@atomic-chat/core 0.5.1

package/dist/runtime/llamacpp/policy.js

@@ -0,0 +1,191 @@
+/**
+ * Small policy helpers of the llama.cpp extension, ported verbatim from
+ * `extensions/llamacpp-upstream-extension/src/{util,index}.ts`: backend-id predicates, the AVX
+ * preflight, backend-mismatch classification, env parsing, readiness timeout, load-error shaping.
+ */
+import { AtomicCoreError } from '../../contracts/index.js';
+export const CPU_NO_AVX_ERROR_CODE = 'CPU_NO_AVX';
+export const CPU_NO_AVX_MESSAGE = "Your CPU is too old to run this model: it doesn't support the AVX instruction set that the bundled engine requires. The app cannot run local models on this processor.";
+export const BACKEND_NOT_CONFIGURED_MESSAGE = 'Llama.cpp backend is not configured (version_backend is missing or invalid). Check Settings → Llama.cpp — Version & Backend, or reinstall the application.';
+/** Readiness wait floor (30 min): large models on cold storage outran the 600 s default. */
+export const MODEL_LOAD_READY_TIMEOUT_FLOOR_SECS = 1800;
+export function stripBom(s) {
+    return s.replace(/\uFEFF/g, '').trim();
+}
+/** The extension's strict build-number parser (`b6325` only; the unified fork tag yields null). */
+export function parseBuildNumberStrict(version) {
+    const m = /^b(\d+)$/.exec(version);
+    return m ? parseInt(m[1], 10) : null;
+}
+/** A concrete `<version>/<backend>`: not empty, not `none`, has a `/`, not the `latest/` sentinel. */
+export function isConcreteVersionBackend(vb) {
+    const v = stripBom(vb ?? '');
+    if (!v || v === 'none')
+        return false;
+    if (!v.includes('/'))
+        return false;
+    if (v.startsWith('latest/'))
+        return false;
+    return true;
+}
+/** `win-cpu-x64`, `linux-cpu-arm64`, … — never the macOS ids. */
+export function isCpuBackend(backend) {
+    return stripBom(backend ?? '')
+        .toLowerCase()
+        .includes('-cpu-');
+}
+export function cpuHasAvx(extensions) {
+    if (!extensions || extensions.length === 0)
+        return false;
+    return extensions.some((e) => {
+        const x = e.toLowerCase();
+        return x === 'avx' || x === 'avx2' || x.startsWith('avx512');
+    });
+}
+/** Block only on a positive no-AVX signal: x86, a CPU backend, a non-empty list lacking AVX. */
+export function isUnsupportedNoAvxCpu(arch, backend, extensions) {
+    const a = (arch ?? '').trim().toLowerCase();
+    const isX86 = a === 'x86_64' || a === 'x86' || a === 'amd64';
+    if (!isX86)
+        return false;
+    if (!isCpuBackend(backend))
+        return false;
+    if (!extensions || extensions.length === 0)
+        return false;
+    return !cpuHasAvx(extensions);
+}
+const GPU_BACKEND_CATEGORIES = new Set([
+    'cuda-cu13',
+    'cuda-cu13.0',
+    'cuda-cu12.4',
+    'cuda-cu12.0',
+    'cuda-cu11.7',
+    'vulkan',
+]);
+export function isGpuBackendCategory(category) {
+    return GPU_BACKEND_CATEGORIES.has(category);
+}
+export function gpuKindOf(category) {
+    if (category.startsWith('cuda-'))
+        return 'cuda';
+    if (category === 'vulkan')
+        return 'vulkan';
+    return 'other';
+}
+/** Zero offloaded layers is conclusive; an absent device is not treated as CPU. */
+export function runtimeRanOnCpu(runtimeDevice) {
+    if (!runtimeDevice)
+        return false;
+    if (runtimeDevice.gpu_layers_offloaded === 0)
+        return true;
+    const primary = (runtimeDevice.primary_device ?? '').trim();
+    if (!primary)
+        return false;
+    return primary === 'CPU' || primary.startsWith('CPU_');
+}
+/**
+ * Compare what the UI shows, what was launched and what the process reports. Precedence: silent
+ * swap, then a GPU build that ran on the CPU (unless `-ngl 0` was requested), then a better tier.
+ */
+export function classifyBackendMismatch(input) {
+    const configured = stripBom(input.configuredBackend ?? '');
+    const effective = stripBom(input.effectiveBackend ?? '') || configured;
+    if (!configured)
+        return { kind: 'ok' };
+    if (effective && effective !== configured)
+        return { kind: 'silent-fallback', configured, effective };
+    const effectiveCategory = input.categoryOf(effective);
+    const cpuOnlyByRequest = input.requestedGpuLayers === 0;
+    if (isGpuBackendCategory(effectiveCategory) && !cpuOnlyByRequest && runtimeRanOnCpu(input.runtimeDevice)) {
+        return {
+            kind: 'runtime-cpu',
+            configured: effective,
+            primaryDevice: (input.runtimeDevice?.primary_device ?? '').trim() || 'CPU',
+            offloaded: input.runtimeDevice?.gpu_layers_offloaded ?? null,
+            total: input.runtimeDevice?.total_layers ?? null,
+            gpuKind: gpuKindOf(effectiveCategory),
+            cudaRuntimeMissing: input.runtimeDevice?.cuda_runtime_missing === true,
+            deviceInitError: input.runtimeDevice?.device_init_error ?? null,
+        };
+    }
+    const ideal = stripBom(input.idealBackend ?? '');
+    if (ideal) {
+        const idealCategory = input.categoryOf(ideal);
+        if (isGpuBackendCategory(idealCategory) && idealCategory !== effectiveCategory) {
+            return { kind: 'suboptimal-config', configured: effective, ideal };
+        }
+    }
+    return { kind: 'ok' };
+}
+/** `KEY=VALUE;KEY2=VALUE2` — keys starting with `LLAMA` are reserved and dropped. */
+export function parseEnvString(envString, target = {}) {
+    envString
+        .split(';')
+        .filter((pair) => pair.trim())
+        .forEach((pair) => {
+        const [key, ...valueParts] = pair.split('=');
+        const cleanKey = key?.trim();
+        if (cleanKey && valueParts.length > 0 && !cleanKey.startsWith('LLAMA')) {
+            target[cleanKey] = valueParts.join('=').trim();
+        }
+    });
+    return target;
+}
+/** Never below the floor; honours a larger configured value; non-numbers read as 600. */
+export function modelLoadReadyTimeoutSecs(configuredTimeoutSecs) {
+    const configured = Number(configuredTimeoutSecs);
+    const base = Number.isFinite(configured) && configured > 0 ? configured : 600;
+    return Math.max(base, MODEL_LOAD_READY_TIMEOUT_FLOOR_SECS);
+}
+/** Readable text of any load error: message + details + `[code]`; never `[object Object]`. */
+export function formatLoadError(err) {
+    if (err instanceof Error) {
+        const e = err;
+        const parts = [err.message];
+        if (typeof e.details === 'string' && e.details.trim())
+            parts.push(e.details.trim());
+        const code = typeof e.code === 'string' && e.code ? ` [${e.code}]` : '';
+        return err.message ? `${parts.join('\n')}${code}` : String(err);
+    }
+    if (err && typeof err === 'object') {
+        const e = err;
+        const parts = [];
+        if (typeof e.message === 'string' && e.message.trim())
+            parts.push(e.message.trim());
+        if (typeof e.details === 'string' && e.details.trim())
+            parts.push(e.details.trim());
+        if (parts.length > 0) {
+            const code = typeof e.code === 'string' && e.code ? ` [${e.code}]` : '';
+            return `${parts.join('\n')}${code}`;
+        }
+        try {
+            const json = JSON.stringify(err);
+            if (json && json !== '{}' && json !== 'null')
+                return json;
+        }
+        catch {
+            /* fall through */
+        }
+    }
+    return String(err);
+}
+/** Codes that describe a recoverable user/environment condition, not a crash (kept out of Sentry). */
+export const RECOVERABLE_LOAD_ERROR_CODES = new Set([
+    'MODEL_FILE_NOT_FOUND',
+    'MODEL_FILE_CORRUPT',
+    'MODEL_SHARDS_INCOMPLETE',
+    'MULTIMODAL_PROJECTOR_LOAD_FAILED',
+    'BINARY_NOT_FOUND',
+    'MODEL_ARCH_NOT_SUPPORTED',
+    'OS_VERSION_UNSUPPORTED',
+    'CPU_NO_AVX',
+]);
+export function isRecoverableLoadError(err) {
+    const code = err?.code;
+    return typeof code === 'string' && RECOVERABLE_LOAD_ERROR_CODES.has(code);
+}
+/** The extension's `codedLoadError`, as an `AtomicCoreError`. */
+export function codedLoadError(code, message, details) {
+    return new AtomicCoreError(code, message, details);
+}
+//# sourceMappingURL=policy.js.map

package/dist/runtime/llamacpp/policy.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"policy.js","sourceRoot":"","sources":["../../../src/runtime/llamacpp/policy.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,OAAO,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAA;AAG1D,MAAM,CAAC,MAAM,qBAAqB,GAAG,YAAqB,CAAA;AAC1D,MAAM,CAAC,MAAM,kBAAkB,GAC7B,wKAAwK,CAAA;AAC1K,MAAM,CAAC,MAAM,8BAA8B,GACzC,4JAA4J,CAAA;AAE9J,4FAA4F;AAC5F,MAAM,CAAC,MAAM,mCAAmC,GAAG,IAAI,CAAA;AAEvD,MAAM,UAAU,QAAQ,CAAC,CAAS;IAChC,OAAO,CAAC,CAAC,OAAO,CAAC,SAAS,EAAE,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;AACxC,CAAC;AAED,mGAAmG;AACnG,MAAM,UAAU,sBAAsB,CAAC,OAAe;IACpD,MAAM,CAAC,GAAG,UAAU,CAAC,IAAI,CAAC,OAAO,CAAC,CAAA;IAClC,OAAO,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAW,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,IAAI,CAAA;AAChD,CAAC;AAED,sGAAsG;AACtG,MAAM,UAAU,wBAAwB,CAAC,EAA6B;IACpE,MAAM,CAAC,GAAG,QAAQ,CAAC,EAAE,IAAI,EAAE,CAAC,CAAA;IAC5B,IAAI,CAAC,CAAC,IAAI,CAAC,KAAK,MAAM;QAAE,OAAO,KAAK,CAAA;IACpC,IAAI,CAAC,CAAC,CAAC,QAAQ,CAAC,GAAG,CAAC;QAAE,OAAO,KAAK,CAAA;IAClC,IAAI,CAAC,CAAC,UAAU,CAAC,SAAS,CAAC;QAAE,OAAO,KAAK,CAAA;IACzC,OAAO,IAAI,CAAA;AACb,CAAC;AAED,iEAAiE;AACjE,MAAM,UAAU,YAAY,CAAC,OAAkC;IAC7D,OAAO,QAAQ,CAAC,OAAO,IAAI,EAAE,CAAC;SAC3B,WAAW,EAAE;SACb,QAAQ,CAAC,OAAO,CAAC,CAAA;AACtB,CAAC;AAED,MAAM,UAAU,SAAS,CAAC,UAAuC;IAC/D,IAAI,CAAC,UAAU,IAAI,UAAU,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,KAAK,CAAA;IACxD,OAAO,UAAU,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE;QAC3B,MAAM,CAAC,GAAG,CAAC,CAAC,WAAW,EAAE,CAAA;QACzB,OAAO,CAAC,KAAK,KAAK,IAAI,CAAC,KAAK,MAAM,IAAI,CAAC,CAAC,UAAU,CAAC,QAAQ,CAAC,CAAA;IAC9D,CAAC,CAAC,CAAA;AACJ,CAAC;AAED,gGAAgG;AAChG,MAAM,UAAU,qBAAqB,CACnC,IAA+B,EAC/B,OAAkC,EAClC,UAAuC;IAEvC,MAAM,CAAC,GAAG,CAAC,IAAI,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAA;IAC3C,MAAM,KAAK,GAAG,CAAC,KAAK,QAAQ,IAAI,CAAC,KAAK,KAAK,IAAI,CAAC,KAAK,OAAO,CAAA;IAC5D,IAAI,CAAC,KAAK;QAAE,OAAO,KAAK,CAAA;IACxB,IAAI,CAAC,YAAY,CAAC,OAAO,CAAC;QAAE,OAAO,KAAK,CAAA;IACxC,IAAI,CAAC,UAAU,IAAI,UAAU,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,KAAK,CAAA;IACxD,OAAO,CAAC,SAAS,CAAC,UAAU,CAAC,CAAA;AAC/B,CAAC;AAED,MAAM,sBAAsB,GAAG,IAAI,GAAG,CAAC;IACrC,WAAW;IACX,aAAa;IACb,aAAa;IACb,aAAa;IACb,aAAa;IACb,QAAQ;CACT,CAAC,CAAA;AAEF,MAAM,UAAU,oBAAoB,CAAC,QAAgB;IACnD,OAAO,sBAAsB,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAA;AAC7C,CAAC;AAID,MAAM,UAAU,SAAS,CAAC,QAAgB;IACxC,IAAI,QAAQ,CAAC,UAAU,CAAC,OAAO,CAAC;QAAE,OAAO,MAAM,CAAA;IAC/C,IAAI,QAAQ,KAAK,QAAQ;QAAE,OAAO,QAAQ,CAAA;IAC1C,OAAO,OAAO,CAAA;AAChB,CAAC;AAED,mFAAmF;AACnF,MAAM,UAAU,eAAe,CAAC,aAA4D;IAC1F,IAAI,CAAC,aAAa;QAAE,OAAO,KAAK,CAAA;IAChC,IAAI,aAAa,CAAC,oBAAoB,KAAK,CAAC;QAAE,OAAO,IAAI,CAAA;IACzD,MAAM,OAAO,GAAG,CAAC,aAAa,CAAC,cAAc,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;IAC3D,IAAI,CAAC,OAAO;QAAE,OAAO,KAAK,CAAA;IAC1B,OAAO,OAAO,KAAK,KAAK,IAAI,OAAO,CAAC,UAAU,CAAC,MAAM,CAAC,CAAA;AACxD,CAAC;AAiBD;;;GAGG;AACH,MAAM,UAAU,uBAAuB,CAAC,KAOvC;IACC,MAAM,UAAU,GAAG,QAAQ,CAAC,KAAK,CAAC,iBAAiB,IAAI,EAAE,CAAC,CAAA;IAC1D,MAAM,SAAS,GAAG,QAAQ,CAAC,KAAK,CAAC,gBAAgB,IAAI,EAAE,CAAC,IAAI,UAAU,CAAA;IACtE,IAAI,CAAC,UAAU;QAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,CAAA;IACtC,IAAI,SAAS,IAAI,SAAS,KAAK,UAAU;QAAE,OAAO,EAAE,IAAI,EAAE,iBAAiB,EAAE,UAAU,EAAE,SAAS,EAAE,CAAA;IACpG,MAAM,iBAAiB,GAAG,KAAK,CAAC,UAAU,CAAC,SAAS,CAAC,CAAA;IACrD,MAAM,gBAAgB,GAAG,KAAK,CAAC,kBAAkB,KAAK,CAAC,CAAA;IACvD,IAAI,oBAAoB,CAAC,iBAAiB,CAAC,IAAI,CAAC,gBAAgB,IAAI,eAAe,CAAC,KAAK,CAAC,aAAa,CAAC,EAAE,CAAC;QACzG,OAAO;YACL,IAAI,EAAE,aAAa;YACnB,UAAU,EAAE,SAAS;YACrB,aAAa,EAAE,CAAC,KAAK,CAAC,aAAa,EAAE,cAAc,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,IAAI,KAAK;YAC1E,SAAS,EAAE,KAAK,CAAC,aAAa,EAAE,oBAAoB,IAAI,IAAI;YAC5D,KAAK,EAAE,KAAK,CAAC,aAAa,EAAE,YAAY,IAAI,IAAI;YAChD,OAAO,EAAE,SAAS,CAAC,iBAAiB,CAAC;YACrC,kBAAkB,EAAE,KAAK,CAAC,aAAa,EAAE,oBAAoB,KAAK,IAAI;YACtE,eAAe,EAAE,KAAK,CAAC,aAAa,EAAE,iBAAiB,IAAI,IAAI;SAChE,CAAA;IACH,CAAC;IACD,MAAM,KAAK,GAAG,QAAQ,CAAC,KAAK,CAAC,YAAY,IAAI,EAAE,CAAC,CAAA;IAChD,IAAI,KAAK,EAAE,CAAC;QACV,MAAM,aAAa,GAAG,KAAK,CAAC,UAAU,CAAC,KAAK,CAAC,CAAA;QAC7C,IAAI,oBAAoB,CAAC,aAAa,CAAC,IAAI,aAAa,KAAK,iBAAiB,EAAE,CAAC;YAC/E,OAAO,EAAE,IAAI,EAAE,mBAAmB,EAAE,UAAU,EAAE,SAAS,EAAE,KAAK,EAAE,CAAA;QACpE,CAAC;IACH,CAAC;IACD,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,CAAA;AACvB,CAAC;AAED,qFAAqF;AACrF,MAAM,UAAU,cAAc,CAC5B,SAAiB,EACjB,SAAiC,EAAE;IAEnC,SAAS;SACN,KAAK,CAAC,GAAG,CAAC;SACV,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC;SAC7B,OAAO,CAAC,CAAC,IAAI,EAAE,EAAE;QAChB,MAAM,CAAC,GAAG,EAAE,GAAG,UAAU,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,CAAA;QAC5C,MAAM,QAAQ,GAAG,GAAG,EAAE,IAAI,EAAE,CAAA;QAC5B,IAAI,QAAQ,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC,IAAI,CAAC,QAAQ,CAAC,UAAU,CAAC,OAAO,CAAC,EAAE,CAAC;YACvE,MAAM,CAAC,QAAQ,CAAC,GAAG,UAAU,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,CAAA;QAChD,CAAC;IACH,CAAC,CAAC,CAAA;IACJ,OAAO,MAAM,CAAA;AACf,CAAC;AAED,yFAAyF;AACzF,MAAM,UAAU,yBAAyB,CAAC,qBAAsC;IAC9E,MAAM,UAAU,GAAG,MAAM,CAAC,qBAAqB,CAAC,CAAA;IAChD,MAAM,IAAI,GAAG,MAAM,CAAC,QAAQ,CAAC,UAAU,CAAC,IAAI,UAAU,GAAG,CAAC,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,GAAG,CAAA;IAC7E,OAAO,IAAI,CAAC,GAAG,CAAC,IAAI,EAAE,mCAAmC,CAAC,CAAA;AAC5D,CAAC;AAED,8FAA8F;AAC9F,MAAM,UAAU,eAAe,CAAC,GAAY;IAC1C,IAAI,GAAG,YAAY,KAAK,EAAE,CAAC;QACzB,MAAM,CAAC,GAAG,GAAoD,CAAA;QAC9D,MAAM,KAAK,GAAG,CAAC,GAAG,CAAC,OAAO,CAAC,CAAA;QAC3B,IAAI,OAAO,CAAC,CAAC,OAAO,KAAK,QAAQ,IAAI,CAAC,CAAC,OAAO,CAAC,IAAI,EAAE;YAAE,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,OAAO,CAAC,IAAI,EAAE,CAAC,CAAA;QACnF,MAAM,IAAI,GAAG,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ,IAAI,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,CAAC,EAAE,CAAA;QACvE,OAAO,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,IAAI,EAAE,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,CAAA;IACjE,CAAC;IACD,IAAI,GAAG,IAAI,OAAO,GAAG,KAAK,QAAQ,EAAE,CAAC;QACnC,MAAM,CAAC,GAAG,GAA+D,CAAA;QACzE,MAAM,KAAK,GAAa,EAAE,CAAA;QAC1B,IAAI,OAAO,CAAC,CAAC,OAAO,KAAK,QAAQ,IAAI,CAAC,CAAC,OAAO,CAAC,IAAI,EAAE;YAAE,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,OAAO,CAAC,IAAI,EAAE,CAAC,CAAA;QACnF,IAAI,OAAO,CAAC,CAAC,OAAO,KAAK,QAAQ,IAAI,CAAC,CAAC,OAAO,CAAC,IAAI,EAAE;YAAE,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,OAAO,CAAC,IAAI,EAAE,CAAC,CAAA;QACnF,IAAI,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YACrB,MAAM,IAAI,GAAG,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ,IAAI,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,CAAC,EAAE,CAAA;YACvE,OAAO,GAAG,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,IAAI,EAAE,CAAA;QACrC,CAAC;QACD,IAAI,CAAC;YACH,MAAM,IAAI,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,CAAA;YAChC,IAAI,IAAI,IAAI,IAAI,KAAK,IAAI,IAAI,IAAI,KAAK,MAAM;gBAAE,OAAO,IAAI,CAAA;QAC3D,CAAC;QAAC,MAAM,CAAC;YACP,kBAAkB;QACpB,CAAC;IACH,CAAC;IACD,OAAO,MAAM,CAAC,GAAG,CAAC,CAAA;AACpB,CAAC;AAED,sGAAsG;AACtG,MAAM,CAAC,MAAM,4BAA4B,GAAwB,IAAI,GAAG,CAAC;IACvE,sBAAsB;IACtB,oBAAoB;IACpB,yBAAyB;IACzB,kCAAkC;IAClC,kBAAkB;IAClB,0BAA0B;IAC1B,wBAAwB;IACxB,YAAY;CACb,CAAC,CAAA;AAEF,MAAM,UAAU,sBAAsB,CAAC,GAAY;IACjD,MAAM,IAAI,GAAI,GAA6C,EAAE,IAAI,CAAA;IACjE,OAAO,OAAO,IAAI,KAAK,QAAQ,IAAI,4BAA4B,CAAC,GAAG,CAAC,IAAI,CAAC,CAAA;AAC3E,CAAC;AAED,iEAAiE;AACjE,MAAM,UAAU,cAAc,CAAC,IAAe,EAAE,OAAe,EAAE,OAAgB;IAC/E,OAAO,IAAI,eAAe,CAAC,IAAI,EAAE,OAAO,EAAE,OAAO,CAAC,CAAA;AACpD,CAAC"}

package/dist/runtime/llamacpp/probe.d.ts

@@ -0,0 +1,13 @@
+/**
+ * Capability probe: does the installed `llama-server` advertise a speculative type in its `-h`
+ * output? Port of `check_spec_type_support` in the plugin's `commands.rs` (5 s budget, substring
+ * match over stdout+stderr).
+ */
+import type { ManagedProcess, SpawnSpec } from '../shared/index.js';
+export declare const DFLASH_SPEC_TYPE = "draft-dflash";
+export declare const PROBE_TIMEOUT_MS = 5000;
+export interface ProbeDeps {
+    spawn?: (spec: SpawnSpec) => ManagedProcess;
+    timeoutMs?: number;
+}
+export declare function checkSpecTypeSupport(exePath: string, specType: string, env: Record<string, string>, cwd: string | undefined, deps?: ProbeDeps): Promise<boolean>;

package/dist/runtime/llamacpp/probe.js

@@ -0,0 +1,35 @@
+/**
+ * Capability probe: does the installed `llama-server` advertise a speculative type in its `-h`
+ * output? Port of `check_spec_type_support` in the plugin's `commands.rs` (5 s budget, substring
+ * match over stdout+stderr).
+ */
+import { AtomicCoreError } from '../../contracts/index.js';
+import { spawnManaged } from '../shared/index.js';
+export const DFLASH_SPEC_TYPE = 'draft-dflash';
+export const PROBE_TIMEOUT_MS = 5000;
+export async function checkSpecTypeSupport(exePath, specType, env, cwd, deps = {}) {
+    const spawn = deps.spawn ?? spawnManaged;
+    const timeoutMs = deps.timeoutMs ?? PROBE_TIMEOUT_MS;
+    let proc;
+    try {
+        proc = spawn({ exe: exePath, args: ['-h'], env, cwd });
+    }
+    catch (e) {
+        throw new AtomicCoreError('MODEL_LOAD_FAILED', 'Could not probe llama.cpp backend capabilities.', e.message);
+    }
+    const timeout = new Promise((r) => setTimeout(() => r('timeout'), timeoutMs).unref());
+    const outcome = await Promise.race([proc.exited.then(() => 'exited'), timeout]);
+    if (outcome === 'timeout') {
+        await proc.terminate(500);
+        throw new AtomicCoreError('MODEL_LOAD_TIMED_OUT', 'Timed out while probing llama.cpp backend capabilities.', `llama-server -h did not finish within ${Math.round(timeoutMs / 1000)}s for ${exePath}`);
+    }
+    const failure = proc.spawnFailure();
+    if (failure) {
+        throw new AtomicCoreError('MODEL_LOAD_FAILED', 'Could not probe llama.cpp backend capabilities.', failure.message);
+    }
+    // let the line readers drain
+    await new Promise((r) => setTimeout(r, 10));
+    const { stdout, stderr } = proc.output();
+    return `${stdout}\n${stderr}`.includes(specType);
+}
+//# sourceMappingURL=probe.js.map

package/dist/runtime/llamacpp/probe.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"probe.js","sourceRoot":"","sources":["../../../src/runtime/llamacpp/probe.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,OAAO,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAA;AAC1D,OAAO,EAAE,YAAY,EAAE,MAAM,oBAAoB,CAAA;AAGjD,MAAM,CAAC,MAAM,gBAAgB,GAAG,cAAc,CAAA;AAC9C,MAAM,CAAC,MAAM,gBAAgB,GAAG,IAAI,CAAA;AAOpC,MAAM,CAAC,KAAK,UAAU,oBAAoB,CACxC,OAAe,EACf,QAAgB,EAChB,GAA2B,EAC3B,GAAuB,EACvB,OAAkB,EAAE;IAEpB,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,IAAI,YAAY,CAAA;IACxC,MAAM,SAAS,GAAG,IAAI,CAAC,SAAS,IAAI,gBAAgB,CAAA;IACpD,IAAI,IAAoB,CAAA;IACxB,IAAI,CAAC;QACH,IAAI,GAAG,KAAK,CAAC,EAAE,GAAG,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,IAAI,CAAC,EAAE,GAAG,EAAE,GAAG,EAAE,CAAC,CAAA;IACxD,CAAC;IAAC,OAAO,CAAC,EAAE,CAAC;QACX,MAAM,IAAI,eAAe,CACvB,mBAAmB,EACnB,iDAAiD,EAChD,CAAW,CAAC,OAAO,CACrB,CAAA;IACH,CAAC;IACD,MAAM,OAAO,GAAG,IAAI,OAAO,CAAY,CAAC,CAAC,EAAE,EAAE,CAAC,UAAU,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC,EAAE,SAAS,CAAC,CAAC,KAAK,EAAE,CAAC,CAAA;IAChG,MAAM,OAAO,GAAG,MAAM,OAAO,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,EAAE,CAAC,QAAiB,CAAC,EAAE,OAAO,CAAC,CAAC,CAAA;IACxF,IAAI,OAAO,KAAK,SAAS,EAAE,CAAC;QAC1B,MAAM,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,CAAA;QACzB,MAAM,IAAI,eAAe,CACvB,sBAAsB,EACtB,yDAAyD,EACzD,yCAAyC,IAAI,CAAC,KAAK,CAAC,SAAS,GAAG,IAAI,CAAC,SAAS,OAAO,EAAE,CACxF,CAAA;IACH,CAAC;IACD,MAAM,OAAO,GAAG,IAAI,CAAC,YAAY,EAAE,CAAA;IACnC,IAAI,OAAO,EAAE,CAAC;QACZ,MAAM,IAAI,eAAe,CACvB,mBAAmB,EACnB,iDAAiD,EACjD,OAAO,CAAC,OAAO,CAChB,CAAA;IACH,CAAC;IACD,6BAA6B;IAC7B,MAAM,IAAI,OAAO,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,UAAU,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAA;IAC3C,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,GAAG,IAAI,CAAC,MAAM,EAAE,CAAA;IACxC,OAAO,GAAG,MAAM,KAAK,MAAM,EAAE,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAA;AAClD,CAAC"}

package/dist/runtime/llamacpp/runtime-device.d.ts

@@ -0,0 +1,33 @@
+/**
+ * Which device the loaded model actually runs on, derived from the `llama-server` startup log.
+ * Port of `src-tauri/plugins/tauri-plugin-llamacpp-upstream/src/runtime_device.rs`.
+ * Pinned by `test/contract/runtime-device.test.ts`.
+ *
+ * `--list-devices` only says what a binary can enumerate; the startup log is the only signal that
+ * tells a healthy GPU load from a CUDA/Vulkan build that silently degraded to CPU.
+ */
+import type { RuntimeDeviceInfo } from '../../contracts/index.js';
+export declare function parseDeviceInitError(line: string): string | undefined;
+/** `load_backend: loaded CUDA backend from /path/libggml-cuda.so` → `CUDA` */
+export declare function parseLoadedBackend(line: string): string | undefined;
+/** `load_tensors: offloaded 33/33 layers to GPU` → [33, 33] */
+export declare function parseOffloadedLayers(line: string): [number, number] | undefined;
+/** `load_tensors: offloading 32 repeating layers to GPU` → 32 */
+export declare function parseRepeatingLayers(line: string): number | undefined;
+export declare function parseSize(text: string): number | undefined;
+/** `load_tensors:        CUDA0 model buffer size =  4155.99 MiB` → ['CUDA0', bytes] */
+export declare function parseModelBuffer(line: string): [string, number] | undefined;
+export declare class RuntimeDeviceAccumulator {
+    private loadedBackends;
+    private buffers;
+    private gpuLayersOffloaded;
+    private totalLayers;
+    private cudaRuntimeMissing;
+    private deviceInitError;
+    /** Recorded at spawn time by the CUDA-path probe, before any log line. */
+    markCudaRuntimeMissing(): void;
+    ingest(line: string): void;
+    snapshot(): RuntimeDeviceInfo;
+}
+/** True when the log carried nothing recognisable; callers must not conclude anything from it. */
+export declare function isInconclusive(info: RuntimeDeviceInfo): boolean;

package/dist/runtime/llamacpp/runtime-device.js

@@ -0,0 +1,169 @@
+/**
+ * Which device the loaded model actually runs on, derived from the `llama-server` startup log.
+ * Port of `src-tauri/plugins/tauri-plugin-llamacpp-upstream/src/runtime_device.rs`.
+ * Pinned by `test/contract/runtime-device.test.ts`.
+ *
+ * `--list-devices` only says what a binary can enumerate; the startup log is the only signal that
+ * tells a healthy GPU load from a CUDA/Vulkan build that silently degraded to CPU.
+ */
+import { floatAsU64, parseRustF64, parseRustI32 } from '../../util/index.js';
+const DEVICE_INIT_ERROR_MARKERS = [
+    'failed to initialize CUDA',
+    'no CUDA devices found',
+    'error while loading shared libraries',
+    'failed to load backend',
+    'ggml_vulkan: No devices found',
+    'no usable GPU found',
+];
+export function parseDeviceInitError(line) {
+    const trimmed = line.trim();
+    return DEVICE_INIT_ERROR_MARKERS.some((m) => trimmed.includes(m)) ? trimmed : undefined;
+}
+/** `load_backend: loaded CUDA backend from /path/libggml-cuda.so` → `CUDA` */
+export function parseLoadedBackend(line) {
+    const i = line.indexOf('load_backend: loaded ');
+    if (i < 0)
+        return undefined;
+    const rest = line.slice(i + 'load_backend: loaded '.length);
+    const j = rest.indexOf(' backend');
+    if (j < 0)
+        return undefined;
+    const name = rest.slice(0, j).trim();
+    return name === '' ? undefined : name;
+}
+/** `load_tensors: offloaded 33/33 layers to GPU` → [33, 33] */
+export function parseOffloadedLayers(line) {
+    if (!line.includes('layers to GPU'))
+        return undefined;
+    const i = line.indexOf('offloaded ');
+    if (i < 0)
+        return undefined;
+    const ratio = line
+        .slice(i + 'offloaded '.length)
+        .split(/\s+/)
+        .find((t) => t !== '');
+    if (ratio === undefined)
+        return undefined;
+    const slash = ratio.indexOf('/');
+    if (slash < 0)
+        return undefined;
+    const offloaded = parseRustI32(ratio.slice(0, slash).trim());
+    const total = parseRustI32(ratio.slice(slash + 1).trim());
+    return offloaded === undefined || total === undefined ? undefined : [offloaded, total];
+}
+/** `load_tensors: offloading 32 repeating layers to GPU` → 32 */
+export function parseRepeatingLayers(line) {
+    if (!line.includes('repeating layers to GPU'))
+        return undefined;
+    const i = line.indexOf('offloading ');
+    if (i < 0)
+        return undefined;
+    const first = line
+        .slice(i + 'offloading '.length)
+        .split(/\s+/)
+        .find((t) => t !== '');
+    return first === undefined ? undefined : parseRustI32(first);
+}
+export function parseSize(text) {
+    const parts = text.split(/\s+/).filter((t) => t !== '');
+    const value = parts[0] === undefined ? undefined : parseRustF64(parts[0]);
+    if (value === undefined)
+        return undefined;
+    const unit = parts[1] ?? 'B';
+    const multiplier = unit === 'GiB' ? 1024 ** 3 : unit === 'MiB' ? 1024 ** 2 : unit === 'KiB' ? 1024 : 1;
+    if (value < 0)
+        return undefined;
+    return floatAsU64(value * multiplier);
+}
+/** `load_tensors:        CUDA0 model buffer size =  4155.99 MiB` → ['CUDA0', bytes] */
+export function parseModelBuffer(line) {
+    const i = line.indexOf(' model buffer size =');
+    if (i < 0)
+        return undefined;
+    const head = line.slice(0, i);
+    const tail = line.slice(i + ' model buffer size ='.length);
+    if (!head.includes('load_tensors:'))
+        return undefined;
+    const colon = head.lastIndexOf(':');
+    if (colon < 0)
+        return undefined;
+    const label = head.slice(colon + 1).trim();
+    if (label === '')
+        return undefined;
+    const bytes = parseSize(tail);
+    return bytes === undefined ? undefined : [label, bytes];
+}
+const isCpuBufferLabel = (label) => label === 'CPU' || label.startsWith('CPU_');
+export class RuntimeDeviceAccumulator {
+    loadedBackends = [];
+    buffers = new Map();
+    gpuLayersOffloaded = null;
+    totalLayers = null;
+    cudaRuntimeMissing = false;
+    deviceInitError = null;
+    /** Recorded at spawn time by the CUDA-path probe, before any log line. */
+    markCudaRuntimeMissing() {
+        this.cudaRuntimeMissing = true;
+    }
+    ingest(line) {
+        if (this.deviceInitError === null) {
+            const err = parseDeviceInitError(line);
+            if (err !== undefined)
+                this.deviceInitError = err;
+        }
+        const backend = parseLoadedBackend(line);
+        if (backend !== undefined) {
+            if (!this.loadedBackends.includes(backend))
+                this.loadedBackends.push(backend);
+            return;
+        }
+        const offloaded = parseOffloadedLayers(line);
+        if (offloaded !== undefined) {
+            this.gpuLayersOffloaded = offloaded[0];
+            this.totalLayers = offloaded[1];
+            return;
+        }
+        const repeating = parseRepeatingLayers(line);
+        if (repeating !== undefined) {
+            if (this.gpuLayersOffloaded === null)
+                this.gpuLayersOffloaded = repeating;
+            return;
+        }
+        const buffer = parseModelBuffer(line);
+        if (buffer !== undefined) {
+            const [label, bytes] = buffer;
+            this.buffers.set(label, Math.max(this.buffers.get(label) ?? 0, bytes));
+        }
+    }
+    snapshot() {
+        let largest;
+        for (const [label, bytes] of this.buffers) {
+            if (isCpuBufferLabel(label) || bytes <= 0)
+                continue;
+            // Larger wins; on a tie the lexicographically smaller label wins (deterministic splits).
+            if (largest === undefined || bytes > largest[1] || (bytes === largest[1] && label < largest[0])) {
+                largest = [label, bytes];
+            }
+        }
+        const nothingOffloaded = this.gpuLayersOffloaded === 0;
+        const primaryDevice = largest !== undefined && !nothingOffloaded ? largest[0] : this.buffers.size > 0 ? 'CPU' : '';
+        return {
+            loaded_backends: [...this.loadedBackends],
+            primary_device: primaryDevice,
+            gpu_layers_offloaded: this.gpuLayersOffloaded,
+            total_layers: this.totalLayers,
+            gpu_buffer_bytes: largest === undefined ? null : largest[1],
+            cuda_runtime_missing: this.cudaRuntimeMissing,
+            device_init_error: this.deviceInitError,
+        };
+    }
+}
+/** True when the log carried nothing recognisable; callers must not conclude anything from it. */
+export function isInconclusive(info) {
+    return (info.loaded_backends.length === 0 &&
+        info.primary_device === '' &&
+        info.gpu_layers_offloaded === null &&
+        !info.cuda_runtime_missing &&
+        info.device_init_error === null);
+}
+//# sourceMappingURL=runtime-device.js.map

package/dist/runtime/llamacpp/runtime-device.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"runtime-device.js","sourceRoot":"","sources":["../../../src/runtime/llamacpp/runtime-device.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAGH,OAAO,EAAE,UAAU,EAAE,YAAY,EAAE,YAAY,EAAE,MAAM,qBAAqB,CAAA;AAE5E,MAAM,yBAAyB,GAAG;IAChC,2BAA2B;IAC3B,uBAAuB;IACvB,sCAAsC;IACtC,wBAAwB;IACxB,+BAA+B;IAC/B,qBAAqB;CACtB,CAAA;AAED,MAAM,UAAU,oBAAoB,CAAC,IAAY;IAC/C,MAAM,OAAO,GAAG,IAAI,CAAC,IAAI,EAAE,CAAA;IAC3B,OAAO,yBAAyB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,OAAO,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,SAAS,CAAA;AACzF,CAAC;AAED,8EAA8E;AAC9E,MAAM,UAAU,kBAAkB,CAAC,IAAY;IAC7C,MAAM,CAAC,GAAG,IAAI,CAAC,OAAO,CAAC,uBAAuB,CAAC,CAAA;IAC/C,IAAI,CAAC,GAAG,CAAC;QAAE,OAAO,SAAS,CAAA;IAC3B,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,GAAG,uBAAuB,CAAC,MAAM,CAAC,CAAA;IAC3D,MAAM,CAAC,GAAG,IAAI,CAAC,OAAO,CAAC,UAAU,CAAC,CAAA;IAClC,IAAI,CAAC,GAAG,CAAC;QAAE,OAAO,SAAS,CAAA;IAC3B,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,EAAE,CAAA;IACpC,OAAO,IAAI,KAAK,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,IAAI,CAAA;AACvC,CAAC;AAED,+DAA+D;AAC/D,MAAM,UAAU,oBAAoB,CAAC,IAAY;IAC/C,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,eAAe,CAAC;QAAE,OAAO,SAAS,CAAA;IACrD,MAAM,CAAC,GAAG,IAAI,CAAC,OAAO,CAAC,YAAY,CAAC,CAAA;IACpC,IAAI,CAAC,GAAG,CAAC;QAAE,OAAO,SAAS,CAAA;IAC3B,MAAM,KAAK,GAAG,IAAI;SACf,KAAK,CAAC,CAAC,GAAG,YAAY,CAAC,MAAM,CAAC;SAC9B,KAAK,CAAC,KAAK,CAAC;SACZ,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,KAAK,EAAE,CAAC,CAAA;IACxB,IAAI,KAAK,KAAK,SAAS;QAAE,OAAO,SAAS,CAAA;IACzC,MAAM,KAAK,GAAG,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,CAAA;IAChC,IAAI,KAAK,GAAG,CAAC;QAAE,OAAO,SAAS,CAAA;IAC/B,MAAM,SAAS,GAAG,YAAY,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,EAAE,KAAK,CAAC,CAAC,IAAI,EAAE,CAAC,CAAA;IAC5D,MAAM,KAAK,GAAG,YAAY,CAAC,KAAK,CAAC,KAAK,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAA;IACzD,OAAO,SAAS,KAAK,SAAS,IAAI,KAAK,KAAK,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,SAAS,EAAE,KAAK,CAAC,CAAA;AACxF,CAAC;AAED,iEAAiE;AACjE,MAAM,UAAU,oBAAoB,CAAC,IAAY;IAC/C,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,yBAAyB,CAAC;QAAE,OAAO,SAAS,CAAA;IAC/D,MAAM,CAAC,GAAG,IAAI,CAAC,OAAO,CAAC,aAAa,CAAC,CAAA;IACrC,IAAI,CAAC,GAAG,CAAC;QAAE,OAAO,SAAS,CAAA;IAC3B,MAAM,KAAK,GAAG,IAAI;SACf,KAAK,CAAC,CAAC,GAAG,aAAa,CAAC,MAAM,CAAC;SAC/B,KAAK,CAAC,KAAK,CAAC;SACZ,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,KAAK,EAAE,CAAC,CAAA;IACxB,OAAO,KAAK,KAAK,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,YAAY,CAAC,KAAK,CAAC,CAAA;AAC9D,CAAC;AAED,MAAM,UAAU,SAAS,CAAC,IAAY;IACpC,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,KAAK,EAAE,CAAC,CAAA;IACvD,MAAM,KAAK,GAAG,KAAK,CAAC,CAAC,CAAC,KAAK,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,YAAY,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAA;IACzE,IAAI,KAAK,KAAK,SAAS;QAAE,OAAO,SAAS,CAAA;IACzC,MAAM,IAAI,GAAG,KAAK,CAAC,CAAC,CAAC,IAAI,GAAG,CAAA;IAC5B,MAAM,UAAU,GAAG,IAAI,KAAK,KAAK,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,KAAK,KAAK,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,KAAK,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAA;IACtG,IAAI,KAAK,GAAG,CAAC;QAAE,OAAO,SAAS,CAAA;IAC/B,OAAO,UAAU,CAAC,KAAK,GAAG,UAAU,CAAC,CAAA;AACvC,CAAC;AAED,uFAAuF;AACvF,MAAM,UAAU,gBAAgB,CAAC,IAAY;IAC3C,MAAM,CAAC,GAAG,IAAI,CAAC,OAAO,CAAC,sBAAsB,CAAC,CAAA;IAC9C,IAAI,CAAC,GAAG,CAAC;QAAE,OAAO,SAAS,CAAA;IAC3B,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAA;IAC7B,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,GAAG,sBAAsB,CAAC,MAAM,CAAC,CAAA;IAC1D,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,eAAe,CAAC;QAAE,OAAO,SAAS,CAAA;IACrD,MAAM,KAAK,GAAG,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,CAAA;IACnC,IAAI,KAAK,GAAG,CAAC;QAAE,OAAO,SAAS,CAAA;IAC/B,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,IAAI,EAAE,CAAA;IAC1C,IAAI,KAAK,KAAK,EAAE;QAAE,OAAO,SAAS,CAAA;IAClC,MAAM,KAAK,GAAG,SAAS,CAAC,IAAI,CAAC,CAAA;IAC7B,OAAO,KAAK,KAAK,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,KAAK,EAAE,KAAK,CAAC,CAAA;AACzD,CAAC;AAED,MAAM,gBAAgB,GAAG,CAAC,KAAa,EAAE,EAAE,CAAC,KAAK,KAAK,KAAK,IAAI,KAAK,CAAC,UAAU,CAAC,MAAM,CAAC,CAAA;AAEvF,MAAM,OAAO,wBAAwB;IAC3B,cAAc,GAAa,EAAE,CAAA;IAC7B,OAAO,GAAG,IAAI,GAAG,EAAkB,CAAA;IACnC,kBAAkB,GAAkB,IAAI,CAAA;IACxC,WAAW,GAAkB,IAAI,CAAA;IACjC,kBAAkB,GAAG,KAAK,CAAA;IAC1B,eAAe,GAAkB,IAAI,CAAA;IAE7C,0EAA0E;IAC1E,sBAAsB;QACpB,IAAI,CAAC,kBAAkB,GAAG,IAAI,CAAA;IAChC,CAAC;IAED,MAAM,CAAC,IAAY;QACjB,IAAI,IAAI,CAAC,eAAe,KAAK,IAAI,EAAE,CAAC;YAClC,MAAM,GAAG,GAAG,oBAAoB,CAAC,IAAI,CAAC,CAAA;YACtC,IAAI,GAAG,KAAK,SAAS;gBAAE,IAAI,CAAC,eAAe,GAAG,GAAG,CAAA;QACnD,CAAC;QACD,MAAM,OAAO,GAAG,kBAAkB,CAAC,IAAI,CAAC,CAAA;QACxC,IAAI,OAAO,KAAK,SAAS,EAAE,CAAC;YAC1B,IAAI,CAAC,IAAI,CAAC,cAAc,CAAC,QAAQ,CAAC,OAAO,CAAC;gBAAE,IAAI,CAAC,cAAc,CAAC,IAAI,CAAC,OAAO,CAAC,CAAA;YAC7E,OAAM;QACR,CAAC;QACD,MAAM,SAAS,GAAG,oBAAoB,CAAC,IAAI,CAAC,CAAA;QAC5C,IAAI,SAAS,KAAK,SAAS,EAAE,CAAC;YAC5B,IAAI,CAAC,kBAAkB,GAAG,SAAS,CAAC,CAAC,CAAC,CAAA;YACtC,IAAI,CAAC,WAAW,GAAG,SAAS,CAAC,CAAC,CAAC,CAAA;YAC/B,OAAM;QACR,CAAC;QACD,MAAM,SAAS,GAAG,oBAAoB,CAAC,IAAI,CAAC,CAAA;QAC5C,IAAI,SAAS,KAAK,SAAS,EAAE,CAAC;YAC5B,IAAI,IAAI,CAAC,kBAAkB,KAAK,IAAI;gBAAE,IAAI,CAAC,kBAAkB,GAAG,SAAS,CAAA;YACzE,OAAM;QACR,CAAC;QACD,MAAM,MAAM,GAAG,gBAAgB,CAAC,IAAI,CAAC,CAAA;QACrC,IAAI,MAAM,KAAK,SAAS,EAAE,CAAC;YACzB,MAAM,CAAC,KAAK,EAAE,KAAK,CAAC,GAAG,MAAM,CAAA;YAC7B,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,KAAK,EAAE,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,KAAK,CAAC,CAAC,CAAA;QACxE,CAAC;IACH,CAAC;IAED,QAAQ;QACN,IAAI,OAAqC,CAAA;QACzC,KAAK,MAAM,CAAC,KAAK,EAAE,KAAK,CAAC,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YAC1C,IAAI,gBAAgB,CAAC,KAAK,CAAC,IAAI,KAAK,IAAI,CAAC;gBAAE,SAAQ;YACnD,yFAAyF;YACzF,IAAI,OAAO,KAAK,SAAS,IAAI,KAAK,GAAG,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,KAAK,OAAO,CAAC,CAAC,CAAC,IAAI,KAAK,GAAG,OAAO,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;gBAChG,OAAO,GAAG,CAAC,KAAK,EAAE,KAAK,CAAC,CAAA;YAC1B,CAAC;QACH,CAAC;QACD,MAAM,gBAAgB,GAAG,IAAI,CAAC,kBAAkB,KAAK,CAAC,CAAA;QACtD,MAAM,aAAa,GACjB,OAAO,KAAK,SAAS,IAAI,CAAC,gBAAgB,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,OAAO,CAAC,IAAI,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAA;QAC9F,OAAO;YACL,eAAe,EAAE,CAAC,GAAG,IAAI,CAAC,cAAc,CAAC;YACzC,cAAc,EAAE,aAAa;YAC7B,oBAAoB,EAAE,IAAI,CAAC,kBAAkB;YAC7C,YAAY,EAAE,IAAI,CAAC,WAAW;YAC9B,gBAAgB,EAAE,OAAO,KAAK,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC;YAC3D,oBAAoB,EAAE,IAAI,CAAC,kBAAkB;YAC7C,iBAAiB,EAAE,IAAI,CAAC,eAAe;SACxC,CAAA;IACH,CAAC;CACF;AAED,kGAAkG;AAClG,MAAM,UAAU,cAAc,CAAC,IAAuB;IACpD,OAAO,CACL,IAAI,CAAC,eAAe,CAAC,MAAM,KAAK,CAAC;QACjC,IAAI,CAAC,cAAc,KAAK,EAAE;QAC1B,IAAI,CAAC,oBAAoB,KAAK,IAAI;QAClC,CAAC,IAAI,CAAC,oBAAoB;QAC1B,IAAI,CAAC,iBAAiB,KAAK,IAAI,CAChC,CAAA;AACH,CAAC"}

package/dist/runtime/llamacpp/runtime.d.ts

@@ -0,0 +1,151 @@
+/**
+ * The live half of the llama.cpp runtime: it turns a load plan into a running `llama-server` and
+ * keeps the session table the app's `model-factory.ts`, the Rust agent and `/v1` all read
+ * (PLAN.md §3.2, §3.3). Policy lives in `load-plan.ts` and `args.ts`; this file owns processes.
+ *
+ * What it guarantees:
+ *  - a session exists only while its process does — the exit watcher removes it and emits
+ *    `session:died`, so nothing dials a dead port;
+ *  - every spawned process is journalled *before* it can be forgotten, so a crashed owner leaves a
+ *    trail the next owner can clean up (PLAN.md §3.4);
+ *  - the two post-failure retries of `performLoad` (drop mmproj, drop MTP) happen here, once each;
+ *  - stderr is fed to the device accumulator while loading, which is the only place llama.cpp says
+ *    which device it actually used;
+ *  - a load the user cancels stops wherever it is — queued, planning, spawning or just ready — kills
+ *    what it started at once, and never falls into a retry.
+ */
+import type { CoreEvents, DeviceInfo, LocalProviderId, RuntimeDeviceInfo, SessionInfo, UnloadResult } from '../../contracts/index.js';
+import type { DataLayout } from '../../config/index.js';
+import type { ProcessJournal } from '../../lock/index.js';
+import type { ModelRegistry } from '../../models/index.js';
+import { spawnAndAwaitReady } from '../shared/index.js';
+import type { ManagedProcess, SpawnSpec, CtxIncreaseResult, LocalRuntime, RecreateResult } from '../shared/index.js';
+import type { LlamacppConfigInput } from './args.js';
+import type { LlamacppEngineSettings, LoadPlanDeps } from './load-plan.js';
+export type EmitFn = <K extends keyof CoreEvents>(name: K, payload: CoreEvents[K]) => void;
+export interface RuntimeSettings {
+    config: LlamacppConfigInput;
+    engine: LlamacppEngineSettings;
+}
+export type { CtxIncreaseResult } from '../shared/index.js';
+export interface LoadOptions {
+    /** Per-model overrides, canonical keys (the `settings` argument of the extension's `load()`). */
+    overrides?: Partial<LlamacppConfigInput>;
+    isEmbedding?: boolean;
+    /** Use this executable instead of resolving one (CLI `--bin`). */
+    exePath?: string;
+    /** Feature-gating tag for an explicit executable; neutral when omitted. */
+    versionBackend?: string;
+    /** Explicit model/projector paths bypass `model.yml` resolution (CLI flags). */
+    modelPath?: string;
+    mmprojPath?: string;
+    /** Exact readiness timeout for CLI parity; provider loads keep the application floor. */
+    timeoutSecs?: number;
+    /** Append backend stdout/stderr to this path for the lifetime of the session. */
+    logPath?: string;
+    /** Relay backend stdout/stderr through authenticated `core:log` events. */
+    verbose?: boolean;
+    /** Bind the server to this port instead of a random free one (CLI `--port`). */
+    port?: number;
+    /** Skip the auto-unload of other text models. */
+    bypassAutoUnload?: boolean;
+    /** Aborted when the user cancels this load; the load rejects with `MODEL_LOAD_CANCELLED`. */
+    signal?: AbortSignal;
+}
+export interface LlamacppRuntimeOptions {
+    layout: DataLayout;
+    registry: ModelRegistry;
+    /** Owner instance that will be recorded as the parent of every spawned process. */
+    instanceId: string;
+    provider?: LocalProviderId;
+    journal?: ProcessJournal | undefined;
+    emit?: EmitFn;
+    platform?: NodeJS.Platform;
+    baseEnv?: NodeJS.ProcessEnv;
+    fetch?: typeof fetch;
+    /** Provider settings for a load; re-read per load so a settings change lands on the next one. */
+    readSettings: () => Promise<RuntimeSettings>;
+    /** Resolve `<version>/<backend>` to an installed executable (the backend service in a full core). */
+    ensureBackendReady?: LoadPlanDeps['ensureBackendReady'] | undefined;
+    resolveLatestBackend?: LoadPlanDeps['resolveLatestBackend'] | undefined;
+    cpuInfo?: LoadPlanDeps['cpuInfo'] | undefined;
+    ensureGemmaMtpDraft?: LoadPlanDeps['ensureGemmaMtpDraft'] | undefined;
+    ensureDflashDraft?: LoadPlanDeps['ensureDflashDraft'] | undefined;
+    /** Model that must never be auto-unloaded. */
+    transcriptionModelId?: string;
+    /**
+     * Read a model's GGUF metadata. A seam because the model's trained context comes from here, and
+     * it is what decides when the context ladder has nowhere left to climb — untestable otherwise
+     * without hand-building a GGUF file.
+     */
+    readGgufMetadata?: LoadPlanDeps['readGgufMetadata'] | undefined;
+    /** Test seam. */
+    spawn?: typeof spawnAndAwaitReady | undefined;
+    probeDevicesWith?: ((spec: SpawnSpec) => ManagedProcess) | undefined;
+}
+export declare const DEVICE_PROBE_TIMEOUT_MS = 10000;
+export declare class LlamacppRuntime implements LocalRuntime {
+    private readonly options;
+    private readonly sessions;
+    private readonly loading;
+    private loadTail;
+    private readonly shutdownController;
+    private closing;
+    private shutdownPromise;
+    private readonly provider;
+    private readonly platform;
+    private readonly emit;
+    private readonly fetchImpl;
+    constructor(options: LlamacppRuntimeOptions);
+    list(): SessionInfo[];
+    findSession(modelId: string): SessionInfo | undefined;
+    getLoadedModels(): string[];
+    getRuntimeDeviceInfo(modelId: string): RuntimeDeviceInfo | undefined;
+    getMaxCtxTrain(modelId: string): number | undefined;
+    isLoading(modelId: string): boolean;
+    /** Load a model, or join the load already in flight for it. */
+    load(modelId: string, opts?: LoadOptions): Promise<SessionInfo>;
+    private loadOnce;
+    private enqueueLoad;
+    private planDeps;
+    /** Ports this runtime already handed out, so a second load never lands on the same one. */
+    private usedPorts;
+    private spawnSession;
+    private journalSpawn;
+    private watchExit;
+    private reportRuntimeDevice;
+    private healthy;
+    /** Stop one session. Unloading something that is not loaded is not an error. */
+    unload(modelId: string): Promise<UnloadResult>;
+    /** The context window a loaded session is actually running with. */
+    getCtxSize(modelId: string): number | undefined;
+    /**
+     * Reload a model with the next context size up the ladder, because a request did not fit.
+     *
+     * Ported from the app's `auto_increase_ctx` handler, including the two cases where it declines:
+     *
+     * - `fit`: with fit on, llama.cpp sizes the window itself at load and `--ctx-size` is not even
+     *   emitted, so a reload would change nothing and cost the user a model load.
+     * - `at_max`: the ladder has reached what the model was trained for. Reloading at the same size
+     *   would loop — the next request would overflow again and ask again.
+     *
+     * The unload is allowed to fail: a session that is already gone is still a reload candidate, and
+     * refusing to reload because the corpse would not die is worse than reloading.
+     */
+    autoIncreaseCtx(modelId: string, reason?: string): Promise<CtxIncreaseResult>;
+    /**
+     * Reload a model at the context it already has, because its engine is unusable: a fatal compute
+     * failure (a Metal OOM during prompt processing) leaves the ggml backend in an error state that
+     * only a new process clears. The window is deliberately not grown — more context would only make
+     * an out-of-memory failure more likely. As with `autoIncreaseCtx`, a failed unload does not stop
+     * the reload.
+     */
+    recreateSession(modelId: string): Promise<RecreateResult>;
+    private unloadSession;
+    unloadAll(): Promise<void>;
+    /** `llama-server --list-devices` against an installed backend. */
+    getDevices(exePath: string): Promise<DeviceInfo[]>;
+    shutdown(): Promise<void>;
+    dispose(): Promise<void>;
+    private assertRunning;
+}