAI,QAAQ,CAAC,IAAI,KAAK,IAAI;QAAE,IAAI,CAAC,IAAI,CAAC,QAAQ,EAAE,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAA;IACtE,IAAI,QAAQ,CAAC,gBAAgB;QAAE,IAAI,CAAC,IAAI,CAAC,oBAAoB,CAAC,CAAA;IAC9D,IAAI,OAAO,CAAC,IAAI,CAAC,UAAU;QAAE,IAAI,CAAC,IAAI,CAAC,uBAAuB,EAAE,IAAI,CAAC,SAAS,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC,CAAC,CAAA;IACvG,IAAI,KAAK,CAAC,OAAO,CAAC;QAAE,IAAI,CAAC,IAAI,CAAC,iBAAiB,CAAC,CAAA;IAChD,IAAI,QAAQ,CAAC,cAAc,KAAK,KAAK,IAAI,aAAa,CAAC,OAAO,CAAC,oBAAoB,CAAC,EAAE,CAAC;QACrF,IAAI,CAAC,IAAI,CAAC,kBAAkB,EAAE,KAAK,CAAC,CAAA;IACtC,CAAC;IACD,IAAI,MAAM,EAAE,WAAW;QAAE,IAAI,CAAC,IAAI,CAAC,2BAA2B,EAAE,oBAAoB,EAAE,MAAM,CAAC,WAAW,CAAC,CAAA;IACzG,IAAI,MAAM,EAAE,gBAAgB;QAAE,IAAI,CAAC,IAAI,CAAC,oBAAoB,EAAE,MAAM,CAAC,gBAAgB,CAAC,CAAA;IACtF,OAAO;QACL,MAAM,EAAE,EAAE,cAAc,EAAE,mBAAmB,EAAE;QAC/C,IAAI;QACJ,GAAG,EAAE;YACH,mBAAmB,EAAE,GAAG;YACxB,YAAY,EAAE,GAAG;YACjB,cAAc,EAAE,GAAG;YACnB,oBAAoB,EAAE,GAAG;YACzB,GAAG,QAAQ,CAAC,OAAO,CAAC,eAAe,CAAC;SACrC;KACF,CAAA;AACH,CAAC;AAED,oGAAoG;AACpG,MAAM,CAAC,MAAM,sBAAsB,GAAG,EAAE,GAAG,MAAM,CAAA;AACjD,MAAM,CAAC,MAAM,yBAAyB,GAAG,MAAM,CAAA;AAE/C,MAAM,UAAU,sBAAsB,CAAC,WAAmB,EAAE,QAAsB;IAChF,IAAI,QAAQ,CAAC,oBAAoB,KAAK,IAAI;QAAE,OAAO,QAAQ,CAAC,oBAAoB,GAAG,IAAI,CAAA;IACvF,OAAO,IAAI,CAAC,IAAI,CAAC,sBAAsB,GAAG,CAAC,WAAW,GAAG,GAAG,CAAC,GAAG,yBAAyB,CAAC,CAAA;AAC5F,CAAC;AAED,gGAAgG;AAChG,sBAAsB;AACtB,gGAAgG;AAEhG,MAAM,GAAG,GACP,iHAAiH,CAAA;AACnH,4FAA4F;AAC5F,MAAM,gBAAgB,GACpB,iJAAiJ,CAAA;AACnJ,MAAM,YAAY,GAChB,oNAAoN,CAAA;AACtN,MAAM,wBAAwB,GAAG,4DAA4D,CAAA;AAC7F,MAAM,OAAO,GAAG,gBAAgB,CAAA;AAEhC,MAAM,KAAK,GAAG,CAAC,MAAc,EAAE,IAAY,EAAU,EAAE,CACrD,IAAI,KAAK,KAAK;IACZ,CAAC,CAAC,MAAM;IACR,CAAC,CAAC,IAAI,KAAK,KAAK;QACd,CAAC,CAAC,MAAM,GAAG,IAAI;QACf,CAAC,CAAC,IAAI,KAAK,KAAK;YACd,CAAC,CAAC,MAAM,GAAG,IAAI,IAAI,CAAC;YACpB,CAAC,CAAC,MAAM,GAAG,GAAG,CAAA;AACtB,MAAM,KAAK,GAAG,CAAC,KAAa,EAAE,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,GAAG,GAAG,CAAC,GAAG,GAAG,CAAA;AAE9D,MAAM,UAAU,gBAAgB,CAAC,GAAW,EAAE,QAAuB;IACnE,MAAM,EAAE,GAAG,YAAY,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;IACjC,IAAI,EAAE,KAAK,IAAI,EAAE,CAAC;QAChB,MAAM,WAAW,GAAG,MAAM,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAA;QACjC,MAAM,IAAI,GAAG,EAAE,CAAC,CAAC,CAAC,KAAK,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAA;QACvD,OAAO;YACL,IAAI,EAAE,eAAe;YACrB,OAAO,EACL,sBAAsB,WAAW,wCAAwC,EAAE,CAAC,CAAC,CAAC,SAAS,EAAE,CAAC,CAAC,CAAC,mBAAmB;gBAC/G,CAAC,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,iBAAiB,IAAI,uDAAuD,CAAC;YACtG,OAAO,EAAE;gBACP,aAAa,EAAE,WAAW;gBAC1B,aAAa,EAAE,MAAM,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;gBAC5B,gBAAgB,EAAE,MAAM,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;gBAC/B,GAAG,CAAC,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,eAAe,EAAE,IAAI,EAAE,CAAC;aACpD;YACD,OAAO,EAAE,EAAE,CAAC,CAAC,CAAC;SACf,CAAA;IACH,CAAC;IACD,MAAM,IAAI,GAAG,gBAAgB,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;IACvC,IAAI,IAAI,KAAK,IAAI,EAAE,CAAC;QAClB,OAAO;YACL,IAAI,EAAE,eAAe;YACrB,OAAO,EACL,eAAe,IAAI,CAAC,CAAC,CAAC,WAAW,IAAI,CAAC,CAAC,CAAC,kDAAkD,IAAI,CAAC,CAAC,CAAC,OAAO;gBACxG,6GAA6G;YAC/G,OAAO,EAAE;gBACP,QAAQ,EAAE,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;gBACzB,SAAS,EAAE,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;gBAC1B,WAAW,EAAE,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;gBAC5B,aAAa,EAAE,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;aAC/B;YACD,OAAO,EAAE,IAAI,CAAC,CAAC,CAAC;SACjB,CAAA;IACH,CAAC;IACD,MAAM,GAAG,GAAG,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;IACzB,IAAI,GAAG,KAAK,IAAI,EAAE,CAAC;QACjB,MAAM,KAAK,GAAG,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,CAAW,CAAC,CAAC,CAAA;QAC5D,MAAM,IAAI,GAAG,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,CAAW,CAAC,CAAC,CAAA;QAC3D,OAAO;YACL,IAAI,EAAE,eAAe;YACrB,OAAO,EAAE,qDAAqD,GAAG,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,CAAC,SAAS,GAAG,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,CAAC,QAAQ;YAC/G,OAAO,EAAE,EAAE,aAAa,EAAE,KAAK,EAAE,QAAQ,EAAE,IAAI,EAAE;YACjD,OAAO,EAAE,GAAG,CAAC,CAAC,CAAC;SAChB,CAAA;IACH,CAAC;IACD,MAAM,IAAI,GAAG,wBAAwB,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;IAC/C,IAAI,IAAI,KAAK,IAAI,EAAE,CAAC;QAClB,OAAO;YACL,IAAI,EAAE,mBAAmB;YACzB,OAAO,EAAE,0DAA0D,IAAI,CAAC,CAAC,CAAC,EAAE,OAAO,CAAC,IAAI,EAAE,EAAE,CAAC,GAAG;YAChG,OAAO,EAAE,IAAI,CAAC,CAAC,CAAC;SACjB,CAAA;IACH,CAAC;IACD,IAAI,OAAO,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC;QACtB,OAAO,EAAE,IAAI,EAAE,eAAe,EAAE,OAAO,EAAE,+CAA+C,EAAE,CAAA;IAC5F,CAAC;IACD,OAAO;QACL,IAAI,EAAE,OAAO;QACb,OAAO,EAAE,cAAc,QAAQ,KAAK,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,cAAc,QAAQ,EAAE,uBAAuB;KAChG,CAAA;AACH,CAAC;AAED,gGAAgG;AAChG,sDAAsD;AACtD,gGAAgG;AAEhG,MAAM,UAAU,gBAAgB,CAAC,OAGhC;IACC,MAAM,MAAM,GAAG,OAAO,CAAC,MAAM,CAAA;IAC7B,OAAO;QACL,KAAK,EAAE,OAAO,CAAC,MAAM,EAAE,WAAW,CAAC;QACnC,SAAS,EAAE,OAAO,CAAC,MAAM,EAAE,gBAAgB,CAAC;QAC5C,iBAAiB,EAAE,MAAM,EAAE,iBAAiB,KAAK,IAAI;QACrD,MAAM,EAAE,KAAK;QACb,UAAU,EAAE,KAAK;QACjB,SAAS,EAAE,KAAK;KACjB,CAAA;AACH,CAAC;AAED,MAAM,CAAC,MAAM,WAAW,GAA4B;IAClD,EAAE,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,sBAAsB,EAAE;IAChD,EAAE,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,iBAAiB,EAAE;IAC3C,EAAE,MAAM,EAAE,KAAK,EAAE,IAAI,EAAE,YAAY,EAAE;CACtC,CAAA;AAED,MAAM,CAAC,MAAM,sBAAsB,GAA4B;IAC7D,EAAE,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,sBAAsB,EAAE;IAChD,EAAE,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,iBAAiB,EAAE;CAC5C,CAAA;AAED,MAAM,UAAU,GAAG,IAAI,GAAG,CAAC,sBAAsB,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAA;AAE7E,SAAS,QAAQ,CAAC,IAA6B,EAAE,GAAW;IAC1D,IAAI,CAAC,CAAC,GAAG,IAAI,IAAI,CAAC,IAAI,IAAI,CAAC,GAAG,CAAC,KAAK,IAAI;QAAE,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC,EAAE,CAAA;IAC7E,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,CAAA;IACvB,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,IAAI,KAAK,IAAI,CAAC,EAAE,CAAC;QACxE,MAAM,IAAI,eAAe,CAAC,kBAAkB,EAAE,GAAG,GAAG,8BAA8B,CAAC,CAAA;IACrF,CAAC;IACD,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,KAAK,EAAE,CAAA;AACjC,CAAC;AAED,4GAA4G;AAC5G,MAAM,UAAU,GAAG,IAAI,GAAG,CAAC;IACzB,WAAW;IACX,aAAa;IACb,OAAO;IACP,WAAW;IACX,OAAO;IACP,aAAa;IACb,WAAW;CACZ,CAAC,CAAA;AAEF,SAAS,YAAY,CAAC,IAA6B;IACjD,MAAM,QAAQ,GAAG,IAAI,CAAC,UAAU,CAAC,CAAA;IACjC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,QAAQ,CAAC;QAAE,OAAO,KAAK,CAAA;IAC1C,OAAO,QAAQ,CAAC,IAAI,CAAC,CAAC,OAAO,EAAE,EAAE;QAC/B,MAAM,OAAO,GACX,OAAO,KAAK,IAAI,IAAI,OAAO,OAAO,KAAK,QAAQ,CAAC,CAAC,CAAE,OAAiC,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI,CAAA;QACrG,OAAO,CACL,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC;YACtB,OAAO,CAAC,IAAI,CACV,CAAC,IAAI,EAAE,EAAE,CACP,IAAI,KAAK,IAAI;gBACb,OAAO,IAAI,KAAK,QAAQ;gBACxB,UAAU,CAAC,GAAG,CAAC,MAAM,CAAE,IAA2B,CAAC,IAAI,CAAC,CAAC,CAC5D,CACF,CAAA;IACH,CAAC,CAAC,CAAA;AACJ,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,sBAAsB,CACpC,KAAa,EACb,IAAa,EACb,QAAsB,EACtB,YAAsC;IAEtC,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,KAAK,CAAC;QAAE,OAAO,IAAI,CAAA;IACvC,IAAI,OAAO,IAAI,KAAK,QAAQ,IAAI,IAAI,KAAK,IAAI,IAAI,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC;QAAE,OAAO,IAAI,CAAA;IACjF,MAAM,MAAM,GAAG,IAA+B,CAAA;IAC9C,MAAM,KAAK,GAAG,OAAO,MAAM,CAAC,OAAO,CAAC,KAAK,QAAQ,CAAC,CAAC,CAAC,cAAc,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,YAAY,CAAA;IACnG,IAAI,YAAY,CAAC,MAAM,CAAC,EAAE,CAAC;QACzB,MAAM,IAAI,qBAAqB,CAAC,GAAG,KAAK,2BAA2B,EAAE,wBAAwB,CAAC,CAAA;IAChG,CAAC;IACD,IAAI,YAAY,KAAK,SAAS,EAAE,CAAC;QAC/B,IAAI,CAAC,YAAY,CAAC,KAAK,IAAI,YAAY,CAAC,MAAM,CAAC,EAAE,CAAC;YAChD,MAAM,IAAI,qBAAqB,CAAC,GAAG,KAAK,iCAAiC,EAAE,wBAAwB,CAAC,CAAA;QACtG,CAAC;QACD,IAAI,CAAC,YAAY,CAAC,iBAAiB,IAAI,uBAAuB,CAAC,MAAM,CAAC,EAAE,CAAC;YACvE,MAAM,IAAI,qBAAqB,CAC7B,GAAG,KAAK,sCAAsC,EAC9C,wBAAwB,CACzB,CAAA;QACH,CAAC;IACH,CAAC;IACD,MAAM,GAAG,GAAG,QAAQ,CAAC,iBAAiB,CAAA;IACtC,IAAI,KAAK,KAAK,iBAAiB,EAAE,CAAC;QAChC,MAAM,KAAK,GAAG,QAAQ,CAAC,MAAM,EAAE,YAAY,CAAC,CAAA;QAC5C,OAAO,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,GAAG,MAAM,EAAE,UAAU,EAAE,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,KAAK,EAAE,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,MAAM,CAAA;IACvF,CAAC;IACD,MAAM,MAAM,GAAG,QAAQ,CAAC,MAAM,EAAE,YAAY,CAAC,CAAA;IAC7C,MAAM,MAAM,GAAG,QAAQ,CAAC,MAAM,EAAE,uBAAuB,CAAC,CAAA;IACxD,MAAM,KAAK,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,KAAK,CAAC,CAAA;IAC3F,uGAAuG;IACvG,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,MAAM,CAAA;IACrC,MAAM,IAAI,GAA4B,EAAE,GAAG,MAAM,EAAE,CAAA;IACnD,OAAO,IAAI,CAAC,YAAY,CAAC,CAAA;IACzB,OAAO,IAAI,CAAC,uBAAuB,CAAC,CAAA;IACpC,IAAI,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,uBAAuB,CAAC,CAAC,CAAC,YAAY,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,GAAG,KAAK,CAAC,CAAA;IACvF,OAAO,IAAI,CAAA;AACb,CAAC;AAED,4GAA4G;AAC5G,MAAM,QAAQ,GAAG;IACf,uFAAuF;IACvF,8EAA8E;IAC9E,gFAAgF;CACjF,CAAA;AAED,SAAS,YAAY,CAAC,IAAY;IAChC,IAAI,CAAC;QACH,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAyD,CAAA;QACvF,MAAM,OAAO,GAAG,MAAM,CAAC,KAAK,EAAE,OAAO,IAAI,MAAM,CAAC,OAAO,CAAA;QACvD,OAAO,OAAO,OAAO,KAAK,QAAQ,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAA;IACnD,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,EAAE,CAAA;IACX,CAAC;AACH,CAAC;AAED,qGAAqG;AACrG,MAAM,UAAU,yBAAyB,CAAC,MAAc,EAAE,IAAY;IACpE,IAAI,MAAM,KAAK,GAAG;QAAE,OAAO,IAAI,CAAA;IAC/B,MAAM,OAAO,GAAG,YAAY,CAAC,IAAI,CAAC,CAAA;IAClC,KAAK,MAAM,OAAO,IAAI,QAAQ,EAAE,CAAC;QAC/B,MAAM,KAAK,GAAG,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,CAAA;QACnC,IAAI,KAAK,KAAK,IAAI;YAAE,SAAQ;QAC5B,MAAM,KAAK,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAA;QAC9B,MAAM,SAAS,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAA;QAClC,OAAO;YACL,KAAK,EAAE;gBACL,OAAO,EAAE,0CAA0C,KAAK,gCAAgC,SAAS,4DAA4D;gBAC7J,IAAI,EAAE,uBAAuB;gBAC7B,KAAK,EAAE,UAAU;gBACjB,IAAI,EAAE,yBAAyB;aAChC;SACF,CAAA;IACH,CAAC;IACD,OAAO,IAAI,CAAA;AACb,CAAC;AAED;;;GAGG;AACH,MAAM,CAAC,MAAM,+BAA+B,GAAwB,IAAI,GAAG,CAAC,CAAC,OAAO,CAAC,CAAC,CAAA;AAEtF,MAAM,UAAU,sBAAsB,CACpC,KAAa,EACb,WAAoB,EACpB,MAAiC;IAEjC,IAAI,KAAK,KAAK,sBAAsB;QAAE,OAAO,IAAI,CAAA;IACjD,MAAM,MAAM,GAAG,MAAM,EAAE,gBAAgB,IAAI,IAAI,CAAA;IAC/C,IAAI,MAAM,KAAK,IAAI,IAAI,CAAC,+BAA+B,CAAC,GAAG,CAAC,MAAM,CAAC;QAAE,OAAO,IAAI,CAAA;IAChF,OAAO,iBAAiB,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,oBAAoB,CAAA;AACrE,CAAC;AAED;;;;GAIG;AACH,MAAM,aAAa,GAAkC;IACnD,EAAE,KAAK,EAAE,qBAAqB,EAAE,OAAO,EAAE,wBAAwB,EAAE;IACnE,EAAE,KAAK,EAAE,qBAAqB,EAAE,OAAO,EAAE,uCAAuC,EAAE;IAClF,EAAE,KAAK,EAAE,qBAAqB,EAAE,OAAO,EAAE,oBAAoB,EAAE;IAC/D,EAAE,KAAK,EAAE,qBAAqB,EAAE,OAAO,EAAE,qCAAqC,EAAE;IAChF,EAAE,KAAK,EAAE,qBAAqB,EAAE,OAAO,EAAE,2BAA2B,EAAE;IACtE,EAAE,KAAK,EAAE,qBAAqB,EAAE,OAAO,EAAE,uBAAuB,EAAE;CACnE,CAAA;AAED,MAAM,CAAC,MAAM,WAAW,GAAqC;IAC3D,EAAE,EAAE,MAAM;IACV,eAAe,EAAE,qCAAqC;IACtD,SAAS,EAAE,EAAE,IAAI,EAAE,SAAS,EAAE,cAAc,EAAE,GAAG,EAAE;IACnD,MAAM,EAAE,WAAW;IACnB,gBAAgB,EAAE,sBAAsB;IACxC,YAAY,EAAE,aAAa;IAC3B,gBAAgB,EAAE,CAAC,GAAG,EAAE,EAAE,CAAC,oBAAoB,CAAC,CAAC,GAAG,IAAI,EAAE,CAA4B,CAAC;IACvF,WAAW,EAAE,eAAe;IAC5B,kBAAkB,EAAE,sBAAsB;IAC1C,YAAY,EAAE,gBAAgB;IAC9B,YAAY,EAAE,gBAAgB;IAC9B,kBAAkB,EAAE,sBAAsB;IAC1C,kBAAkB,EAAE,sBAAsB;IAC1C,gBAAgB,EAAE,CAAC,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,EAAE,CAAC,yBAAyB,CAAC,MAAM,EAAE,IAAI,CAAC;IACnF,mGAAmG;IACnG,2CAA2C;IAC3C,UAAU,EAAE,CAAC,QAAQ,EAAE,EAAE,CAAC;QACxB,QAAQ,CAAC,cAAc;QACvB,QAAQ,CAAC,iBAAiB;QAC1B,QAAQ,CAAC,YAAY;QACrB,QAAQ,CAAC,mBAAmB;QAC5B,QAAQ,CAAC,sBAAsB;QAC/B,QAAQ,CAAC,WAAW;QACpB,QAAQ,CAAC,cAAc;QACvB,QAAQ,CAAC,sBAAsB;QAC/B,QAAQ,CAAC,qBAAqB;QAC9B,QAAQ,CAAC,cAAc;QACvB,QAAQ,CAAC,KAAK;QACd,QAAQ,CAAC,IAAI;QACb,QAAQ,CAAC,gBAAgB;QACzB,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,UAAU,CAAC;KACpC;CACF,CAAA"}

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

@@ -0,0 +1,16 @@
+import type { ManagedEngineSpec } from '../managed-engines/spec.js';
+import type { ManagedTextCapabilities } from '../managed-text/index.js';
+import type { SessionRoutePolicy } from '../shared/index.js';
+import { type VllmSettings } from './settings.js';
+export declare const VLLM_ENGINE_ID = "vllm";
+/**
+ * What the public server (`:1337`) must know about a `vllm` session to route it honestly: only the
+ * routes the adapter declares, tool calls and JSON output only as the session can, and vLLM's context
+ * overflow answered as `context_length_exceeded`. `null` capabilities (a session the core did not
+ * start) keep the routes and the mapping and leave tools and JSON output to the engine.
+ */
+export declare function vllmRoutePolicy(capabilities: ManagedTextCapabilities | null, limits?: {
+    contextLength: number;
+    maxOutputTokens: number;
+} | null): SessionRoutePolicy;
+export declare const VLLM_ENGINE: ManagedEngineSpec<VllmSettings>;

package/dist/runtime/vllm/engine.js

@@ -0,0 +1,37 @@
+/**
+ * vLLM as a managed engine (change `add-vllm-runtime`, task 3.4, design D3): its adapter, settings,
+ * memory rule and launch plan, route policy, and its descriptor in conf (`runtimes/vllm.json`, read
+ * only once conf main publishes it — until then the provider is unsupported, design D15).
+ */
+import { defaultDescriptorUrl } from '../environment/index.js';
+import { VLLM_ROUTES, mapVllmContextLengthError, vllmAdapter } from './adapter.js';
+import { vllmCheckEngine, vllmLaunchPlan } from './memory.js';
+import { vllmSettings } from './settings.js';
+export const VLLM_ENGINE_ID = 'vllm';
+/**
+ * What the public server (`:1337`) must know about a `vllm` session to route it honestly: only the
+ * routes the adapter declares, tool calls and JSON output only as the session can, and vLLM's context
+ * overflow answered as `context_length_exceeded`. `null` capabilities (a session the core did not
+ * start) keep the routes and the mapping and leave tools and JSON output to the engine.
+ */
+export function vllmRoutePolicy(capabilities, limits = null) {
+    return {
+        routes: VLLM_ROUTES,
+        tools: capabilities?.tools ?? true,
+        structuredOutput: capabilities?.structured_output ?? true,
+        mapError: mapVllmContextLengthError,
+        ...(limits === null ? {} : limits),
+    };
+}
+export const VLLM_ENGINE = {
+    engine_id: VLLM_ENGINE_ID,
+    provider: VLLM_ENGINE_ID,
+    label: 'vLLM',
+    descriptor: { engine_id: VLLM_ENGINE_ID, label: 'vLLM', url: defaultDescriptorUrl(VLLM_ENGINE_ID) },
+    adapter: vllmAdapter,
+    settings: vllmSettings,
+    check: vllmCheckEngine,
+    launchPlan: vllmLaunchPlan,
+    routePolicy: vllmRoutePolicy,
+};
+//# sourceMappingURL=engine.js.map

package/dist/runtime/vllm/engine.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"engine.js","sourceRoot":"","sources":["../../../src/runtime/vllm/engine.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AACH,OAAO,EAAE,oBAAoB,EAAE,MAAM,yBAAyB,CAAA;AAI9D,OAAO,EAAE,WAAW,EAAE,yBAAyB,EAAE,WAAW,EAAE,MAAM,cAAc,CAAA;AAClF,OAAO,EAAE,eAAe,EAAE,cAAc,EAAE,MAAM,aAAa,CAAA;AAC7D,OAAO,EAAE,YAAY,EAAqB,MAAM,eAAe,CAAA;AAE/D,MAAM,CAAC,MAAM,cAAc,GAAG,MAAM,CAAA;AAEpC;;;;;GAKG;AACH,MAAM,UAAU,eAAe,CAC7B,YAA4C,EAC5C,SAAoE,IAAI;IAExE,OAAO;QACL,MAAM,EAAE,WAAW;QACnB,KAAK,EAAE,YAAY,EAAE,KAAK,IAAI,IAAI;QAClC,gBAAgB,EAAE,YAAY,EAAE,iBAAiB,IAAI,IAAI;QACzD,QAAQ,EAAE,yBAAyB;QACnC,GAAG,CAAC,MAAM,KAAK,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,MAAM,CAAC;KACnC,CAAA;AACH,CAAC;AAED,MAAM,CAAC,MAAM,WAAW,GAAoC;IAC1D,SAAS,EAAE,cAAc;IACzB,QAAQ,EAAE,cAAc;IACxB,KAAK,EAAE,MAAM;IACb,UAAU,EAAE,EAAE,SAAS,EAAE,cAAc,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,oBAAoB,CAAC,cAAc,CAAC,EAAE;IACnG,OAAO,EAAE,WAAW;IACpB,QAAQ,EAAE,YAAY;IACtB,KAAK,EAAE,eAAe;IACtB,UAAU,EAAE,cAAc;IAC1B,WAAW,EAAE,eAAe;CAC7B,CAAA"}

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

@@ -0,0 +1,11 @@
+/**
+ * The `vllm` provider (change `add-vllm-runtime`): its settings (`settings.ts`), the
+ * `ManagedTextAdapter` for `vllm serve` (`adapter.ts`), its memory model (`memory.ts`) and its spec in
+ * the managed engine registry (`engine.ts`).
+ *
+ * Public API of this module is exported from this file only.
+ */
+export * from './settings.js';
+export * from './adapter.js';
+export * from './memory.js';
+export * from './engine.js';

package/dist/runtime/vllm/index.js

@@ -0,0 +1,12 @@
+/**
+ * The `vllm` provider (change `add-vllm-runtime`): its settings (`settings.ts`), the
+ * `ManagedTextAdapter` for `vllm serve` (`adapter.ts`), its memory model (`memory.ts`) and its spec in
+ * the managed engine registry (`engine.ts`).
+ *
+ * Public API of this module is exported from this file only.
+ */
+export * from './settings.js';
+export * from './adapter.js';
+export * from './memory.js';
+export * from './engine.js';
+//# sourceMappingURL=index.js.map

package/dist/runtime/vllm/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../src/runtime/vllm/index.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AACH,cAAc,eAAe,CAAA;AAC7B,cAAc,cAAc,CAAA;AAC5B,cAAc,aAAa,CAAA;AAC3B,cAAc,aAAa,CAAA"}

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

@@ -0,0 +1,69 @@
+/**
+ * vLLM's memory, decided by core (change `add-vllm-runtime`, task 3.3, design D9; spec `vllm-runtime`,
+ * "Память vLLM задаёт core"). vLLM takes a share of the card's *total* memory (`--gpu-memory-
+ * utilization`, 0.92 by default) and, before it starts, refuses to run when the card's free memory —
+ * measured after its own CUDA context exists — is below that share; so on any desktop whose card holds
+ * a display and a browser, its default fails at start. Core instead:
+ *
+ * - gives vLLM (unless the memory share setting fixes it) the share of the card that is free right
+ *   before the container is created, less a margin larger than vLLM's own CUDA context: `min(0.95, (free − 1.5 GiB) / total)` (on a card with
+ *   the host's memory, never more than half of it). Within that share vLLM places the weights and
+ *   activations and gives the rest to the KV cache, sized by vLLM itself — it knows every model's KV
+ *   shape, hybrid ones (Qwen3.5's linear attention) included. Core passes `--kv-cache-memory-bytes`
+ *   only when the KV cache size setting fixes it. (Core sized the KV cache in bytes at first, from the
+ *   attention layers only; for Qwen3.5 that was a third of what vLLM needs, and every load failed —
+ *   Windows acceptance, 2026-10-06. The 1.5 GiB margin comes from the same run: the context took 0.87
+ *   GiB under WSL on an RTX 4070 Laptop, more than the first 512 MiB.)
+ * - checks, before any container, that the weights, an estimate of the KV cache (context × concurrent
+ *   requests × the model's per-token KV bytes over its attention layers, or the KV cache size setting)
+ *   and vLLM's own overhead fit the free memory — the same rule the compatibility check
+ *   (`POST /models/vllm/check`) applies. It is an estimate for the verdict and the refusal before a
+ *   container; it no longer sizes anything vLLM runs with.
+ *
+ * The overhead (~2 GiB: CUDA context, activations, sampler, graphs) and the margins are first
+ * estimates; the live run of task 6.1 records the numbers (`summary.json`).
+ */
+import type { GpuFacts } from '../../contracts/index.js';
+import { type CheckpointMemoryInputs, type HostMemory, type JsonObject, type KvReserveBasis, type ManagedCheckEngine, type MemoryNeed } from '../managed-models/index.js';
+import type { VllmLaunchPlan } from './adapter.js';
+import type { VllmSettings } from './settings.js';
+/** What vLLM holds beyond weights and KV cache: CUDA context, activations, sampler, graphs. */
+export declare const VLLM_ENGINE_OVERHEAD_BYTES: number;
+/** Counted as needed in the memory check: what other processes may still take meanwhile. */
+export declare const VLLM_FREE_MEMORY_MARGIN_BYTES: number;
+/**
+ * Left out of the share vLLM is given: more than vLLM's own CUDA context, which already holds memory
+ * when vLLM checks the share at start (0.87 GiB under WSL on an RTX 4070 Laptop, 2026-10-06).
+ */
+export declare const VLLM_START_CHECK_MARGIN_BYTES: number;
+/** The highest share of a card vLLM is ever given. */
+export declare const VLLM_MAX_GPU_MEMORY_UTILIZATION = 0.95;
+/** On a card with the host's memory: contexts the KV cache holds, as for TensorRT-LLM (design D9). */
+export declare const VLLM_UNIFIED_KV_CONTEXTS = 2;
+/** The tokens the KV cache estimate covers: context × requests (two contexts on unified memory). */
+export declare function vllmKvTokens(settings: VllmSettings, unifiedMemory: boolean): number;
+/**
+ * The KV cache the memory check counts: the KV cache size setting when set, else an estimate from the
+ * model's attention layers — and whether the model's shape sized it. vLLM sizes its own KV cache
+ * unless the setting fixes it, so this only feeds the check's verdict.
+ */
+export declare function vllmKvCacheBytes(checkpoint: {
+    configJson: JsonObject;
+    hfQuantConfigJson: JsonObject | null;
+    weightBytesTotal?: number;
+}, settings: VllmSettings, gpu: GpuFacts, unifiedMemory: boolean): {
+    bytes: number;
+    basis: KvReserveBasis;
+};
+/**
+ * vLLM's memory rule on `gpu`: weights on the card + KV cache in bytes + its own overhead (CUDA context
+ * included) + a margin for what other processes may take meanwhile; on a card with the host's memory,
+ * whatever of `MemAvailable` lies beyond half the host's memory counts as needed too.
+ */
+export declare function vllmMemoryNeed(gpu: GpuFacts, checkpoint: CheckpointMemoryInputs, settings: VllmSettings, host: HostMemory): MemoryNeed;
+/** `--gpu-memory-utilization` from the card as it stands now (see the file banner). */
+export declare function vllmGpuMemoryUtilization(gpu: GpuFacts, host: HostMemory): number;
+/** The launch plan `beforeCreate` hands the adapter, from the card re-probed right before the container. */
+export declare function vllmLaunchPlan(settings: VllmSettings, checkpoint: CheckpointMemoryInputs, gpu: GpuFacts, host: HostMemory): VllmLaunchPlan;
+/** The check skeleton's hooks for vLLM: its memory rule, and no checkpoint quirks of its own. */
+export declare function vllmCheckEngine(settings: VllmSettings): ManagedCheckEngine;