@atomic-chat/core 0.18.4 → 0.19.0

package/dist/runtime/litert-lm/host/args.d.ts

@@ -0,0 +1,47 @@
+/**
+ * The arguments of the hidden `litert-host` mode (openspec change `add-litert-lm-runtime`, task 5.1).
+ * Core is the only caller, so the parse is strict: anything unknown is a bug on core's side.
+ *
+ *   litert-host probe   --lib <build>
+ *   litert-host inspect --lib <build> <file>
+ *   litert-host serve   --lib <build> --engine-version <v> --model <file> --backend cpu|gpu|npu
+ *                       --max-tokens <n> --port <p, 0 for any> --cache <dir> --id <model id>
+ *                       [--dispatch <dir> --ov-libs <dir>] [--speculative] [--reasoning]
+ */
+export declare const LITERT_LM_BACKENDS: readonly ["cpu", "gpu", "npu"];
+export type LiteRtLmBackend = (typeof LITERT_LM_BACKENDS)[number];
+export type LiteRtHostArgs = {
+    mode: 'probe';
+    lib: string;
+} | {
+    mode: 'inspect';
+    lib: string;
+    file: string;
+} | {
+    mode: 'serve';
+    /** The build directory, holding `lib/`. */
+    lib: string;
+    /** The build's LiteRT-LM version, checked against the file's `min_runtime_version`. */
+    engineVersion: string;
+    model: string;
+    backend: LiteRtLmBackend;
+    /** The engine's `max_num_tokens`, already cut to what the file allows. */
+    maxTokens: number;
+    port: number;
+    /** The library's `cache_dir`; created when missing. */
+    cache: string;
+    /** The id `/v1/models` lists: core's session id. */
+    id: string;
+    /** NPU only: the build's `vendors/intel_openvino/dispatch`. */
+    dispatch?: string;
+    /** NPU only: the prepared OpenVINO libraries. */
+    ovLibs?: string;
+    /** Use the file's draft model when it has one. */
+    speculative: boolean;
+    /** Think by default (`model.yml` `reasoning`); a request may still turn it off. */
+    reasoning: boolean;
+};
+export declare class LiteRtHostUsageError extends Error {
+    constructor(message: string);
+}
+export declare function parseLiteRtHostArgs(argv: readonly string[]): LiteRtHostArgs;

package/dist/runtime/litert-lm/host/args.js

@@ -0,0 +1,108 @@
+/**
+ * The arguments of the hidden `litert-host` mode (openspec change `add-litert-lm-runtime`, task 5.1).
+ * Core is the only caller, so the parse is strict: anything unknown is a bug on core's side.
+ *
+ *   litert-host probe   --lib <build>
+ *   litert-host inspect --lib <build> <file>
+ *   litert-host serve   --lib <build> --engine-version <v> --model <file> --backend cpu|gpu|npu
+ *                       --max-tokens <n> --port <p, 0 for any> --cache <dir> --id <model id>
+ *                       [--dispatch <dir> --ov-libs <dir>] [--speculative] [--reasoning]
+ */
+import { parseArgs } from 'node:util';
+export const LITERT_LM_BACKENDS = ['cpu', 'gpu', 'npu'];
+export class LiteRtHostUsageError extends Error {
+    constructor(message) {
+        super(message);
+        this.name = 'LiteRtHostUsageError';
+    }
+}
+const required = (values, key) => {
+    const value = values[key];
+    if (typeof value !== 'string' || value === '')
+        throw new LiteRtHostUsageError(`--${key} is required`);
+    return value;
+};
+const VERSION_RE = /^\d+\.\d+\.\d+$/;
+function wholeNumber(values, key, min, max) {
+    const raw = required(values, key);
+    const value = Number(raw);
+    if (!Number.isSafeInteger(value) || value < min || value > max)
+        throw new LiteRtHostUsageError(`--${key} ${raw} is not a whole number from ${min} to ${max}`);
+    return value;
+}
+export function parseLiteRtHostArgs(argv) {
+    const [mode, ...rest] = argv;
+    try {
+        if (mode === 'probe') {
+            const { values } = parseArgs({ args: rest, options: { lib: { type: 'string' } }, strict: true });
+            return { mode, lib: required(values, 'lib') };
+        }
+        if (mode === 'inspect') {
+            const { values, positionals } = parseArgs({
+                args: rest,
+                options: { lib: { type: 'string' } },
+                strict: true,
+                allowPositionals: true,
+            });
+            const lib = required(values, 'lib');
+            if (positionals.length !== 1 || !positionals[0])
+                throw new LiteRtHostUsageError('inspect takes one model file');
+            return { mode, lib, file: positionals[0] };
+        }
+        if (mode === 'serve') {
+            const { values } = parseArgs({
+                args: rest,
+                options: {
+                    'lib': { type: 'string' },
+                    'engine-version': { type: 'string' },
+                    'model': { type: 'string' },
+                    'backend': { type: 'string' },
+                    'max-tokens': { type: 'string' },
+                    'port': { type: 'string' },
+                    'cache': { type: 'string' },
+                    'id': { type: 'string' },
+                    'dispatch': { type: 'string' },
+                    'ov-libs': { type: 'string' },
+                    'speculative': { type: 'boolean' },
+                    'reasoning': { type: 'boolean' },
+                },
+                strict: true,
+            });
+            const lib = required(values, 'lib');
+            const engineVersion = required(values, 'engine-version');
+            if (!VERSION_RE.test(engineVersion))
+                throw new LiteRtHostUsageError(`--engine-version ${engineVersion} is not X.Y.Z`);
+            const backend = required(values, 'backend');
+            if (!LITERT_LM_BACKENDS.some((known) => known === backend))
+                throw new LiteRtHostUsageError(`--backend ${backend} is not cpu, gpu or npu`);
+            const dispatch = values['dispatch'];
+            const ovLibs = values['ov-libs'];
+            if (backend === 'npu' && (!dispatch || !ovLibs))
+                throw new LiteRtHostUsageError('--backend npu needs --dispatch and --ov-libs');
+            if (backend !== 'npu' && (dispatch !== undefined || ovLibs !== undefined))
+                throw new LiteRtHostUsageError('--dispatch and --ov-libs are only for --backend npu');
+            return {
+                mode,
+                lib,
+                engineVersion,
+                model: required(values, 'model'),
+                backend: backend,
+                maxTokens: wholeNumber(values, 'max-tokens', 1, 16_777_216),
+                port: wholeNumber(values, 'port', 0, 65535),
+                cache: required(values, 'cache'),
+                id: required(values, 'id'),
+                ...(dispatch ? { dispatch } : {}),
+                ...(ovLibs ? { ovLibs } : {}),
+                speculative: values['speculative'] === true,
+                reasoning: values['reasoning'] === true,
+            };
+        }
+    }
+    catch (error) {
+        if (error instanceof LiteRtHostUsageError)
+            throw error;
+        throw new LiteRtHostUsageError(error instanceof Error ? error.message : String(error));
+    }
+    throw new LiteRtHostUsageError(`litert-host: unknown mode ${JSON.stringify(mode ?? '')} (probe, inspect, serve)`);
+}
+//# sourceMappingURL=args.js.map

package/dist/runtime/litert-lm/host/args.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"args.js","sourceRoot":"","sources":["../../../../src/runtime/litert-lm/host/args.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAEH,OAAO,EAAE,SAAS,EAAE,MAAM,WAAW,CAAA;AAErC,MAAM,CAAC,MAAM,kBAAkB,GAAG,CAAC,KAAK,EAAE,KAAK,EAAE,KAAK,CAAU,CAAA;AA+BhE,MAAM,OAAO,oBAAqB,SAAQ,KAAK;IAC7C,YAAY,OAAe;QACzB,KAAK,CAAC,OAAO,CAAC,CAAA;QACd,IAAI,CAAC,IAAI,GAAG,sBAAsB,CAAA;IACpC,CAAC;CACF;AAED,MAAM,QAAQ,GAAG,CAAC,MAA+B,EAAE,GAAW,EAAU,EAAE;IACxE,MAAM,KAAK,GAAG,MAAM,CAAC,GAAG,CAAC,CAAA;IACzB,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,EAAE;QAAE,MAAM,IAAI,oBAAoB,CAAC,KAAK,GAAG,cAAc,CAAC,CAAA;IACrG,OAAO,KAAK,CAAA;AACd,CAAC,CAAA;AAED,MAAM,UAAU,GAAG,iBAAiB,CAAA;AAEpC,SAAS,WAAW,CAAC,MAA+B,EAAE,GAAW,EAAE,GAAW,EAAE,GAAW;IACzF,MAAM,GAAG,GAAG,QAAQ,CAAC,MAAM,EAAE,GAAG,CAAC,CAAA;IACjC,MAAM,KAAK,GAAG,MAAM,CAAC,GAAG,CAAC,CAAA;IACzB,IAAI,CAAC,MAAM,CAAC,aAAa,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,GAAG,IAAI,KAAK,GAAG,GAAG;QAC5D,MAAM,IAAI,oBAAoB,CAAC,KAAK,GAAG,IAAI,GAAG,+BAA+B,GAAG,OAAO,GAAG,EAAE,CAAC,CAAA;IAC/F,OAAO,KAAK,CAAA;AACd,CAAC;AAED,MAAM,UAAU,mBAAmB,CAAC,IAAuB;IACzD,MAAM,CAAC,IAAI,EAAE,GAAG,IAAI,CAAC,GAAG,IAAI,CAAA;IAC5B,IAAI,CAAC;QACH,IAAI,IAAI,KAAK,OAAO,EAAE,CAAC;YACrB,MAAM,EAAE,MAAM,EAAE,GAAG,SAAS,CAAC,EAAE,IAAI,EAAE,IAAI,EAAE,OAAO,EAAE,EAAE,GAAG,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,EAAE,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,CAAA;YAChG,OAAO,EAAE,IAAI,EAAE,GAAG,EAAE,QAAQ,CAAC,MAAM,EAAE,KAAK,CAAC,EAAE,CAAA;QAC/C,CAAC;QACD,IAAI,IAAI,KAAK,SAAS,EAAE,CAAC;YACvB,MAAM,EAAE,MAAM,EAAE,WAAW,EAAE,GAAG,SAAS,CAAC;gBACxC,IAAI,EAAE,IAAI;gBACV,OAAO,EAAE,EAAE,GAAG,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE,EAAE;gBACpC,MAAM,EAAE,IAAI;gBACZ,gBAAgB,EAAE,IAAI;aACvB,CAAC,CAAA;YACF,MAAM,GAAG,GAAG,QAAQ,CAAC,MAAM,EAAE,KAAK,CAAC,CAAA;YACnC,IAAI,WAAW,CAAC,MAAM,KAAK,CAAC,IAAI,CAAC,WAAW,CAAC,CAAC,CAAC;gBAC7C,MAAM,IAAI,oBAAoB,CAAC,8BAA8B,CAAC,CAAA;YAChE,OAAO,EAAE,IAAI,EAAE,GAAG,EAAE,IAAI,EAAE,WAAW,CAAC,CAAC,CAAC,EAAE,CAAA;QAC5C,CAAC;QACD,IAAI,IAAI,KAAK,OAAO,EAAE,CAAC;YACrB,MAAM,EAAE,MAAM,EAAE,GAAG,SAAS,CAAC;gBAC3B,IAAI,EAAE,IAAI;gBACV,OAAO,EAAE;oBACP,KAAK,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE;oBACzB,gBAAgB,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE;oBACpC,OAAO,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE;oBAC3B,SAAS,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE;oBAC7B,YAAY,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE;oBAChC,MAAM,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE;oBAC1B,OAAO,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE;oBAC3B,IAAI,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE;oBACxB,UAAU,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE;oBAC9B,SAAS,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE;oBAC7B,aAAa,EAAE,EAAE,IAAI,EAAE,SAAS,EAAE;oBAClC,WAAW,EAAE,EAAE,IAAI,EAAE,SAAS,EAAE;iBACjC;gBACD,MAAM,EAAE,IAAI;aACb,CAAC,CAAA;YACF,MAAM,GAAG,GAAG,QAAQ,CAAC,MAAM,EAAE,KAAK,CAAC,CAAA;YACnC,MAAM,aAAa,GAAG,QAAQ,CAAC,MAAM,EAAE,gBAAgB,CAAC,CAAA;YACxD,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,aAAa,CAAC;gBACjC,MAAM,IAAI,oBAAoB,CAAC,oBAAoB,aAAa,eAAe,CAAC,CAAA;YAClF,MAAM,OAAO,GAAG,QAAQ,CAAC,MAAM,EAAE,SAAS,CAAC,CAAA;YAC3C,IAAI,CAAC,kBAAkB,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,KAAK,OAAO,CAAC;gBACxD,MAAM,IAAI,oBAAoB,CAAC,aAAa,OAAO,yBAAyB,CAAC,CAAA;YAC/E,MAAM,QAAQ,GAAG,MAAM,CAAC,UAAU,CAAC,CAAA;YACnC,MAAM,MAAM,GAAG,MAAM,CAAC,SAAS,CAAC,CAAA;YAChC,IAAI,OAAO,KAAK,KAAK,IAAI,CAAC,CAAC,QAAQ,IAAI,CAAC,MAAM,CAAC;gBAC7C,MAAM,IAAI,oBAAoB,CAAC,8CAA8C,CAAC,CAAA;YAChF,IAAI,OAAO,KAAK,KAAK,IAAI,CAAC,QAAQ,KAAK,SAAS,IAAI,MAAM,KAAK,SAAS,CAAC;gBACvE,MAAM,IAAI,oBAAoB,CAAC,qDAAqD,CAAC,CAAA;YACvF,OAAO;gBACL,IAAI;gBACJ,GAAG;gBACH,aAAa;gBACb,KAAK,EAAE,QAAQ,CAAC,MAAM,EAAE,OAAO,CAAC;gBAChC,OAAO,EAAE,OAA0B;gBACnC,SAAS,EAAE,WAAW,CAAC,MAAM,EAAE,YAAY,EAAE,CAAC,EAAE,UAAU,CAAC;gBAC3D,IAAI,EAAE,WAAW,CAAC,MAAM,EAAE,MAAM,EAAE,CAAC,EAAE,KAAK,CAAC;gBAC3C,KAAK,EAAE,QAAQ,CAAC,MAAM,EAAE,OAAO,CAAC;gBAChC,EAAE,EAAE,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC;gBAC1B,GAAG,CAAC,QAAQ,CAAC,CAAC,CAAC,EAAE,QAAQ,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBACjC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBAC7B,WAAW,EAAE,MAAM,CAAC,aAAa,CAAC,KAAK,IAAI;gBAC3C,SAAS,EAAE,MAAM,CAAC,WAAW,CAAC,KAAK,IAAI;aACxC,CAAA;QACH,CAAC;IACH,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,IAAI,KAAK,YAAY,oBAAoB;YAAE,MAAM,KAAK,CAAA;QACtD,MAAM,IAAI,oBAAoB,CAAC,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAA;IACxF,CAAC;IACD,MAAM,IAAI,oBAAoB,CAC5B,6BAA6B,IAAI,CAAC,SAAS,CAAC,IAAI,IAAI,EAAE,CAAC,0BAA0B,CAClF,CAAA;AACH,CAAC"}

package/dist/runtime/litert-lm/host/ffi-bun.d.ts

@@ -0,0 +1,16 @@
+/**
+ * `LiteRtLibrary` over `bun:ffi` — the second and last file in `src/` that may import `bun:*` (change
+ * `add-litert-lm-runtime`, design D1 and decision 3.6; allow-listed in both runtime-agnostic gates). It
+ * is imported dynamically by the `litert-host` mode alone, so the npm package still loads under Node.
+ *
+ * Streaming is the hard part (spike S2). The library calls a stream callback on a C++ thread of its
+ * own and frees the chunk as soon as the callback returns. A plain `JSCallback` cannot run there; a
+ * `threadsafe` one only queues the call, so the JS thread would read freed memory, and even with the
+ * chunk copied first Bun's threadsafe callbacks crashed about one run in five under a fast stream.
+ * So the callback is C: a shim compiled at start by the TinyCC built into Bun (`cc`) copies each chunk
+ * on the library's thread into a FIFO under a mutex, and JS drains the FIFO on a 2 ms timer that runs
+ * only while a generation is under way. The shim links nothing (`-nostdlib`: TinyCC finds no libc
+ * without development packages on Linux); `malloc` and the mutex come in as pointers from `dlsym`.
+ */
+import type { LoadLiteRtLibrary } from './library.js';
+export declare const loadBunLiteRtLibrary: LoadLiteRtLibrary;