@atomic-chat/core 0.13.6 → 0.18.3

package/dist/runtime/sglang/adapter.js

@@ -0,0 +1,393 @@
+/**
+ * The `sglang` engine's `ManagedTextAdapter` (change `add-sglang-runtime`, task 2.2, design D3, D5;
+ * spec `sglang-runtime`): how core runs `python3 -m sglang.launch_server` from the `lmsysorg/sglang`
+ * image on the shared managed-text lifecycle. Everything here follows SGLang v0.5.21's source (the
+ * rulings of task 2.2 in the change's `rulings/core.md` say where).
+ *
+ * - **argv**: `--model-path <model> --served-model-name <id> --host 0.0.0.0 --port 30000
+ *   --context-length <ctx> --max-running-requests <n> --mem-fraction-static <f>`, then only what
+ *   differs from SGLang's own: the KV cache limit, the prefill chunk, `--disable-radix-cache`,
+ *   `--dtype`, `--random-seed`, the CUDA graph mode (D5), `--kv-cache-dtype fp8_e4m3` on compute
+ *   capability 8.9 and newer, the family's parsers, and `--json-model-override-args
+ *   {"language_model_only":true}` for a multimodal checkpoint. `f` is core's (`plan`, from the card's
+ *   free memory right before the container, design D4); within it SGLang sizes the KV cache itself.
+ *   Never `--trust-remote-code` (no code from a model repository runs), `--log-requests` (prompts would
+ *   land in the logs core serves), `--api-key`/`--admin-api-key` (the session gateway checks the key)
+ *   or `--enable-metrics`.
+ * - **env**: Hugging Face offline (every file is already in the store), no tracking, the output limit
+ *   for a request that names none (`SGLANG_MAX_NEW_TOKENS_LIMIT`), and every compile cache — SGLang's
+ *   own and its JIT kernels', Triton's, inductor's, FlashInfer's, CUDA's, TileLang's — under the engine
+ *   cache, so a second start of the same model reuses the first one's work.
+ * - **readiness** `GET /health`, then the shared generation probe; **routes** chat completions,
+ *   completions and the model list; requests capped at the output setting, given the sampling defaults
+ *   they do not set, refused for tools without a parser, structured output the family does not declare,
+ *   and images; SGLang's context overflow answered as OpenAI's `context_length_exceeded` with both
+ *   numbers.
+ *
+ * Log lines, exit wording and error bodies follow SGLang v0.5.21's source
+ * (`test/helpers/sglang-log-fixtures.ts`); the live run of task 4.1 confirms or replaces them.
+ */
+import { AtomicCoreError } from '../../contracts/index.js';
+import { supportsFp8KvCache } from '../managed-models/index.js';
+import { withGenerationDefaults } from '../managed-text/generation-defaults.js';
+import { MANAGED_TEXT_ADAPTER_CONTRACT_VERSION, ManagedRequestRefusal, asksForStructuredOutput, asksForTools, capRequestedOutput, carriesMedia, reasoningIntoContent, thinkingRequested, } from '../managed-text/index.js';
+import { validateSglangSettings } from './settings.js';
+const GiB = 1024 ** 3;
+/** The port `sglang.launch_server` listens on inside its container (SGLang's own default). */
+export const SGLANG_CONTAINER_PORT = 30000;
+function isPlan(value) {
+    if (value === null || typeof value !== 'object')
+        return false;
+    const plan = value;
+    return (typeof plan['memFractionStatic'] === 'number' &&
+        plan['memFractionStatic'] > 0 &&
+        plan['memFractionStatic'] < 1 &&
+        typeof plan['multimodal'] === 'boolean');
+}
+/**
+ * Design D5 in SGLang v0.5.21's flags: `auto` captures decode graphs only up to the requests served at
+ * once (`--cuda-graph-max-bs-decode`), so they cost little memory; `on` is SGLang's own range; `off`
+ * disables both phases. `--disable-cuda-graph` is deprecated in v0.5.21 and never used.
+ */
+function cudaGraphArgs(settings) {
+    switch (settings.cuda_graphs) {
+        case 'auto':
+            return ['--cuda-graph-max-bs-decode', String(settings.max_running_requests)];
+        case 'on':
+            return [];
+        case 'off':
+            return ['--disable-decode-cuda-graph', '--disable-prefill-cuda-graph'];
+    }
+}
+/**
+ * Every compile cache of the engine, under the engine cache (`ManagedLaunchContext.engineCachePath`).
+ * SGLang points the third-party ones at `SGLANG_CACHE_DIR` itself only when nothing imported them yet,
+ * and its JIT kernels at `~/.cache/sglang/jit` whatever that says, so each is set here.
+ */
+function cacheEnv(cache) {
+    return {
+        SGLANG_CACHE_DIR: `${cache}/sglang`,
+        SGLANG_JIT_CACHE_DIR: `${cache}/sglang/jit`,
+        TRITON_CACHE_DIR: `${cache}/triton`,
+        TORCHINDUCTOR_CACHE_DIR: `${cache}/inductor`,
+        FLASHINFER_WORKSPACE_BASE: `${cache}/flashinfer`,
+        CUDA_CACHE_PATH: `${cache}/nv`,
+        TILELANG_CACHE_DIR: `${cache}/tilelang`,
+    };
+}
+export function buildSglangLaunch(context) {
+    if (!isPlan(context.plan)) {
+        throw new AtomicCoreError('INVALID_ARGUMENT', 'An SGLang launch needs core’s memory plan (its static memory share), computed right before the container is created.');
+    }
+    const { settings, family } = context;
+    const argv = [
+        'python3',
+        '-m',
+        'sglang.launch_server',
+        '--model-path',
+        context.modelPath,
+        '--served-model-name',
+        context.modelId,
+        '--host',
+        '0.0.0.0',
+        '--port',
+        String(SGLANG_CONTAINER_PORT),
+        '--context-length',
+        String(settings.context_length),
+        '--max-running-requests',
+        String(settings.max_running_requests),
+        '--mem-fraction-static',
+        String(Number(context.plan.memFractionStatic.toFixed(4))),
+    ];
+    if (settings.max_total_tokens !== null)
+        argv.push('--max-total-tokens', String(settings.max_total_tokens));
+    if (settings.chunked_prefill_size !== null) {
+        argv.push('--chunked-prefill-size', String(settings.chunked_prefill_size));
+    }
+    if (!settings.enable_prefix_caching)
+        argv.push('--disable-radix-cache');
+    if (settings.dtype !== 'auto')
+        argv.push('--dtype', settings.dtype);
+    if (settings.seed !== null)
+        argv.push('--random-seed', String(settings.seed));
+    argv.push(...cudaGraphArgs(settings));
+    if (settings.kv_cache_dtype === 'fp8' && supportsFp8KvCache(context.gpuComputeCapability)) {
+        argv.push('--kv-cache-dtype', 'fp8_e4m3');
+    }
+    // The checkpoint's own "text only" declaration, which every class of the descriptor honours (no
+    // multimodal reservation, media refused); SGLang's --language-model-only flag accepts only three
+    // architectures, none of them ours.
+    if (context.plan.multimodal) {
+        argv.push('--json-model-override-args', JSON.stringify({ language_model_only: true }));
+    }
+    if (family?.tool_parser)
+        argv.push('--tool-call-parser', family.tool_parser);
+    if (family?.reasoning_parser)
+        argv.push('--reasoning-parser', family.reasoning_parser);
+    return {
+        engine: { container_port: SGLANG_CONTAINER_PORT },
+        argv,
+        env: {
+            HF_HUB_OFFLINE: '1',
+            TRANSFORMERS_OFFLINE: '1',
+            DO_NOT_TRACK: '1',
+            // A chat request that names no output gets this, then whatever room the context has left:
+            // SGLang's scheduler clips rather than refuses, so a long prompt is not refused for the cap's
+            // room. (Completions keep OpenAI's default of 16, which SGLang applies before this.)
+            SGLANG_MAX_NEW_TOKENS_LIMIT: String(settings.max_output_tokens),
+            ...cacheEnv(context.engineCachePath),
+        },
+    };
+}
+/** A base for engine start-up, the first kernel compilation and graph capture; more per GiB of weights. */
+export const SGLANG_READINESS_BASE_MS = 10 * 60_000;
+export const SGLANG_READINESS_PER_GIB_MS = 20_000;
+export function sglangReadinessTimeoutMs(weightBytes, settings) {
+    if (settings.load_timeout_seconds !== null)
+        return settings.load_timeout_seconds * 1000;
+    return Math.ceil(SGLANG_READINESS_BASE_MS + (weightBytes / GiB) * SGLANG_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/;
+/** `mem_cache/kv_cache_configurator.py`: the weights took all of the static share. */
+const NO_KV_MEMORY = /Loaded weights leave no GPU memory for the KV cache under --mem-fraction-static=([\d.]+)\. Raise --mem-fraction-static above ([\d.]+)/;
+/** `model_executor/pool_configurator.py`: a KV pool of no tokens. */
+const EMPTY_POOL = /Not enough memory\. Please try to increase --mem-fraction-static\.(?: Current value: mem_fraction_static=([\d.]+))?/;
+/** `configs/model_config.py::_derive_context_length`: the setting is longer than the model's own context. */
+const CONTEXT_TOO_LONG = /context_length \((\d+)\) is greater than the derived context_length \((\d+)\)/;
+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 classifySglangExit(log, exitCode) {
+    const noKv = NO_KV_MEMORY.exec(log);
+    if (noKv !== null) {
+        return {
+            kind: 'out-of-memory',
+            message: `The model's weights left no GPU memory for the KV cache: SGLang kept a share of ${noKv[1]} of the ` +
+                `free memory and needed more than ${noKv[2]}. Free GPU memory, or choose a smaller model.`,
+            numbers: { mem_fraction_static: Number(noKv[1]), minimum_mem_fraction_static: Number(noKv[2]) },
+            excerpt: noKv[0],
+        };
+    }
+    const empty = EMPTY_POOL.exec(log);
+    if (empty !== null) {
+        return {
+            kind: 'out-of-memory',
+            message: `The KV cache got no room after the model's weights` +
+                (empty[1] === undefined ? '' : ` within SGLang's memory share of ${empty[1]}`) +
+                '. Free GPU memory, or choose a smaller model.',
+            numbers: empty[1] === undefined ? {} : { mem_fraction_static: Number(empty[1]) },
+            excerpt: empty[0],
+        };
+    }
+    const oom = OOM.exec(log);
+    if (oom !== null) {
+        return {
+            kind: 'out-of-memory',
+            message: `The GPU ran out of memory: SGLang tried to allocate ${oom[1]} ${oom[2]} with ${oom[3]} ${oom[4]} free.`,
+            numbers: {
+                requested_gib: round(toGiB(Number(oom[1]), oom[2])),
+                free_gib: round(toGiB(Number(oom[3]), oom[4])),
+            },
+            excerpt: oom[0],
+        };
+    }
+    const tooLong = CONTEXT_TOO_LONG.exec(log);
+    if (tooLong !== null) {
+        return {
+            kind: 'other',
+            message: `The context length ${tooLong[1]} is longer than the model supports (${tooLong[2]} tokens). ` +
+                `Lower the context length in the SGLang settings to ${tooLong[2]} or less.`,
+            numbers: { context_length: Number(tooLong[1]), model_context_length: Number(tooLong[2]) },
+            excerpt: tooLong[0],
+        };
+    }
+    const arch = UNSUPPORTED_ARCHITECTURE.exec(log);
+    if (arch !== null) {
+        return {
+            kind: 'unsupported-model',
+            message: `This SGLang release cannot run the model's architecture: ${arch[1]?.replace(/'/g, '')}.`,
+            excerpt: arch[0],
+        };
+    }
+    if (ANY_OOM.test(log)) {
+        return { kind: 'out-of-memory', message: 'The GPU ran out of memory while SGLang started.' };
+    }
+    return {
+        kind: 'other',
+        message: `SGLang exited${exitCode === null ? '' : ` with code ${exitCode}`} before it was ready.`,
+    };
+}
+// ---------------------------------------------------------------------------------------------
+// Capabilities, routes, request and response rewrites
+// ---------------------------------------------------------------------------------------------
+export function sglangCapabilities(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,
+    };
+}
+/** Everything else SGLang serves — `/server_info`, `/flush_cache`, `/metrics`, embeddings… — is the gateway's 404. */
+export const SGLANG_ROUTES = [
+    { method: 'POST', path: '/v1/chat/completions' },
+    { method: 'POST', path: '/v1/completions' },
+    { method: 'GET', path: '/v1/models' },
+];
+export const SGLANG_REWRITABLE_ROUTES = [
+    { method: 'POST', path: '/v1/chat/completions' },
+    { method: 'POST', path: '/v1/completions' },
+];
+const REWRITABLE = new Set(SGLANG_REWRITABLE_ROUTES.map((route) => route.path));
+/**
+ * Caps the output a request asks for at `max_output_tokens` (one key, as the client sent it); a chat
+ * request that asks for none is capped by `SGLANG_MAX_NEW_TOKENS_LIMIT` instead, within the room the
+ * context has left (a completions request that names none gets OpenAI's 16, SGLang's own default). Writes the sampling defaults a request does not set: SGLang's OpenAI path fills
+ * every sampling value before `--preferred-sampling-params` would apply. 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 sglangRewriteRequestBody(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');
+        }
+    }
+    return capRequestedOutput(route, withGenerationDefaults(client, settings.generation), settings.max_output_tokens);
+}
+/**
+ * SGLang's 400 wording for a request over the context (`managers/tokenizer_manager.py`), with the
+ * limit and the request: the input alone, and the input with the output it asked for.
+ */
+const OVERFLOW = [
+    {
+        pattern: /The input \((\d+) tokens\) is longer than the model's context length \((\d+) tokens\)/,
+        limit: 2,
+        requested: 1,
+    },
+    {
+        pattern: /maximum context length of (\d+) tokens\. You requested a total of (\d+) tokens/,
+        limit: 1,
+        requested: 2,
+    },
+];
+function errorMessage(body) {
+    try {
+        const parsed = JSON.parse(body);
+        const message = parsed.error?.message ?? parsed.message;
+        return typeof message === 'string' ? message : '';
+    }
+    catch {
+        return '';
+    }
+}
+/** SGLang's context overflow as OpenAI's `context_length_exceeded`; anything else relayed as it is. */
+export function mapSglangContextLengthError(status, body) {
+    if (status !== 400)
+        return null;
+    const message = errorMessage(body);
+    for (const { pattern, limit, requested } of OVERFLOW) {
+        const match = pattern.exec(message);
+        if (match === null)
+            continue;
+        return {
+            error: {
+                message: `This model's maximum context length is ${match[limit]} tokens, but the request has ${match[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 D3: the same
+ * hook as vLLM's). SGLang's `qwen3` parser splits only on the tags it sees when the request turned
+ * thinking off, so the hook is a no-op then; it moves what a model thought by its template's default.
+ */
+export const SGLANG_REASONING_AT_START_PARSERS = new Set(['qwen3']);
+export function sglangRewriteResponseFor(route, requestBody, family) {
+    if (route !== '/v1/chat/completions')
+        return null;
+    const parser = family?.reasoning_parser ?? null;
+    if (parser === null || !SGLANG_REASONING_AT_START_PARSERS.has(parser))
+        return null;
+    return thinkingRequested(requestBody) ? null : reasoningIntoContent;
+}
+/**
+ * `initializing-engine` from the first line that shows SGLang working on the model: reading weights,
+ * sizing the KV pool, capturing CUDA graphs. Before them the load stays in `starting-container`
+ * (Python, CUDA and the tokenizer and scheduler processes coming up).
+ */
+const STAGE_MARKERS = [
+    { stage: 'initializing-engine', pattern: /Load weight begin/ },
+    { stage: 'initializing-engine', pattern: /Load weight end/ },
+    { stage: 'initializing-engine', pattern: /Memory pool end/ },
+    { stage: 'initializing-engine', pattern: /Capture .* begin/ },
+    { stage: 'initializing-engine', pattern: /max_total_num_tokens=/ },
+];
+export const sglangAdapter = {
+    id: 'sglang',
+    contractVersion: MANAGED_TEXT_ADAPTER_CONTRACT_VERSION,
+    readiness: { path: '/health', expectedStatus: 200 },
+    // `/health` alone is not proof the engine can serve (spec: ready only after one generated token).
+    generationProbe: {
+        modelsPath: '/v1/models',
+        path: '/v1/chat/completions',
+        body: { messages: [{ role: 'user', content: 'Hi' }], max_tokens: 1, stream: false },
+    },
+    routes: SGLANG_ROUTES,
+    rewritableRoutes: SGLANG_REWRITABLE_ROUTES,
+    stageMarkers: STAGE_MARKERS,
+    validateSettings: (raw) => validateSglangSettings((raw ?? {})),
+    buildLaunch: buildSglangLaunch,
+    readinessTimeoutMs: sglangReadinessTimeoutMs,
+    classifyExit: classifySglangExit,
+    capabilities: sglangCapabilities,
+    rewriteRequestBody: sglangRewriteRequestBody,
+    rewriteResponseFor: sglangRewriteResponseFor,
+    mapErrorResponse: (_route, status, body) => mapSglangContextLengthError(status, body),
+    // Everything SGLang is started with — the output limit too, an environment variable of its launch —
+    // and the sampling defaults its gateway writes; only the load timeout is a load's alone.
+    restartKey: (settings) => [
+        settings.context_length,
+        settings.max_output_tokens,
+        settings.max_running_requests,
+        settings.mem_fraction_static,
+        settings.max_total_tokens,
+        settings.cuda_graphs,
+        settings.kv_cache_dtype,
+        settings.enable_prefix_caching,
+        settings.chunked_prefill_size,
+        settings.dtype,
+        settings.seed,
+        JSON.stringify(settings.generation),
+    ],
+};
+//# sourceMappingURL=adapter.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"adapter.js","sourceRoot":"","sources":["../../../src/runtime/sglang/adapter.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4BG;AACH,OAAO,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAA;AAE1D,OAAO,EAAE,kBAAkB,EAAE,MAAM,4BAA4B,CAAA;AAC/D,OAAO,EAAE,sBAAsB,EAAE,MAAM,wCAAwC,CAAA;AAC/E,OAAO,EACL,qCAAqC,EACrC,qBAAqB,EACrB,uBAAuB,EACvB,YAAY,EACZ,kBAAkB,EAClB,YAAY,EACZ,oBAAoB,EACpB,iBAAiB,GAQlB,MAAM,0BAA0B,CAAA;AACjC,OAAO,EAAE,sBAAsB,EAAuB,MAAM,eAAe,CAAA;AAE3E,MAAM,GAAG,GAAG,IAAI,IAAI,CAAC,CAAA;AAErB,8FAA8F;AAC9F,MAAM,CAAC,MAAM,qBAAqB,GAAG,KAAK,CAAA;AAgB1C,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,OAAO,IAAI,CAAC,mBAAmB,CAAC,KAAK,QAAQ;QAC7C,IAAI,CAAC,mBAAmB,CAAC,GAAG,CAAC;QAC7B,IAAI,CAAC,mBAAmB,CAAC,GAAG,CAAC;QAC7B,OAAO,IAAI,CAAC,YAAY,CAAC,KAAK,SAAS,CACxC,CAAA;AACH,CAAC;AAED;;;;GAIG;AACH,SAAS,aAAa,CAAC,QAAwB;IAC7C,QAAQ,QAAQ,CAAC,WAAW,EAAE,CAAC;QAC7B,KAAK,MAAM;YACT,OAAO,CAAC,4BAA4B,EAAE,MAAM,CAAC,QAAQ,CAAC,oBAAoB,CAAC,CAAC,CAAA;QAC9E,KAAK,IAAI;YACP,OAAO,EAAE,CAAA;QACX,KAAK,KAAK;YACR,OAAO,CAAC,6BAA6B,EAAE,8BAA8B,CAAC,CAAA;IAC1E,CAAC;AACH,CAAC;AAED;;;;GAIG;AACH,SAAS,QAAQ,CAAC,KAAa;IAC7B,OAAO;QACL,gBAAgB,EAAE,GAAG,KAAK,SAAS;QACnC,oBAAoB,EAAE,GAAG,KAAK,aAAa;QAC3C,gBAAgB,EAAE,GAAG,KAAK,SAAS;QACnC,uBAAuB,EAAE,GAAG,KAAK,WAAW;QAC5C,yBAAyB,EAAE,GAAG,KAAK,aAAa;QAChD,eAAe,EAAE,GAAG,KAAK,KAAK;QAC9B,kBAAkB,EAAE,GAAG,KAAK,WAAW;KACxC,CAAA;AACH,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,OAA6C;IAC7E,IAAI,CAAC,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC;QAC1B,MAAM,IAAI,eAAe,CACvB,kBAAkB,EAClB,sHAAsH,CACvH,CAAA;IACH,CAAC;IACD,MAAM,EAAE,QAAQ,EAAE,MAAM,EAAE,GAAG,OAAO,CAAA;IACpC,MAAM,IAAI,GAAG;QACX,SAAS;QACT,IAAI;QACJ,sBAAsB;QACtB,cAAc;QACd,OAAO,CAAC,SAAS;QACjB,qBAAqB;QACrB,OAAO,CAAC,OAAO;QACf,QAAQ;QACR,SAAS;QACT,QAAQ;QACR,MAAM,CAAC,qBAAqB,CAAC;QAC7B,kBAAkB;QAClB,MAAM,CAAC,QAAQ,CAAC,cAAc,CAAC;QAC/B,wBAAwB;QACxB,MAAM,CAAC,QAAQ,CAAC,oBAAoB,CAAC;QACrC,uBAAuB;QACvB,MAAM,CAAC,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,iBAAiB,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC;KAC1D,CAAA;IACD,IAAI,QAAQ,CAAC,gBAAgB,KAAK,IAAI;QAAE,IAAI,CAAC,IAAI,CAAC,oBAAoB,EAAE,MAAM,CAAC,QAAQ,CAAC,gBAAgB,CAAC,CAAC,CAAA;IAC1G,IAAI,QAAQ,CAAC,oBAAoB,KAAK,IAAI,EAAE,CAAC;QAC3C,IAAI,CAAC,IAAI,CAAC,wBAAwB,EAAE,MAAM,CAAC,QAAQ,CAAC,oBAAoB,CAAC,CAAC,CAAA;IAC5E,CAAC;IACD,IAAI,CAAC,QAAQ,CAAC,qBAAqB;QAAE,IAAI,CAAC,IAAI,CAAC,uBAAuB,CAAC,CAAA;IACvE,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,eAAe,EAAE,MAAM,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAA;IAC7E,IAAI,CAAC,IAAI,CAAC,GAAG,aAAa,CAAC,QAAQ,CAAC,CAAC,CAAA;IACrC,IAAI,QAAQ,CAAC,cAAc,KAAK,KAAK,IAAI,kBAAkB,CAAC,OAAO,CAAC,oBAAoB,CAAC,EAAE,CAAC;QAC1F,IAAI,CAAC,IAAI,CAAC,kBAAkB,EAAE,UAAU,CAAC,CAAA;IAC3C,CAAC;IACD,gGAAgG;IAChG,iGAAiG;IACjG,oCAAoC;IACpC,IAAI,OAAO,CAAC,IAAI,CAAC,UAAU,EAAE,CAAC;QAC5B,IAAI,CAAC,IAAI,CAAC,4BAA4B,EAAE,IAAI,CAAC,SAAS,CAAC,EAAE,mBAAmB,EAAE,IAAI,EAAE,CAAC,CAAC,CAAA;IACxF,CAAC;IACD,IAAI,MAAM,EAAE,WAAW;QAAE,IAAI,CAAC,IAAI,CAAC,oBAAoB,EAAE,MAAM,CAAC,WAAW,CAAC,CAAA;IAC5E,IAAI,MAAM,EAAE,gBAAgB;QAAE,IAAI,CAAC,IAAI,CAAC,oBAAoB,EAAE,MAAM,CAAC,gBAAgB,CAAC,CAAA;IACtF,OAAO;QACL,MAAM,EAAE,EAAE,cAAc,EAAE,qBAAqB,EAAE;QACjD,IAAI;QACJ,GAAG,EAAE;YACH,cAAc,EAAE,GAAG;YACnB,oBAAoB,EAAE,GAAG;YACzB,YAAY,EAAE,GAAG;YACjB,0FAA0F;YAC1F,8FAA8F;YAC9F,qFAAqF;YACrF,2BAA2B,EAAE,MAAM,CAAC,QAAQ,CAAC,iBAAiB,CAAC;YAC/D,GAAG,QAAQ,CAAC,OAAO,CAAC,eAAe,CAAC;SACrC;KACF,CAAA;AACH,CAAC;AAED,2GAA2G;AAC3G,MAAM,CAAC,MAAM,wBAAwB,GAAG,EAAE,GAAG,MAAM,CAAA;AACnD,MAAM,CAAC,MAAM,2BAA2B,GAAG,MAAM,CAAA;AAEjD,MAAM,UAAU,wBAAwB,CAAC,WAAmB,EAAE,QAAwB;IACpF,IAAI,QAAQ,CAAC,oBAAoB,KAAK,IAAI;QAAE,OAAO,QAAQ,CAAC,oBAAoB,GAAG,IAAI,CAAA;IACvF,OAAO,IAAI,CAAC,IAAI,CAAC,wBAAwB,GAAG,CAAC,WAAW,GAAG,GAAG,CAAC,GAAG,2BAA2B,CAAC,CAAA;AAChG,CAAC;AAED,gGAAgG;AAChG,sBAAsB;AACtB,gGAAgG;AAEhG,MAAM,GAAG,GACP,iHAAiH,CAAA;AACnH,sFAAsF;AACtF,MAAM,YAAY,GAChB,uIAAuI,CAAA;AACzI,qEAAqE;AACrE,MAAM,UAAU,GACd,qHAAqH,CAAA;AACvH,6GAA6G;AAC7G,MAAM,gBAAgB,GAAG,+EAA+E,CAAA;AACxG,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,kBAAkB,CAAC,GAAW,EAAE,QAAuB;IACrE,MAAM,IAAI,GAAG,YAAY,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;IACnC,IAAI,IAAI,KAAK,IAAI,EAAE,CAAC;QAClB,OAAO;YACL,IAAI,EAAE,eAAe;YACrB,OAAO,EACL,mFAAmF,IAAI,CAAC,CAAC,CAAC,UAAU;gBACpG,oCAAoC,IAAI,CAAC,CAAC,CAAC,+CAA+C;YAC5F,OAAO,EAAE,EAAE,mBAAmB,EAAE,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,2BAA2B,EAAE,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE;YAC/F,OAAO,EAAE,IAAI,CAAC,CAAC,CAAC;SACjB,CAAA;IACH,CAAC;IACD,MAAM,KAAK,GAAG,UAAU,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;IAClC,IAAI,KAAK,KAAK,IAAI,EAAE,CAAC;QACnB,OAAO;YACL,IAAI,EAAE,eAAe;YACrB,OAAO,EACL,oDAAoD;gBACpD,CAAC,KAAK,CAAC,CAAC,CAAC,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,oCAAoC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC;gBAC9E,+CAA+C;YACjD,OAAO,EAAE,KAAK,CAAC,CAAC,CAAC,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,mBAAmB,EAAE,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE;YAChF,OAAO,EAAE,KAAK,CAAC,CAAC,CAAC;SAClB,CAAA;IACH,CAAC;IACD,MAAM,GAAG,GAAG,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;IACzB,IAAI,GAAG,KAAK,IAAI,EAAE,CAAC;QACjB,OAAO;YACL,IAAI,EAAE,eAAe;YACrB,OAAO,EAAE,uDAAuD,GAAG,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,CAAC,SAAS,GAAG,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,CAAC,QAAQ;YACjH,OAAO,EAAE;gBACP,aAAa,EAAE,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,CAAW,CAAC,CAAC;gBAC7D,QAAQ,EAAE,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,CAAW,CAAC,CAAC;aACzD;YACD,OAAO,EAAE,GAAG,CAAC,CAAC,CAAC;SAChB,CAAA;IACH,CAAC;IACD,MAAM,OAAO,GAAG,gBAAgB,CAAC,IAAI,CAAC,GAAG,CAAC,CAAA;IAC1C,IAAI,OAAO,KAAK,IAAI,EAAE,CAAC;QACrB,OAAO;YACL,IAAI,EAAE,OAAO;YACb,OAAO,EACL,sBAAsB,OAAO,CAAC,CAAC,CAAC,uCAAuC,OAAO,CAAC,CAAC,CAAC,YAAY;gBAC7F,sDAAsD,OAAO,CAAC,CAAC,CAAC,WAAW;YAC7E,OAAO,EAAE,EAAE,cAAc,EAAE,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,EAAE,oBAAoB,EAAE,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,EAAE;YACzF,OAAO,EAAE,OAAO,CAAC,CAAC,CAAC;SACpB,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,4DAA4D,IAAI,CAAC,CAAC,CAAC,EAAE,OAAO,CAAC,IAAI,EAAE,EAAE,CAAC,GAAG;YAClG,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,iDAAiD,EAAE,CAAA;IAC9F,CAAC;IACD,OAAO;QACL,IAAI,EAAE,OAAO;QACb,OAAO,EAAE,gBAAgB,QAAQ,KAAK,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,cAAc,QAAQ,EAAE,uBAAuB;KAClG,CAAA;AACH,CAAC;AAED,gGAAgG;AAChG,sDAAsD;AACtD,gGAAgG;AAEhG,MAAM,UAAU,kBAAkB,CAAC,OAGlC;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,sHAAsH;AACtH,MAAM,CAAC,MAAM,aAAa,GAA4B;IACpD,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,wBAAwB,GAA4B;IAC/D,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,wBAAwB,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAA;AAE/E;;;;;;;GAOG;AACH,MAAM,UAAU,wBAAwB,CACtC,KAAa,EACb,IAAa,EACb,QAAwB,EACxB,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,OAAO,kBAAkB,CACvB,KAAK,EACL,sBAAsB,CAAC,MAAM,EAAE,QAAQ,CAAC,UAAU,CAAC,EACnD,QAAQ,CAAC,iBAAiB,CAC3B,CAAA;AACH,CAAC;AAED;;;GAGG;AACH,MAAM,QAAQ,GAAqE;IACjF;QACE,OAAO,EAAE,uFAAuF;QAChG,KAAK,EAAE,CAAC;QACR,SAAS,EAAE,CAAC;KACb;IACD;QACE,OAAO,EAAE,gFAAgF;QACzF,KAAK,EAAE,CAAC;QACR,SAAS,EAAE,CAAC;KACb;CACF,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,uGAAuG;AACvG,MAAM,UAAU,2BAA2B,CAAC,MAAc,EAAE,IAAY;IACtE,IAAI,MAAM,KAAK,GAAG;QAAE,OAAO,IAAI,CAAA;IAC/B,MAAM,OAAO,GAAG,YAAY,CAAC,IAAI,CAAC,CAAA;IAClC,KAAK,MAAM,EAAE,OAAO,EAAE,KAAK,EAAE,SAAS,EAAE,IAAI,QAAQ,EAAE,CAAC;QACrD,MAAM,KAAK,GAAG,OAAO,CAAC,IAAI,CAAC,OAAO,CAAC,CAAA;QACnC,IAAI,KAAK,KAAK,IAAI;YAAE,SAAQ;QAC5B,OAAO;YACL,KAAK,EAAE;gBACL,OAAO,EAAE,0CAA0C,KAAK,CAAC,KAAK,CAAC,gCAAgC,KAAK,CAAC,SAAS,CAAC,4DAA4D;gBAC3K,IAAI,EAAE,uBAAuB;gBAC7B,KAAK,EAAE,UAAU;gBACjB,IAAI,EAAE,yBAAyB;aAChC;SACF,CAAA;IACH,CAAC;IACD,OAAO,IAAI,CAAA;AACb,CAAC;AAED;;;;GAIG;AACH,MAAM,CAAC,MAAM,iCAAiC,GAAwB,IAAI,GAAG,CAAC,CAAC,OAAO,CAAC,CAAC,CAAA;AAExF,MAAM,UAAU,wBAAwB,CACtC,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,iCAAiC,CAAC,GAAG,CAAC,MAAM,CAAC;QAAE,OAAO,IAAI,CAAA;IAClF,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,mBAAmB,EAAE;IAC9D,EAAE,KAAK,EAAE,qBAAqB,EAAE,OAAO,EAAE,iBAAiB,EAAE;IAC5D,EAAE,KAAK,EAAE,qBAAqB,EAAE,OAAO,EAAE,iBAAiB,EAAE;IAC5D,EAAE,KAAK,EAAE,qBAAqB,EAAE,OAAO,EAAE,kBAAkB,EAAE;IAC7D,EAAE,KAAK,EAAE,qBAAqB,EAAE,OAAO,EAAE,uBAAuB,EAAE;CACnE,CAAA;AAED,MAAM,CAAC,MAAM,aAAa,GAAuC;IAC/D,EAAE,EAAE,QAAQ;IACZ,eAAe,EAAE,qCAAqC;IACtD,SAAS,EAAE,EAAE,IAAI,EAAE,SAAS,EAAE,cAAc,EAAE,GAAG,EAAE;IACnD,kGAAkG;IAClG,eAAe,EAAE;QACf,UAAU,EAAE,YAAY;QACxB,IAAI,EAAE,sBAAsB;QAC5B,IAAI,EAAE,EAAE,QAAQ,EAAE,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,MAAM,EAAE,KAAK,EAAE;KACpF;IACD,MAAM,EAAE,aAAa;IACrB,gBAAgB,EAAE,wBAAwB;IAC1C,YAAY,EAAE,aAAa;IAC3B,gBAAgB,EAAE,CAAC,GAAG,EAAE,EAAE,CAAC,sBAAsB,CAAC,CAAC,GAAG,IAAI,EAAE,CAA4B,CAAC;IACzF,WAAW,EAAE,iBAAiB;IAC9B,kBAAkB,EAAE,wBAAwB;IAC5C,YAAY,EAAE,kBAAkB;IAChC,YAAY,EAAE,kBAAkB;IAChC,kBAAkB,EAAE,wBAAwB;IAC5C,kBAAkB,EAAE,wBAAwB;IAC5C,gBAAgB,EAAE,CAAC,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,EAAE,CAAC,2BAA2B,CAAC,MAAM,EAAE,IAAI,CAAC;IACrF,oGAAoG;IACpG,yFAAyF;IACzF,UAAU,EAAE,CAAC,QAAQ,EAAE,EAAE,CAAC;QACxB,QAAQ,CAAC,cAAc;QACvB,QAAQ,CAAC,iBAAiB;QAC1B,QAAQ,CAAC,oBAAoB;QAC7B,QAAQ,CAAC,mBAAmB;QAC5B,QAAQ,CAAC,gBAAgB;QACzB,QAAQ,CAAC,WAAW;QACpB,QAAQ,CAAC,cAAc;QACvB,QAAQ,CAAC,qBAAqB;QAC9B,QAAQ,CAAC,oBAAoB;QAC7B,QAAQ,CAAC,KAAK;QACd,QAAQ,CAAC,IAAI;QACb,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,UAAU,CAAC;KACpC;CACF,CAAA"}

package/dist/runtime/sglang/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 SglangSettings } from './settings.js';
+export declare const SGLANG_ENGINE_ID = "sglang";
+/**
+ * What the public server (`:1337`) must know about an `sglang` session to route it honestly: only the
+ * routes the adapter declares, tool calls and JSON output only as the session can, and SGLang'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 sglangRoutePolicy(capabilities: ManagedTextCapabilities | null, limits?: {
+    contextLength: number;
+    maxOutputTokens: number;
+} | null): SessionRoutePolicy;
+export declare const SGLANG_ENGINE: ManagedEngineSpec<SglangSettings>;

package/dist/runtime/sglang/engine.js

@@ -0,0 +1,38 @@
+/**
+ * SGLang as a managed engine (change `add-sglang-runtime`, task 2.4, design D1): its adapter,
+ * settings, memory rule and launch plan, route policy, and its descriptor in conf
+ * (`runtimes/sglang.json`, read only once conf main publishes it — until then the provider is
+ * unsupported, design D9).
+ */
+import { defaultDescriptorUrl } from '../environment/index.js';
+import { SGLANG_ROUTES, mapSglangContextLengthError, sglangAdapter } from './adapter.js';
+import { sglangCheckEngine, sglangLaunchPlan } from './memory.js';
+import { sglangSettings } from './settings.js';
+export const SGLANG_ENGINE_ID = 'sglang';
+/**
+ * What the public server (`:1337`) must know about an `sglang` session to route it honestly: only the
+ * routes the adapter declares, tool calls and JSON output only as the session can, and SGLang'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 sglangRoutePolicy(capabilities, limits = null) {
+    return {
+        routes: SGLANG_ROUTES,
+        tools: capabilities?.tools ?? true,
+        structuredOutput: capabilities?.structured_output ?? true,
+        mapError: mapSglangContextLengthError,
+        ...(limits === null ? {} : limits),
+    };
+}
+export const SGLANG_ENGINE = {
+    engine_id: SGLANG_ENGINE_ID,
+    provider: SGLANG_ENGINE_ID,
+    label: 'SGLang',
+    descriptor: { engine_id: SGLANG_ENGINE_ID, label: 'SGLang', url: defaultDescriptorUrl(SGLANG_ENGINE_ID) },
+    adapter: sglangAdapter,
+    settings: sglangSettings,
+    check: sglangCheckEngine,
+    launchPlan: sglangLaunchPlan,
+    routePolicy: sglangRoutePolicy,
+};
+//# sourceMappingURL=engine.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"engine.js","sourceRoot":"","sources":["../../../src/runtime/sglang/engine.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,OAAO,EAAE,oBAAoB,EAAE,MAAM,yBAAyB,CAAA;AAI9D,OAAO,EAAE,aAAa,EAAE,2BAA2B,EAAE,aAAa,EAAE,MAAM,cAAc,CAAA;AACxF,OAAO,EAAE,iBAAiB,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAA;AACjE,OAAO,EAAE,cAAc,EAAuB,MAAM,eAAe,CAAA;AAEnE,MAAM,CAAC,MAAM,gBAAgB,GAAG,QAAQ,CAAA;AAExC;;;;;GAKG;AACH,MAAM,UAAU,iBAAiB,CAC/B,YAA4C,EAC5C,SAAoE,IAAI;IAExE,OAAO;QACL,MAAM,EAAE,aAAa;QACrB,KAAK,EAAE,YAAY,EAAE,KAAK,IAAI,IAAI;QAClC,gBAAgB,EAAE,YAAY,EAAE,iBAAiB,IAAI,IAAI;QACzD,QAAQ,EAAE,2BAA2B;QACrC,GAAG,CAAC,MAAM,KAAK,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,MAAM,CAAC;KACnC,CAAA;AACH,CAAC;AAED,MAAM,CAAC,MAAM,aAAa,GAAsC;IAC9D,SAAS,EAAE,gBAAgB;IAC3B,QAAQ,EAAE,gBAAgB;IAC1B,KAAK,EAAE,QAAQ;IACf,UAAU,EAAE,EAAE,SAAS,EAAE,gBAAgB,EAAE,KAAK,EAAE,QAAQ,EAAE,GAAG,EAAE,oBAAoB,CAAC,gBAAgB,CAAC,EAAE;IACzG,OAAO,EAAE,aAAa;IACtB,QAAQ,EAAE,cAAc;IACxB,KAAK,EAAE,iBAAiB;IACxB,UAAU,EAAE,gBAAgB;IAC5B,WAAW,EAAE,iBAAiB;CAC/B,CAAA"}

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

@@ -0,0 +1,11 @@
+/**
+ * The `sglang` provider (change `add-sglang-runtime`): its settings (`settings.ts`), the
+ * `ManagedTextAdapter` for `sglang.launch_server` (`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/sglang/index.js

@@ -0,0 +1,12 @@
+/**
+ * The `sglang` provider (change `add-sglang-runtime`): its settings (`settings.ts`), the
+ * `ManagedTextAdapter` for `sglang.launch_server` (`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/sglang/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../src/runtime/sglang/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/sglang/memory.d.ts

@@ -0,0 +1,64 @@
+/**
+ * SGLang's memory, decided by core (change `add-sglang-runtime`, task 2.3, design D4; spec
+ * `sglang-runtime`, "Память SGLang задаёт core"). SGLang sizes its budget from the memory it finds
+ * free at start-up, once its own CUDA context exists: it keeps `--mem-fraction-static` of it for the
+ * weights and the KV cache, and the rest is its room for activations and CUDA graphs
+ * (`mem_cache/kv_cache_configurator.py::_profile_available_bytes`). Its own default share is computed
+ * from the card's *total* memory, so on a card a desktop already holds part of, that room comes out
+ * smaller than SGLang meant. Core instead:
+ *
+ * - gives SGLang (unless the share setting fixes it) `f = (F − R) / F`, clamped to [0.5, 0.95], where
+ *   `F` is what the card has free right before the container is created and `R` the room SGLang needs
+ *   (2 GiB with CUDA graphs, 1.25 GiB without): SGLang then keeps about `R` whatever other processes
+ *   took. On a card with the host's memory the static share holds at most half of the host's memory:
+ *   `(min(F, half the host) − R) / F`. Within its share SGLang sizes the KV cache itself — hybrid
+ *   models' linear-attention state included — and core passes `--max-total-tokens` only from the
+ *   setting.
+ * - checks, before any container, that the weights, an estimate of the KV cache (context × parallel
+ *   requests × the model's per-token KV bytes over its attention layers, or the KV token limit) and
+ *   SGLang's own overhead fit the free memory — the same rule `POST /models/sglang/check` applies. It
+ *   is an estimate for the verdict and the refusal before a container; it sizes nothing SGLang runs
+ *   with.
+ *
+ * The reserve, the overhead (~2 GiB: CUDA context, activations, graphs) and the margin are first
+ * estimates; the live run of task 4.1 records the numbers (`summary.json`, `rulings/core.md`).
+ */
+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 { SglangLaunchPlan } from './adapter.js';
+import type { SglangSettings } from './settings.js';
+/** SGLang's room for activations and CUDA graphs captured up to the parallel requests (design D4, D5). */
+export declare const SGLANG_RUNTIME_RESERVE_BYTES: number;
+/** The same room with CUDA graphs off. */
+export declare const SGLANG_RUNTIME_RESERVE_NO_GRAPHS_BYTES: number;
+/** What SGLang holds beyond weights and KV cache: CUDA context, activations, graphs. */
+export declare const SGLANG_ENGINE_OVERHEAD_BYTES: number;
+/** Counted as needed in the memory check: what other processes may still take meanwhile. */
+export declare const SGLANG_FREE_MEMORY_MARGIN_BYTES: number;
+export declare const SGLANG_MIN_MEM_FRACTION_STATIC = 0.5;
+export declare const SGLANG_MAX_MEM_FRACTION_STATIC = 0.95;
+/**
+ * The KV cache the memory check counts: the KV token limit when set, else context × parallel
+ * requests, × the bytes per token over the model's attention layers — and whether the model's shape
+ * sized it. Without a shape: a quarter of the weights, at least 1 GiB.
+ */
+export declare function sglangKvCacheBytes(checkpoint: {
+    configJson: JsonObject;
+    hfQuantConfigJson: JsonObject | null;
+    weightBytesTotal?: number;
+}, settings: SglangSettings, gpu: GpuFacts): {
+    bytes: number;
+    basis: KvReserveBasis;
+};
+/**
+ * SGLang's memory rule on `gpu`: weights + KV cache estimate + its own overhead + 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 sglangMemoryNeed(gpu: GpuFacts, checkpoint: CheckpointMemoryInputs, settings: SglangSettings, host: HostMemory): MemoryNeed;
+/** `--mem-fraction-static` from the card as it stands now (see the file banner). */
+export declare function sglangMemFractionStatic(settings: SglangSettings, gpu: GpuFacts, host: HostMemory): number;
+/** The launch plan `beforeCreate` hands the adapter, from the card re-probed right before the container. */
+export declare function sglangLaunchPlan(settings: SglangSettings, checkpoint: CheckpointMemoryInputs, gpu: GpuFacts, host: HostMemory): SglangLaunchPlan;
+/** The check skeleton's hooks for SGLang: its memory rule, and no checkpoint quirks of its own. */
+export declare function sglangCheckEngine(settings: SglangSettings): ManagedCheckEngine;