@atomic-chat/core 0.5.1

package/dist/runtime/llamacpp/errors.js

@@ -0,0 +1,137 @@
+/**
+ * Classification of a failed `llama-server` process into the app's error contract. Port of
+ * `src-tauri/plugins/tauri-plugin-llamacpp-upstream/src/error.rs` (`LlamacppError::from_stderr`,
+ * `from_exit_status`, `from_process_output`). Pinned by `test/contract/errors.test.ts`. The
+ * TurboQuant fork's `error.rs` differs in one rule, `wrong number of tensors` (app commit
+ * `ec1fd3ea7`), so the provider is a parameter.
+ *
+ * The cascade is ordered: the first matching group wins. Messages are the app's user-facing strings
+ * and are part of the contract. `details` is always set (the raw stream), even when empty.
+ */
+import { AtomicCoreError } from '../../contracts/index.js';
+const SIGNAL_NUMBERS = {
+    SIGHUP: 1,
+    SIGINT: 2,
+    SIGQUIT: 3,
+    SIGILL: 4,
+    SIGTRAP: 5,
+    SIGABRT: 6,
+    SIGBUS: 7,
+    SIGFPE: 8,
+    SIGKILL: 9,
+    SIGUSR1: 10,
+    SIGSEGV: 11,
+    SIGUSR2: 12,
+    SIGPIPE: 13,
+    SIGALRM: 14,
+    SIGTERM: 15,
+};
+/** Normalise a Node signal name to the Unix number Rust sees. */
+export function signalNumber(signal) {
+    if (signal === null)
+        return null;
+    if (typeof signal === 'number')
+        return signal;
+    return SIGNAL_NUMBERS[signal] ?? null;
+}
+/** Windows: STATUS_ACCESS_VIOLATION / STATUS_STACK_OVERFLOW / STATUS_STACK_BUFFER_OVERRUN. */
+const WINDOWS_CRASH_CODES = new Set([0xc000_0005, 0xc000_00fd, 0xc000_0409]);
+/** A hard native crash (segfault / abort) rather than a normal non-zero exit. */
+export function isCrashExit(exit, platform) {
+    if (platform === 'win32') {
+        return exit.code !== null && WINDOWS_CRASH_CODES.has(exit.code >>> 0);
+    }
+    const sig = signalNumber(exit.signal);
+    return sig === 11 || sig === 6;
+}
+const includesAny = (lower, needles) => needles.some((n) => lower.includes(n));
+/**
+ * TurboQuant (`llamacpp`) can know an architecture and still expect a tensor layout that differs
+ * from a valid upstream GGUF: re-downloading cannot fix that, and the same file loads in stock
+ * llama.cpp, so there the mismatch is an unsupported architecture. Upstream it still means a
+ * damaged file.
+ */
+const TENSOR_COUNT_MISMATCH = 'wrong number of tensors';
+const tensorLayoutIsArchitecture = (provider) => provider === 'llamacpp';
+const CASCADE = [
+    {
+        code: 'OS_VERSION_UNSUPPORTED',
+        message: "The model engine couldn't start because it requires a newer version of macOS than the one on this Mac.",
+        match: (l) => l.includes('dyld') && l.includes('symbol not found'),
+    },
+    {
+        code: 'OUT_OF_MEMORY',
+        message: 'Out of memory. The model requires more RAM or VRAM than available.',
+        match: (l) => includesAny(l, [
+            'out of memory',
+            'failed to allocate',
+            'insufficient memory',
+            'erroroutofdevicememory',
+            'kiogpucommandbuffercallbackerroroutofmemory',
+            'cuda_error_out_of_memory',
+        ]),
+    },
+    {
+        code: 'MODEL_ARCH_NOT_SUPPORTED',
+        message: "The model's architecture or format is not supported by this version of the backend.",
+        match: (l, provider) => includesAny(l, [
+            'error loading model architecture',
+            'unknown model architecture',
+            'error loading model hyperparameters',
+            'key not found in model',
+        ]) ||
+            (tensorLayoutIsArchitecture(provider) && l.includes(TENSOR_COUNT_MISMATCH)),
+    },
+    {
+        code: 'MULTIMODAL_PROJECTOR_LOAD_FAILED',
+        message: "This model's multimodal projector isn't supported by the current llama.cpp backend. Vision/audio is unavailable for this model on this backend.",
+        match: (l) => l.includes('unknown projector type'),
+    },
+    {
+        code: 'MODEL_FILE_CORRUPT',
+        message: 'The model file appears to be incomplete or corrupted. Try deleting and re-downloading the model.',
+        match: (l, provider) => includesAny(l, [
+            'corrupted or incomplete',
+            'invalid magic',
+            'unexpectedly reached end of file',
+            'failed to read tensor',
+        ]) ||
+            (!tensorLayoutIsArchitecture(provider) && l.includes(TENSOR_COUNT_MISMATCH)),
+    },
+];
+export const GENERIC_PROCESS_ERROR_MESSAGE = 'The model process encountered an unexpected error.';
+export const CRASH_MESSAGE = 'The model process crashed unexpectedly (access violation / segfault). This usually means the model is incompatible with this backend, or its speculative-decoding (MTP) configuration is unsupported here.';
+/** `LlamacppError::from_stderr`: ordered substring cascade over the lowercased stream. */
+export function classifyStderr(stderr, provider = 'llamacpp-upstream') {
+    const lower = stderr.toLowerCase();
+    for (const rule of CASCADE) {
+        if (rule.match(lower, provider))
+            return new AtomicCoreError(rule.code, rule.message, stderr);
+    }
+    return new AtomicCoreError('LLAMA_CPP_PROCESS_ERROR', GENERIC_PROCESS_ERROR_MESSAGE, stderr);
+}
+/** `LlamacppError::from_exit_status`: a recognised crash upgrades the generic error's message. */
+export function classifyExit(exit, stderr, platform, provider = 'llamacpp-upstream') {
+    const base = classifyStderr(stderr, provider);
+    if (base.code !== 'LLAMA_CPP_PROCESS_ERROR' || !isCrashExit(exit, platform))
+        return base;
+    return new AtomicCoreError('LLAMA_CPP_PROCESS_ERROR', CRASH_MESSAGE, stderr);
+}
+/**
+ * `LlamacppError::from_process_output`: stderr classification wins; otherwise try stdout (several
+ * llama.cpp builds report loader failures there); otherwise keep the exit-status error but hand the
+ * caller stdout as details when stderr is blank.
+ */
+export function classifyProcessOutput(exit, stderr, stdout, platform, provider = 'llamacpp-upstream') {
+    const base = classifyExit(exit, stderr, platform, provider);
+    if (base.code !== 'LLAMA_CPP_PROCESS_ERROR')
+        return base;
+    const fromStdout = classifyStderr(stdout, provider);
+    if (fromStdout.code !== 'LLAMA_CPP_PROCESS_ERROR')
+        return fromStdout;
+    if (stderr.trim() === '' && stdout.trim() !== '') {
+        return new AtomicCoreError(base.code, base.message, stdout);
+    }
+    return base;
+}
+//# sourceMappingURL=errors.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"errors.js","sourceRoot":"","sources":["../../../src/runtime/llamacpp/errors.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAEH,OAAO,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAA;AAU1D,MAAM,cAAc,GAA2B;IAC7C,MAAM,EAAE,CAAC;IACT,MAAM,EAAE,CAAC;IACT,OAAO,EAAE,CAAC;IACV,MAAM,EAAE,CAAC;IACT,OAAO,EAAE,CAAC;IACV,OAAO,EAAE,CAAC;IACV,MAAM,EAAE,CAAC;IACT,MAAM,EAAE,CAAC;IACT,OAAO,EAAE,CAAC;IACV,OAAO,EAAE,EAAE;IACX,OAAO,EAAE,EAAE;IACX,OAAO,EAAE,EAAE;IACX,OAAO,EAAE,EAAE;IACX,OAAO,EAAE,EAAE;IACX,OAAO,EAAE,EAAE;CACZ,CAAA;AAED,iEAAiE;AACjE,MAAM,UAAU,YAAY,CAAC,MAA8B;IACzD,IAAI,MAAM,KAAK,IAAI;QAAE,OAAO,IAAI,CAAA;IAChC,IAAI,OAAO,MAAM,KAAK,QAAQ;QAAE,OAAO,MAAM,CAAA;IAC7C,OAAO,cAAc,CAAC,MAAM,CAAC,IAAI,IAAI,CAAA;AACvC,CAAC;AAED,8FAA8F;AAC9F,MAAM,mBAAmB,GAAG,IAAI,GAAG,CAAC,CAAC,WAAW,EAAE,WAAW,EAAE,WAAW,CAAC,CAAC,CAAA;AAE5E,iFAAiF;AACjF,MAAM,UAAU,WAAW,CAAC,IAAc,EAAE,QAAyB;IACnE,IAAI,QAAQ,KAAK,OAAO,EAAE,CAAC;QACzB,OAAO,IAAI,CAAC,IAAI,KAAK,IAAI,IAAI,mBAAmB,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,KAAK,CAAC,CAAC,CAAA;IACvE,CAAC;IACD,MAAM,GAAG,GAAG,YAAY,CAAC,IAAI,CAAC,MAAM,CAAC,CAAA;IACrC,OAAO,GAAG,KAAK,EAAE,IAAI,GAAG,KAAK,CAAC,CAAA;AAChC,CAAC;AAQD,MAAM,WAAW,GAAG,CAAC,KAAa,EAAE,OAAiB,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAA;AAEhG;;;;;GAKG;AACH,MAAM,qBAAqB,GAAG,yBAAyB,CAAA;AACvD,MAAM,0BAA0B,GAAG,CAAC,QAAyB,EAAE,EAAE,CAAC,QAAQ,KAAK,UAAU,CAAA;AAEzF,MAAM,OAAO,GAAW;IACtB;QACE,IAAI,EAAE,wBAAwB;QAC9B,OAAO,EACL,wGAAwG;QAC1G,KAAK,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,QAAQ,CAAC,kBAAkB,CAAC;KACnE;IACD;QACE,IAAI,EAAE,eAAe;QACrB,OAAO,EAAE,oEAAoE;QAC7E,KAAK,EAAE,CAAC,CAAC,EAAE,EAAE,CACX,WAAW,CAAC,CAAC,EAAE;YACb,eAAe;YACf,oBAAoB;YACpB,qBAAqB;YACrB,wBAAwB;YACxB,6CAA6C;YAC7C,0BAA0B;SAC3B,CAAC;KACL;IACD;QACE,IAAI,EAAE,0BAA0B;QAChC,OAAO,EAAE,qFAAqF;QAC9F,KAAK,EAAE,CAAC,CAAC,EAAE,QAAQ,EAAE,EAAE,CACrB,WAAW,CAAC,CAAC,EAAE;YACb,kCAAkC;YAClC,4BAA4B;YAC5B,qCAAqC;YACrC,wBAAwB;SACzB,CAAC;YACF,CAAC,0BAA0B,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,QAAQ,CAAC,qBAAqB,CAAC,CAAC;KAC9E;IACD;QACE,IAAI,EAAE,kCAAkC;QACxC,OAAO,EACL,iJAAiJ;QACnJ,KAAK,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAC,wBAAwB,CAAC;KACnD;IACD;QACE,IAAI,EAAE,oBAAoB;QAC1B,OAAO,EACL,kGAAkG;QACpG,KAAK,EAAE,CAAC,CAAC,EAAE,QAAQ,EAAE,EAAE,CACrB,WAAW,CAAC,CAAC,EAAE;YACb,yBAAyB;YACzB,eAAe;YACf,kCAAkC;YAClC,uBAAuB;SACxB,CAAC;YACF,CAAC,CAAC,0BAA0B,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,QAAQ,CAAC,qBAAqB,CAAC,CAAC;KAC/E;CACF,CAAA;AAED,MAAM,CAAC,MAAM,6BAA6B,GAAG,oDAAoD,CAAA;AACjG,MAAM,CAAC,MAAM,aAAa,GACxB,4MAA4M,CAAA;AAE9M,0FAA0F;AAC1F,MAAM,UAAU,cAAc,CAC5B,MAAc,EACd,WAA4B,mBAAmB;IAE/C,MAAM,KAAK,GAAG,MAAM,CAAC,WAAW,EAAE,CAAA;IAClC,KAAK,MAAM,IAAI,IAAI,OAAO,EAAE,CAAC;QAC3B,IAAI,IAAI,CAAC,KAAK,CAAC,KAAK,EAAE,QAAQ,CAAC;YAAE,OAAO,IAAI,eAAe,CAAC,IAAI,CAAC,IAAI,EAAE,IAAI,CAAC,OAAO,EAAE,MAAM,CAAC,CAAA;IAC9F,CAAC;IACD,OAAO,IAAI,eAAe,CAAC,yBAAyB,EAAE,6BAA6B,EAAE,MAAM,CAAC,CAAA;AAC9F,CAAC;AAED,kGAAkG;AAClG,MAAM,UAAU,YAAY,CAC1B,IAAc,EACd,MAAc,EACd,QAAyB,EACzB,WAA4B,mBAAmB;IAE/C,MAAM,IAAI,GAAG,cAAc,CAAC,MAAM,EAAE,QAAQ,CAAC,CAAA;IAC7C,IAAI,IAAI,CAAC,IAAI,KAAK,yBAAyB,IAAI,CAAC,WAAW,CAAC,IAAI,EAAE,QAAQ,CAAC;QAAE,OAAO,IAAI,CAAA;IACxF,OAAO,IAAI,eAAe,CAAC,yBAAyB,EAAE,aAAa,EAAE,MAAM,CAAC,CAAA;AAC9E,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,qBAAqB,CACnC,IAAc,EACd,MAAc,EACd,MAAc,EACd,QAAyB,EACzB,WAA4B,mBAAmB;IAE/C,MAAM,IAAI,GAAG,YAAY,CAAC,IAAI,EAAE,MAAM,EAAE,QAAQ,EAAE,QAAQ,CAAC,CAAA;IAC3D,IAAI,IAAI,CAAC,IAAI,KAAK,yBAAyB;QAAE,OAAO,IAAI,CAAA;IACxD,MAAM,UAAU,GAAG,cAAc,CAAC,MAAM,EAAE,QAAQ,CAAC,CAAA;IACnD,IAAI,UAAU,CAAC,IAAI,KAAK,yBAAyB;QAAE,OAAO,UAAU,CAAA;IACpE,IAAI,MAAM,CAAC,IAAI,EAAE,KAAK,EAAE,IAAI,MAAM,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CAAC;QACjD,OAAO,IAAI,eAAe,CAAC,IAAI,CAAC,IAAI,EAAE,IAAI,CAAC,OAAO,EAAE,MAAM,CAAC,CAAA;IAC7D,CAAC;IACD,OAAO,IAAI,CAAA;AACb,CAAC"}

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

@@ -0,0 +1,14 @@
+/**
+ * llama-server runtime, provider-parameterised (`llamacpp-upstream` | `llamacpp`).
+ *
+ * Ported from: args.rs, error.rs, runtime_device.rs, device.rs, commands.rs:594 (probe),
+ * index.ts:4489-5249 (load plan). See PLAN.md §3.2.
+ */
+export * from './args.js';
+export * from './errors.js';
+export * from './runtime-device.js';
+export * from './devices.js';
+export * from './policy.js';
+export * from './probe.js';
+export * from './load-plan.js';
+export * from './runtime.js';

package/dist/runtime/llamacpp/index.js

@@ -0,0 +1,15 @@
+/**
+ * llama-server runtime, provider-parameterised (`llamacpp-upstream` | `llamacpp`).
+ *
+ * Ported from: args.rs, error.rs, runtime_device.rs, device.rs, commands.rs:594 (probe),
+ * index.ts:4489-5249 (load plan). See PLAN.md §3.2.
+ */
+export * from './args.js';
+export * from './errors.js';
+export * from './runtime-device.js';
+export * from './devices.js';
+export * from './policy.js';
+export * from './probe.js';
+export * from './load-plan.js';
+export * from './runtime.js';
+//# sourceMappingURL=index.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../src/runtime/llamacpp/index.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,cAAc,WAAW,CAAA;AACzB,cAAc,aAAa,CAAA;AAC3B,cAAc,qBAAqB,CAAA;AACnC,cAAc,cAAc,CAAA;AAC5B,cAAc,aAAa,CAAA;AAC3B,cAAc,YAAY,CAAA;AAC1B,cAAc,gBAAgB,CAAA;AAC9B,cAAc,cAAc,CAAA"}

package/dist/runtime/llamacpp/load-plan.d.ts

@@ -0,0 +1,113 @@
+/**
+ * The load plan: everything `performLoad()` in `extensions/llamacpp-upstream-extension/src/index.ts`
+ * (lines 4735-5249) decides before it hands a config to the process spawner, as a pure orchestration
+ * over injected facts. The 23 steps (PLAN.md §8.1) are numbered in comments.
+ *
+ * Owns: settings merge, `latest/` sentinel resolution, flash-attn clamp for old builds, AVX preflight,
+ * model.yml → paths (first shard, mmproj), artifact validation, Llama 3 template override, MTP/DFlash
+ * draft resolution and capability gates, `dflash_n_max`, mutual exclusion, ctx clamp, env, port and
+ * API key. Does NOT spawn; `nextRetry()` encodes the two post-failure retries.
+ */
+import type { LlamacppConfig, LocalProviderId } from '../../contracts/index.js';
+import type { ModelYmlDocument } from '../../models/index.js';
+import type { LlamacppConfigInput } from './args.js';
+/** Engine-level settings the extension reads beside `LlamacppConfig`. */
+export interface LlamacppEngineSettings {
+    /** Seconds; feeds `LLAMA_ARG_TIMEOUT` and the readiness wait. */
+    timeout: number | string;
+    /** `KEY=VALUE;…` user environment. */
+    llamacpp_env: string;
+    /** DFlash block size → `dflash_n_max = block - 1`. */
+    dflash_block_size?: number | string;
+}
+export interface LoadPlanInput {
+    provider: LocalProviderId;
+    modelId: string;
+    /** The provider's settings (34 keys) already canonicalised, plus per-model defaults. */
+    config: LlamacppConfigInput;
+    engine: LlamacppEngineSettings;
+    /** Per-model overrides (`settings` argument of `load()`), canonical keys. */
+    overrides?: Partial<LlamacppConfig> | undefined;
+    /** Explicit CLI paths bypass the installed-model registry. */
+    modelPath?: string | undefined;
+    mmprojPath?: string | undefined;
+    /** CLI `--timeout` is exact; provider loads keep the application readiness floor. */
+    timeoutSecs?: number | undefined;
+    isEmbedding: boolean;
+    dataFolder: string;
+    /** Model ids that must never be auto-unloaded (the transcription companion). */
+    transcriptionModelId?: string;
+}
+export interface LoadPlanDeps {
+    joinData: (relative: string) => string;
+    readModelYml: (modelId: string) => Promise<ModelYmlDocument>;
+    /** Resolve the `latest/<backend>` sentinel; `undefined` when offline with nothing installed. */
+    resolveLatestBackend: (backend: string) => Promise<string | undefined>;
+    /** Download/install if needed; may return a different pair after a compatible-fallback. */
+    ensureBackendReady: (backend: string, version: string) => Promise<{
+        version: string;
+        backend: string;
+        exePath: string;
+    }>;
+    /** CPU facts for the AVX preflight; `undefined` when the probe failed. */
+    cpuInfo: () => Promise<{
+        arch: string;
+        extensions: string[];
+    } | undefined>;
+    exists: (path: string) => Promise<boolean>;
+    fileSize: (path: string) => Promise<number | undefined>;
+    readGgufMetadata: (path: string) => Promise<Record<string, string> | undefined>;
+    randomPort: () => Promise<number>;
+    checkGemmaMtpSupport: (modelId: string) => boolean;
+    ensureGemmaMtpDraft: (modelId: string) => Promise<void>;
+    checkDflashSupport: (modelId: string) => boolean;
+    ensureDflashDraft: (modelId: string) => Promise<void>;
+    backendSupportsDflashSpec: (exePath: string, env: Record<string, string>) => Promise<boolean>;
+    resolveLlama3TemplateOverride: (modelId: string, embedded: string | undefined) => string | undefined;
+    strictSystemGuardSignature?: string;
+    apiSecret?: string;
+    warn?: (msg: string) => void;
+}
+export interface LoadPlan {
+    provider: LocalProviderId;
+    modelId: string;
+    version: string;
+    backend: string;
+    exePath: string;
+    config: LlamacppConfig;
+    port: number;
+    apiKey: string;
+    env: Record<string, string>;
+    modelPath: string;
+    mmprojPath: string | undefined;
+    isEmbedding: boolean;
+    timeoutSecs: number;
+    maxCtxTrain: number | undefined;
+    warnings: string[];
+}
+/** Which loaded text sessions to unload before loading `modelId` (auto-unload policy). */
+export declare function autoUnloadTargets(sessions: Array<{
+    model_id: string;
+    is_embedding: boolean;
+}>, opts: {
+    autoUnload: boolean;
+    isEmbedding: boolean;
+    bypassAutoUnload: boolean;
+    transcriptionModelId?: string;
+}): string[];
+export declare function planLlamaLoad(input: LoadPlanInput, deps: LoadPlanDeps): Promise<LoadPlan>;
+export type RetryPlan = {
+    kind: 'text-only';
+    plan: LoadPlan;
+    notify: 'multimodal-disabled';
+} | {
+    kind: 'without-mtp';
+    plan: LoadPlan;
+};
+/**
+ * Steps 21-22: after a failed spawn, decide whether to retry once. A retried plan is never
+ * retried again for the same reason (the caller passes the plan it just tried).
+ */
+export declare function nextRetry(error: unknown, plan: LoadPlan, transcriptionModelId?: string): RetryPlan | undefined;
+/** `{arch}.context_length` from the GGUF, or `undefined` (with a warning) when unreadable. */
+export declare function resolveModelMaxCtxTrain(modelPath: string, deps: Pick<LoadPlanDeps, 'readGgufMetadata'>, warn?: (m: string) => void): Promise<number | undefined>;

package/dist/runtime/llamacpp/load-plan.js

@@ -0,0 +1,259 @@
+/**
+ * The load plan: everything `performLoad()` in `extensions/llamacpp-upstream-extension/src/index.ts`
+ * (lines 4735-5249) decides before it hands a config to the process spawner, as a pure orchestration
+ * over injected facts. The 23 steps (PLAN.md §8.1) are numbered in comments.
+ *
+ * Owns: settings merge, `latest/` sentinel resolution, flash-attn clamp for old builds, AVX preflight,
+ * model.yml → paths (first shard, mmproj), artifact validation, Llama 3 template override, MTP/DFlash
+ * draft resolution and capability gates, `dflash_n_max`, mutual exclusion, ctx clamp, env, port and
+ * API key. Does NOT spawn; `nextRetry()` encodes the two post-failure retries.
+ */
+import { AtomicCoreError } from '../../contracts/index.js';
+import { effectiveCtxSize, firstGgufShardPath, ggufShardSetPaths, isMtpCapable, matchesMtpLoadFailure, parseGgufShard, } from '../../models/index.js';
+import { generateApiKey } from '../shared/index.js';
+import { withLlamacppDefaults } from './args.js';
+import { BACKEND_NOT_CONFIGURED_MESSAGE, CPU_NO_AVX_ERROR_CODE, CPU_NO_AVX_MESSAGE, codedLoadError, formatLoadError, isCpuBackend, isUnsupportedNoAvxCpu, modelLoadReadyTimeoutSecs, parseBuildNumberStrict, parseEnvString, stripBom, } from './policy.js';
+/** Which loaded text sessions to unload before loading `modelId` (auto-unload policy). */
+export function autoUnloadTargets(sessions, opts) {
+    if (!opts.autoUnload || opts.isEmbedding || opts.bypassAutoUnload)
+        return [];
+    return sessions
+        .filter((s) => s.is_embedding === false && s.model_id !== opts.transcriptionModelId)
+        .map((s) => s.model_id);
+}
+export async function planLlamaLoad(input, deps) {
+    const warnings = [];
+    const warn = (m) => {
+        warnings.push(m);
+        deps.warn?.(m);
+    };
+    const { modelId } = input;
+    // 2. settings merge (per-model overrides win)
+    const cfg = withLlamacppDefaults({ ...input.config, ...(input.overrides ?? {}) });
+    const env = {};
+    // 3. `latest/<backend>` sentinel
+    if (stripBom(cfg.version_backend || '').startsWith('latest/')) {
+        const sentinelBackend = stripBom(cfg.version_backend.split('/')[1] ?? '');
+        const resolved = await deps.resolveLatestBackend(sentinelBackend);
+        if (resolved)
+            cfg.version_backend = resolved;
+        else
+            warn(`Could not resolve latest sentinel for '${sentinelBackend}' (offline and no installed copy of the family).`);
+    }
+    // 4. split
+    let [version, backend] = cfg.version_backend.split('/');
+    if (!version || !backend)
+        throw new AtomicCoreError('INVALID_ARGUMENT', BACKEND_NOT_CONFIGURED_MESSAGE);
+    // 5. flash-attn compat for pre-b6325 builds
+    if (cfg.flash_attn === 'auto' && !backend.startsWith('ik')) {
+        const buildNum = parseBuildNumberStrict(version);
+        if (buildNum !== null && buildNum < 6325)
+            cfg.flash_attn = 'off';
+    }
+    // 6. AVX preflight (CPU backends only; a failed probe never blocks)
+    if (isCpuBackend(backend)) {
+        const cpu = await deps.cpuInfo().catch(() => undefined);
+        if (isUnsupportedNoAvxCpu(cpu?.arch ?? '', backend, cpu?.extensions ?? null)) {
+            throw codedLoadError(CPU_NO_AVX_ERROR_CODE, CPU_NO_AVX_MESSAGE);
+        }
+    }
+    // 7. backend ready (may swap to a compatible installed pair)
+    const ready = await deps.ensureBackendReady(backend, version);
+    version = ready.version;
+    backend = ready.backend;
+    const exePath = ready.exePath;
+    // 8-9. model.yml, port, key, env
+    const modelConfig = input.modelPath
+        ? {
+            model_path: input.modelPath,
+            ...(input.mmprojPath ? { mmproj_path: input.mmprojPath } : {}),
+            name: modelId,
+            size_bytes: 0,
+        }
+        : await deps.readModelYml(modelId);
+    const port = await deps.randomPort();
+    const apiKey = generateApiKey(modelId, port, deps.apiSecret);
+    const timeoutSecs = input.timeoutSecs ?? modelLoadReadyTimeoutSecs(input.engine.timeout);
+    env['LLAMA_API_KEY'] = apiKey;
+    env['LLAMA_ARG_TIMEOUT'] = String(input.timeoutSecs ?? input.engine.timeout);
+    if (input.engine.llamacpp_env)
+        parseEnvString(input.engine.llamacpp_env, env);
+    // 10. paths: first shard, mmproj
+    const modelPath = await resolveShardedModelPath(deps.joinData(modelConfig.model_path), deps, warn);
+    const configuredMmproj = input.mmprojPath ?? modelConfig.mmproj_path;
+    const mmprojPath = configuredMmproj ? deps.joinData(configuredMmproj) : undefined;
+    // 11. artifact validation
+    await assertCompleteGguf(modelPath, modelConfig.model_size_bytes, deps);
+    if (mmprojPath)
+        await assertCompleteGguf(mmprojPath, modelConfig.mmproj_size_bytes, deps);
+    // 12. Llama 3.x template override
+    if (!cfg.chat_template?.trim()) {
+        try {
+            const embedded = (await deps.readGgufMetadata(modelPath))?.['tokenizer.chat_template'];
+            const override = deps.resolveLlama3TemplateOverride(modelId, embedded);
+            if (override) {
+                cfg.chat_template = override;
+                warn(`Overriding strict embedded chat_template for "${modelId}" with the canonical Meta Llama 3.x template.`);
+            }
+            else if (deps.strictSystemGuardSignature && embedded?.includes(deps.strictSystemGuardSignature)) {
+                warn(`Model "${modelId}" has a strict system-message guard in its embedded chat_template but is not a recognized Llama 3.x format; leaving the template untouched.`);
+            }
+        }
+        catch (e) {
+            warn(`chat_template override probe failed for "${modelId}": ${formatLoadError(e)}`);
+        }
+    }
+    // TurboQuant has no speculative decoding: a value imported from an upstream copy of the settings
+    // must not make its load look for, or download, a draft model.
+    if (input.provider === 'llamacpp') {
+        cfg.mtp = false;
+        cfg.dflash = false;
+    }
+    // 13. Gemma MTP draft head
+    let mtpDraftRel = modelConfig.mtp_draft_path;
+    if (cfg.mtp && !mtpDraftRel && deps.checkGemmaMtpSupport(modelId)) {
+        try {
+            await deps.ensureGemmaMtpDraft(modelId);
+            mtpDraftRel = (await deps.readModelYml(modelId)).mtp_draft_path;
+        }
+        catch (e) {
+            warn(`Failed to ensure Gemma MTP draft head for ${modelId}; loading without MTP: ${formatLoadError(e)}`);
+        }
+    }
+    cfg.mtp_draft_path = cfg.mtp && mtpDraftRel ? deps.joinData(mtpDraftRel) : '';
+    if (cfg.mtp) {
+        const meta = await deps.readGgufMetadata(modelPath).catch((e) => {
+            warn(`Embedded MTP metadata probe failed for "${modelId}"; loading without built-in MTP: ${formatLoadError(e)}`);
+            return undefined;
+        });
+        if (!isMtpCapable(meta, cfg.mtp_draft_path)) {
+            warn(`MTP is enabled but model "${modelId}" has no MTP layers and no draft head; loading without MTP.`);
+            cfg.mtp = false;
+        }
+    }
+    // 14. DFlash binary probe
+    cfg.dflash_spec_supported = cfg.dflash ? await deps.backendSupportsDflashSpec(exePath, env) : false;
+    if (cfg.dflash && !cfg.dflash_spec_supported) {
+        warn(`DFlash is enabled but this Llama.cpp backend does not support draft-dflash; loading "${modelId}" without DFlash.`);
+        cfg.dflash = false;
+    }
+    // 15. DFlash draft
+    let dflashDraftRel = modelConfig.dflash_draft_path;
+    if (cfg.dflash && !dflashDraftRel && deps.checkDflashSupport(modelId)) {
+        try {
+            await deps.ensureDflashDraft(modelId);
+            dflashDraftRel = (await deps.readModelYml(modelId)).dflash_draft_path;
+        }
+        catch (e) {
+            warn(`Failed to ensure DFlash draft for ${modelId}; loading without DFlash: ${formatLoadError(e)}`);
+        }
+    }
+    cfg.dflash_draft_path = cfg.dflash && dflashDraftRel ? deps.joinData(dflashDraftRel) : '';
+    if (cfg.dflash && cfg.dflash_draft_path.length === 0) {
+        warn(`DFlash is enabled but model "${modelId}" has no resolvable draft; loading without DFlash.`);
+        cfg.dflash = false;
+    }
+    // 16. dflash_n_max
+    if (cfg.dflash) {
+        const blockSize = Number(input.engine.dflash_block_size);
+        cfg.dflash_n_max =
+            Number.isFinite(blockSize) && blockSize > 1 ? Math.max(Math.floor(blockSize) - 1, 1) : 0;
+    }
+    else
+        cfg.dflash_n_max = 0;
+    // 17. mutual exclusion — DFlash wins
+    if (cfg.dflash && cfg.mtp) {
+        warn(`Both MTP and DFlash are enabled for "${modelId}"; applying DFlash only.`);
+        cfg.mtp = false;
+        cfg.mtp_draft_path = '';
+    }
+    // 18. ctx clamp to the trained context
+    const maxCtxTrain = await resolveModelMaxCtxTrain(modelPath, deps, warn);
+    const clamped = effectiveCtxSize(cfg.ctx_size, maxCtxTrain);
+    if (clamped !== undefined && clamped !== cfg.ctx_size) {
+        warn(`Requested ctx_size ${cfg.ctx_size} exceeds the model's trained context; clamping to ${clamped}.`);
+        cfg.ctx_size = clamped;
+    }
+    // 19. legacy stringly `fit`
+    if (typeof cfg.fit === 'string')
+        cfg.fit = true;
+    return {
+        provider: input.provider,
+        modelId,
+        version,
+        backend,
+        exePath,
+        config: cfg,
+        port,
+        apiKey,
+        env,
+        modelPath,
+        mmprojPath,
+        isEmbedding: input.isEmbedding,
+        timeoutSecs,
+        maxCtxTrain,
+        warnings,
+    };
+}
+/**
+ * Steps 21-22: after a failed spawn, decide whether to retry once. A retried plan is never
+ * retried again for the same reason (the caller passes the plan it just tried).
+ */
+export function nextRetry(error, plan, transcriptionModelId) {
+    const code = error?.code;
+    if (plan.mmprojPath &&
+        code === 'MULTIMODAL_PROJECTOR_LOAD_FAILED' &&
+        plan.modelId !== transcriptionModelId) {
+        return { kind: 'text-only', plan: { ...plan, mmprojPath: undefined }, notify: 'multimodal-disabled' };
+    }
+    if (plan.config.mtp && matchesMtpLoadFailure(formatLoadError(error))) {
+        return {
+            kind: 'without-mtp',
+            plan: { ...plan, config: { ...plan.config, mtp: false, mtp_draft_path: '' } },
+        };
+    }
+    return undefined;
+}
+async function resolveShardedModelPath(modelPath, deps, warn) {
+    const shard = parseGgufShard(modelPath);
+    if (!shard)
+        return modelPath;
+    const set = ggufShardSetPaths(modelPath);
+    const missing = [];
+    for (const p of set)
+        if (!(await deps.exists(p)))
+            missing.push(p);
+    if (missing.length) {
+        throw codedLoadError('MODEL_SHARDS_INCOMPLETE', `This model is split into ${shard.total} parts and ${missing.length} of them are missing on disk. Re-download the model to get the complete set.`);
+    }
+    const first = firstGgufShardPath(modelPath);
+    if (first !== modelPath)
+        warn(`Model is shard ${shard.index}/${shard.total}; loading the first shard so llama.cpp can assemble the set.`);
+    return first;
+}
+async function assertCompleteGguf(filePath, expectedSize, deps) {
+    const size = await deps.fileSize(filePath).catch(() => undefined);
+    if (size === undefined) {
+        throw codedLoadError('MODEL_FILE_NOT_FOUND', `The specified model file does not exist or is not accessible: ${filePath}`);
+    }
+    if (typeof expectedSize === 'number' && expectedSize > 0 && size < expectedSize) {
+        throw codedLoadError('MODEL_FILE_CORRUPT', `The model file is incomplete (${size} of ${expectedSize} bytes), likely from an interrupted download: ${filePath}`);
+    }
+}
+/** `{arch}.context_length` from the GGUF, or `undefined` (with a warning) when unreadable. */
+export async function resolveModelMaxCtxTrain(modelPath, deps, warn = () => { }) {
+    try {
+        const meta = await deps.readGgufMetadata(modelPath);
+        const arch = meta?.['general.architecture'];
+        if (typeof arch !== 'string' || !arch)
+            return undefined;
+        const raw = meta?.[`${arch}.context_length`];
+        const parsed = raw == null ? NaN : parseInt(String(raw), 10);
+        return Number.isFinite(parsed) && parsed > 0 ? parsed : undefined;
+    }
+    catch (e) {
+        warn(`Failed to resolve max ctx_train from GGUF at ${modelPath}: ${formatLoadError(e)}`);
+        return undefined;
+    }
+}
+//# sourceMappingURL=load-plan.js.map

package/dist/runtime/llamacpp/load-plan.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"load-plan.js","sourceRoot":"","sources":["../../../src/runtime/llamacpp/load-plan.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAEH,OAAO,EAAE,eAAe,EAAE,MAAM,0BAA0B,CAAA;AAG1D,OAAO,EACL,gBAAgB,EAChB,kBAAkB,EAClB,iBAAiB,EACjB,YAAY,EACZ,qBAAqB,EACrB,cAAc,GACf,MAAM,uBAAuB,CAAA;AAC9B,OAAO,EAAE,cAAc,EAAE,MAAM,oBAAoB,CAAA;AACnD,OAAO,EAAE,oBAAoB,EAAE,MAAM,WAAW,CAAA;AAEhD,OAAO,EACL,8BAA8B,EAC9B,qBAAqB,EACrB,kBAAkB,EAClB,cAAc,EACd,eAAe,EACf,YAAY,EACZ,qBAAqB,EACrB,yBAAyB,EACzB,sBAAsB,EACtB,cAAc,EACd,QAAQ,GACT,MAAM,aAAa,CAAA;AA4EpB,0FAA0F;AAC1F,MAAM,UAAU,iBAAiB,CAC/B,QAA4D,EAC5D,IAKC;IAED,IAAI,CAAC,IAAI,CAAC,UAAU,IAAI,IAAI,CAAC,WAAW,IAAI,IAAI,CAAC,gBAAgB;QAAE,OAAO,EAAE,CAAA;IAC5E,OAAO,QAAQ;SACZ,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,YAAY,KAAK,KAAK,IAAI,CAAC,CAAC,QAAQ,KAAK,IAAI,CAAC,oBAAoB,CAAC;SACnF,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAA;AAC3B,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,aAAa,CAAC,KAAoB,EAAE,IAAkB;IAC1E,MAAM,QAAQ,GAAa,EAAE,CAAA;IAC7B,MAAM,IAAI,GAAG,CAAC,CAAS,EAAE,EAAE;QACzB,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,CAAA;QAChB,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC,CAAA;IAChB,CAAC,CAAA;IACD,MAAM,EAAE,OAAO,EAAE,GAAG,KAAK,CAAA;IAEzB,8CAA8C;IAC9C,MAAM,GAAG,GAAG,oBAAoB,CAAC,EAAE,GAAG,KAAK,CAAC,MAAM,EAAE,GAAG,CAAC,KAAK,CAAC,SAAS,IAAI,EAAE,CAAC,EAAyB,CAAC,CAAA;IACxG,MAAM,GAAG,GAA2B,EAAE,CAAA;IAEtC,iCAAiC;IACjC,IAAI,QAAQ,CAAC,GAAG,CAAC,eAAe,IAAI,EAAE,CAAC,CAAC,UAAU,CAAC,SAAS,CAAC,EAAE,CAAC;QAC9D,MAAM,eAAe,GAAG,QAAQ,CAAC,GAAG,CAAC,eAAe,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAA;QACzE,MAAM,QAAQ,GAAG,MAAM,IAAI,CAAC,oBAAoB,CAAC,eAAe,CAAC,CAAA;QACjE,IAAI,QAAQ;YAAE,GAAG,CAAC,eAAe,GAAG,QAAQ,CAAA;;YAE1C,IAAI,CACF,0CAA0C,eAAe,kDAAkD,CAC5G,CAAA;IACL,CAAC;IAED,WAAW;IACX,IAAI,CAAC,OAAO,EAAE,OAAO,CAAC,GAAG,GAAG,CAAC,eAAe,CAAC,KAAK,CAAC,GAAG,CAA6C,CAAA;IACnG,IAAI,CAAC,OAAO,IAAI,CAAC,OAAO;QAAE,MAAM,IAAI,eAAe,CAAC,kBAAkB,EAAE,8BAA8B,CAAC,CAAA;IAEvG,4CAA4C;IAC5C,IAAI,GAAG,CAAC,UAAU,KAAK,MAAM,IAAI,CAAC,OAAO,CAAC,UAAU,CAAC,IAAI,CAAC,EAAE,CAAC;QAC3D,MAAM,QAAQ,GAAG,sBAAsB,CAAC,OAAO,CAAC,CAAA;QAChD,IAAI,QAAQ,KAAK,IAAI,IAAI,QAAQ,GAAG,IAAI;YAAE,GAAG,CAAC,UAAU,GAAG,KAAK,CAAA;IAClE,CAAC;IAED,oEAAoE;IACpE,IAAI,YAAY,CAAC,OAAO,CAAC,EAAE,CAAC;QAC1B,MAAM,GAAG,GAAG,MAAM,IAAI,CAAC,OAAO,EAAE,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAA;QACvD,IAAI,qBAAqB,CAAC,GAAG,EAAE,IAAI,IAAI,EAAE,EAAE,OAAO,EAAE,GAAG,EAAE,UAAU,IAAI,IAAI,CAAC,EAAE,CAAC;YAC7E,MAAM,cAAc,CAAC,qBAAqB,EAAE,kBAAkB,CAAC,CAAA;QACjE,CAAC;IACH,CAAC;IAED,6DAA6D;IAC7D,MAAM,KAAK,GAAG,MAAM,IAAI,CAAC,kBAAkB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAA;IAC7D,OAAO,GAAG,KAAK,CAAC,OAAO,CAAA;IACvB,OAAO,GAAG,KAAK,CAAC,OAAO,CAAA;IACvB,MAAM,OAAO,GAAG,KAAK,CAAC,OAAO,CAAA;IAE7B,iCAAiC;IACjC,MAAM,WAAW,GAAqB,KAAK,CAAC,SAAS;QACnD,CAAC,CAAC;YACE,UAAU,EAAE,KAAK,CAAC,SAAS;YAC3B,GAAG,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC,CAAC,EAAE,WAAW,EAAE,KAAK,CAAC,UAAU,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YAC9D,IAAI,EAAE,OAAO;YACb,UAAU,EAAE,CAAC;SACd;QACH,CAAC,CAAC,MAAM,IAAI,CAAC,YAAY,CAAC,OAAO,CAAC,CAAA;IACpC,MAAM,IAAI,GAAG,MAAM,IAAI,CAAC,UAAU,EAAE,CAAA;IACpC,MAAM,MAAM,GAAG,cAAc,CAAC,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC,CAAA;IAC5D,MAAM,WAAW,GAAG,KAAK,CAAC,WAAW,IAAI,yBAAyB,CAAC,KAAK,CAAC,MAAM,CAAC,OAAO,CAAC,CAAA;IACxF,GAAG,CAAC,eAAe,CAAC,GAAG,MAAM,CAAA;IAC7B,GAAG,CAAC,mBAAmB,CAAC,GAAG,MAAM,CAAC,KAAK,CAAC,WAAW,IAAI,KAAK,CAAC,MAAM,CAAC,OAAO,CAAC,CAAA;IAC5E,IAAI,KAAK,CAAC,MAAM,CAAC,YAAY;QAAE,cAAc,CAAC,KAAK,CAAC,MAAM,CAAC,YAAY,EAAE,GAAG,CAAC,CAAA;IAE7E,iCAAiC;IACjC,MAAM,SAAS,GAAG,MAAM,uBAAuB,CAAC,IAAI,CAAC,QAAQ,CAAC,WAAW,CAAC,UAAU,CAAC,EAAE,IAAI,EAAE,IAAI,CAAC,CAAA;IAClG,MAAM,gBAAgB,GAAG,KAAK,CAAC,UAAU,IAAI,WAAW,CAAC,WAAW,CAAA;IACpE,MAAM,UAAU,GAAG,gBAAgB,CAAC,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,gBAAgB,CAAC,CAAC,CAAC,CAAC,SAAS,CAAA;IAEjF,0BAA0B;IAC1B,MAAM,kBAAkB,CAAC,SAAS,EAAE,WAAW,CAAC,gBAAgB,EAAE,IAAI,CAAC,CAAA;IACvE,IAAI,UAAU;QAAE,MAAM,kBAAkB,CAAC,UAAU,EAAE,WAAW,CAAC,iBAAiB,EAAE,IAAI,CAAC,CAAA;IAEzF,kCAAkC;IAClC,IAAI,CAAC,GAAG,CAAC,aAAa,EAAE,IAAI,EAAE,EAAE,CAAC;QAC/B,IAAI,CAAC;YACH,MAAM,QAAQ,GAAG,CAAC,MAAM,IAAI,CAAC,gBAAgB,CAAC,SAAS,CAAC,CAAC,EAAE,CAAC,yBAAyB,CAAC,CAAA;YACtF,MAAM,QAAQ,GAAG,IAAI,CAAC,6BAA6B,CAAC,OAAO,EAAE,QAAQ,CAAC,CAAA;YACtE,IAAI,QAAQ,EAAE,CAAC;gBACb,GAAG,CAAC,aAAa,GAAG,QAAQ,CAAA;gBAC5B,IAAI,CACF,iDAAiD,OAAO,+CAA+C,CACxG,CAAA;YACH,CAAC;iBAAM,IAAI,IAAI,CAAC,0BAA0B,IAAI,QAAQ,EAAE,QAAQ,CAAC,IAAI,CAAC,0BAA0B,CAAC,EAAE,CAAC;gBAClG,IAAI,CACF,UAAU,OAAO,6IAA6I,CAC/J,CAAA;YACH,CAAC;QACH,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,IAAI,CAAC,4CAA4C,OAAO,MAAM,eAAe,CAAC,CAAC,CAAC,EAAE,CAAC,CAAA;QACrF,CAAC;IACH,CAAC;IAED,iGAAiG;IACjG,+DAA+D;IAC/D,IAAI,KAAK,CAAC,QAAQ,KAAK,UAAU,EAAE,CAAC;QAClC,GAAG,CAAC,GAAG,GAAG,KAAK,CAAA;QACf,GAAG,CAAC,MAAM,GAAG,KAAK,CAAA;IACpB,CAAC;IAED,2BAA2B;IAC3B,IAAI,WAAW,GAAG,WAAW,CAAC,cAAc,CAAA;IAC5C,IAAI,GAAG,CAAC,GAAG,IAAI,CAAC,WAAW,IAAI,IAAI,CAAC,oBAAoB,CAAC,OAAO,CAAC,EAAE,CAAC;QAClE,IAAI,CAAC;YACH,MAAM,IAAI,CAAC,mBAAmB,CAAC,OAAO,CAAC,CAAA;YACvC,WAAW,GAAG,CAAC,MAAM,IAAI,CAAC,YAAY,CAAC,OAAO,CAAC,CAAC,CAAC,cAAc,CAAA;QACjE,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,IAAI,CAAC,6CAA6C,OAAO,0BAA0B,eAAe,CAAC,CAAC,CAAC,EAAE,CAAC,CAAA;QAC1G,CAAC;IACH,CAAC;IACD,GAAG,CAAC,cAAc,GAAG,GAAG,CAAC,GAAG,IAAI,WAAW,CAAC,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,EAAE,CAAA;IAC7E,IAAI,GAAG,CAAC,GAAG,EAAE,CAAC;QACZ,MAAM,IAAI,GAAG,MAAM,IAAI,CAAC,gBAAgB,CAAC,SAAS,CAAC,CAAC,KAAK,CAAC,CAAC,CAAU,EAAE,EAAE;YACvE,IAAI,CACF,2CAA2C,OAAO,oCAAoC,eAAe,CAAC,CAAC,CAAC,EAAE,CAC3G,CAAA;YACD,OAAO,SAAS,CAAA;QAClB,CAAC,CAAC,CAAA;QACF,IAAI,CAAC,YAAY,CAAC,IAAI,EAAE,GAAG,CAAC,cAAc,CAAC,EAAE,CAAC;YAC5C,IAAI,CAAC,6BAA6B,OAAO,6DAA6D,CAAC,CAAA;YACvG,GAAG,CAAC,GAAG,GAAG,KAAK,CAAA;QACjB,CAAC;IACH,CAAC;IAED,0BAA0B;IAC1B,GAAG,CAAC,qBAAqB,GAAG,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,IAAI,CAAC,yBAAyB,CAAC,OAAO,EAAE,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAA;IACnG,IAAI,GAAG,CAAC,MAAM,IAAI,CAAC,GAAG,CAAC,qBAAqB,EAAE,CAAC;QAC7C,IAAI,CACF,wFAAwF,OAAO,mBAAmB,CACnH,CAAA;QACD,GAAG,CAAC,MAAM,GAAG,KAAK,CAAA;IACpB,CAAC;IAED,mBAAmB;IACnB,IAAI,cAAc,GAAG,WAAW,CAAC,iBAAiB,CAAA;IAClD,IAAI,GAAG,CAAC,MAAM,IAAI,CAAC,cAAc,IAAI,IAAI,CAAC,kBAAkB,CAAC,OAAO,CAAC,EAAE,CAAC;QACtE,IAAI,CAAC;YACH,MAAM,IAAI,CAAC,iBAAiB,CAAC,OAAO,CAAC,CAAA;YACrC,cAAc,GAAG,CAAC,MAAM,IAAI,CAAC,YAAY,CAAC,OAAO,CAAC,CAAC,CAAC,iBAAiB,CAAA;QACvE,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,IAAI,CAAC,qCAAqC,OAAO,6BAA6B,eAAe,CAAC,CAAC,CAAC,EAAE,CAAC,CAAA;QACrG,CAAC;IACH,CAAC;IACD,GAAG,CAAC,iBAAiB,GAAG,GAAG,CAAC,MAAM,IAAI,cAAc,CAAC,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,cAAc,CAAC,CAAC,CAAC,CAAC,EAAE,CAAA;IACzF,IAAI,GAAG,CAAC,MAAM,IAAI,GAAG,CAAC,iBAAiB,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACrD,IAAI,CAAC,gCAAgC,OAAO,oDAAoD,CAAC,CAAA;QACjG,GAAG,CAAC,MAAM,GAAG,KAAK,CAAA;IACpB,CAAC;IAED,mBAAmB;IACnB,IAAI,GAAG,CAAC,MAAM,EAAE,CAAC;QACf,MAAM,SAAS,GAAG,MAAM,CAAC,KAAK,CAAC,MAAM,CAAC,iBAAiB,CAAC,CAAA;QACxD,GAAG,CAAC,YAAY;YACd,MAAM,CAAC,QAAQ,CAAC,SAAS,CAAC,IAAI,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAA;IAC5F,CAAC;;QAAM,GAAG,CAAC,YAAY,GAAG,CAAC,CAAA;IAE3B,qCAAqC;IACrC,IAAI,GAAG,CAAC,MAAM,IAAI,GAAG,CAAC,GAAG,EAAE,CAAC;QAC1B,IAAI,CAAC,wCAAwC,OAAO,0BAA0B,CAAC,CAAA;QAC/E,GAAG,CAAC,GAAG,GAAG,KAAK,CAAA;QACf,GAAG,CAAC,cAAc,GAAG,EAAE,CAAA;IACzB,CAAC;IAED,uCAAuC;IACvC,MAAM,WAAW,GAAG,MAAM,uBAAuB,CAAC,SAAS,EAAE,IAAI,EAAE,IAAI,CAAC,CAAA;IACxE,MAAM,OAAO,GAAG,gBAAgB,CAAC,GAAG,CAAC,QAAQ,EAAE,WAAW,CAAC,CAAA;IAC3D,IAAI,OAAO,KAAK,SAAS,IAAI,OAAO,KAAK,GAAG,CAAC,QAAQ,EAAE,CAAC;QACtD,IAAI,CAAC,sBAAsB,GAAG,CAAC,QAAQ,qDAAqD,OAAO,GAAG,CAAC,CAAA;QACvG,GAAG,CAAC,QAAQ,GAAG,OAAO,CAAA;IACxB,CAAC;IAED,4BAA4B;IAC5B,IAAI,OAAQ,GAAwB,CAAC,GAAG,KAAK,QAAQ;QAAE,GAAG,CAAC,GAAG,GAAG,IAAI,CAAA;IAErE,OAAO;QACL,QAAQ,EAAE,KAAK,CAAC,QAAQ;QACxB,OAAO;QACP,OAAO;QACP,OAAO;QACP,OAAO;QACP,MAAM,EAAE,GAAG;QACX,IAAI;QACJ,MAAM;QACN,GAAG;QACH,SAAS;QACT,UAAU;QACV,WAAW,EAAE,KAAK,CAAC,WAAW;QAC9B,WAAW;QACX,WAAW;QACX,QAAQ;KACT,CAAA;AACH,CAAC;AAMD;;;GAGG;AACH,MAAM,UAAU,SAAS,CACvB,KAAc,EACd,IAAc,EACd,oBAA6B;IAE7B,MAAM,IAAI,GAAI,KAAwC,EAAE,IAAI,CAAA;IAC5D,IACE,IAAI,CAAC,UAAU;QACf,IAAI,KAAK,kCAAkC;QAC3C,IAAI,CAAC,OAAO,KAAK,oBAAoB,EACrC,CAAC;QACD,OAAO,EAAE,IAAI,EAAE,WAAW,EAAE,IAAI,EAAE,EAAE,GAAG,IAAI,EAAE,UAAU,EAAE,SAAS,EAAE,EAAE,MAAM,EAAE,qBAAqB,EAAE,CAAA;IACvG,CAAC;IACD,IAAI,IAAI,CAAC,MAAM,CAAC,GAAG,IAAI,qBAAqB,CAAC,eAAe,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC;QACrE,OAAO;YACL,IAAI,EAAE,aAAa;YACnB,IAAI,EAAE,EAAE,GAAG,IAAI,EAAE,MAAM,EAAE,EAAE,GAAG,IAAI,CAAC,MAAM,EAAE,GAAG,EAAE,KAAK,EAAE,cAAc,EAAE,EAAE,EAAE,EAAE;SAC9E,CAAA;IACH,CAAC;IACD,OAAO,SAAS,CAAA;AAClB,CAAC;AAED,KAAK,UAAU,uBAAuB,CACpC,SAAiB,EACjB,IAAkB,EAClB,IAAyB;IAEzB,MAAM,KAAK,GAAG,cAAc,CAAC,SAAS,CAAC,CAAA;IACvC,IAAI,CAAC,KAAK;QAAE,OAAO,SAAS,CAAA;IAC5B,MAAM,GAAG,GAAG,iBAAiB,CAAC,SAAS,CAAC,CAAA;IACxC,MAAM,OAAO,GAAa,EAAE,CAAA;IAC5B,KAAK,MAAM,CAAC,IAAI,GAAG;QAAE,IAAI,CAAC,CAAC,MAAM,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;YAAE,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,CAAA;IACjE,IAAI,OAAO,CAAC,MAAM,EAAE,CAAC;QACnB,MAAM,cAAc,CAClB,yBAAyB,EACzB,4BAA4B,KAAK,CAAC,KAAK,cAAc,OAAO,CAAC,MAAM,8EAA8E,CAClJ,CAAA;IACH,CAAC;IACD,MAAM,KAAK,GAAG,kBAAkB,CAAC,SAAS,CAAC,CAAA;IAC3C,IAAI,KAAK,KAAK,SAAS;QACrB,IAAI,CACF,kBAAkB,KAAK,CAAC,KAAK,IAAI,KAAK,CAAC,KAAK,8DAA8D,CAC3G,CAAA;IACH,OAAO,KAAK,CAAA;AACd,CAAC;AAED,KAAK,UAAU,kBAAkB,CAC/B,QAAgB,EAChB,YAAgC,EAChC,IAAkB;IAElB,MAAM,IAAI,GAAG,MAAM,IAAI,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAA;IACjE,IAAI,IAAI,KAAK,SAAS,EAAE,CAAC;QACvB,MAAM,cAAc,CAClB,sBAAsB,EACtB,iEAAiE,QAAQ,EAAE,CAC5E,CAAA;IACH,CAAC;IACD,IAAI,OAAO,YAAY,KAAK,QAAQ,IAAI,YAAY,GAAG,CAAC,IAAI,IAAI,GAAG,YAAY,EAAE,CAAC;QAChF,MAAM,cAAc,CAClB,oBAAoB,EACpB,iCAAiC,IAAI,OAAO,YAAY,iDAAiD,QAAQ,EAAE,CACpH,CAAA;IACH,CAAC;AACH,CAAC;AAED,8FAA8F;AAC9F,MAAM,CAAC,KAAK,UAAU,uBAAuB,CAC3C,SAAiB,EACjB,IAA4C,EAC5C,OAA4B,GAAG,EAAE,GAAE,CAAC;IAEpC,IAAI,CAAC;QACH,MAAM,IAAI,GAAG,MAAM,IAAI,CAAC,gBAAgB,CAAC,SAAS,CAAC,CAAA;QACnD,MAAM,IAAI,GAAG,IAAI,EAAE,CAAC,sBAAsB,CAAC,CAAA;QAC3C,IAAI,OAAO,IAAI,KAAK,QAAQ,IAAI,CAAC,IAAI;YAAE,OAAO,SAAS,CAAA;QACvD,MAAM,GAAG,GAAG,IAAI,EAAE,CAAC,GAAG,IAAI,iBAAiB,CAAC,CAAA;QAC5C,MAAM,MAAM,GAAG,GAAG,IAAI,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,GAAG,CAAC,EAAE,EAAE,CAAC,CAAA;QAC5D,OAAO,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,CAAA;IACnE,CAAC;IAAC,OAAO,CAAC,EAAE,CAAC;QACX,IAAI,CAAC,gDAAgD,SAAS,KAAK,eAAe,CAAC,CAAC,CAAC,EAAE,CAAC,CAAA;QACxF,OAAO,SAAS,CAAA;IAClB,CAAC;AACH,CAAC"}

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

@@ -0,0 +1,70 @@
+/**
+ * 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';
+import type { ErrorCode, RuntimeDeviceInfo } from '../../contracts/index.js';
+export declare const CPU_NO_AVX_ERROR_CODE: "CPU_NO_AVX";
+export declare 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 declare const BACKEND_NOT_CONFIGURED_MESSAGE = "Llama.cpp backend is not configured (version_backend is missing or invalid). Check Settings \u2192 Llama.cpp \u2014 Version & Backend, or reinstall the application.";
+/** Readiness wait floor (30 min): large models on cold storage outran the 600 s default. */
+export declare const MODEL_LOAD_READY_TIMEOUT_FLOOR_SECS = 1800;
+export declare function stripBom(s: string): string;
+/** The extension's strict build-number parser (`b6325` only; the unified fork tag yields null). */
+export declare function parseBuildNumberStrict(version: string): number | null;
+/** A concrete `<version>/<backend>`: not empty, not `none`, has a `/`, not the `latest/` sentinel. */
+export declare function isConcreteVersionBackend(vb: string | undefined | null): boolean;
+/** `win-cpu-x64`, `linux-cpu-arm64`, … — never the macOS ids. */
+export declare function isCpuBackend(backend: string | undefined | null): boolean;
+export declare function cpuHasAvx(extensions: string[] | undefined | null): boolean;
+/** Block only on a positive no-AVX signal: x86, a CPU backend, a non-empty list lacking AVX. */
+export declare function isUnsupportedNoAvxCpu(arch: string | undefined | null, backend: string | undefined | null, extensions: string[] | undefined | null): boolean;
+export declare function isGpuBackendCategory(category: string): boolean;
+export type GpuKind = 'cuda' | 'vulkan' | 'other';
+export declare function gpuKindOf(category: string): GpuKind;
+/** Zero offloaded layers is conclusive; an absent device is not treated as CPU. */
+export declare function runtimeRanOnCpu(runtimeDevice: Partial<RuntimeDeviceInfo> | undefined | null): boolean;
+export type BackendMismatch = {
+    kind: 'ok';
+} | {
+    kind: 'silent-fallback';
+    configured: string;
+    effective: string;
+} | {
+    kind: 'runtime-cpu';
+    configured: string;
+    primaryDevice: string;
+    offloaded: number | null;
+    total: number | null;
+    gpuKind: GpuKind;
+    cudaRuntimeMissing: boolean;
+    deviceInitError: string | null;
+} | {
+    kind: 'suboptimal-config';
+    configured: string;
+    ideal: string;
+};
+/**
+ * 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 declare function classifyBackendMismatch(input: {
+    configuredBackend: string | undefined | null;
+    effectiveBackend: string | undefined | null;
+    runtimeDevice?: Partial<RuntimeDeviceInfo> | null;
+    idealBackend?: string | null;
+    requestedGpuLayers?: number | null;
+    categoryOf: (backend: string) => string;
+}): BackendMismatch;
+/** `KEY=VALUE;KEY2=VALUE2` — keys starting with `LLAMA` are reserved and dropped. */
+export declare function parseEnvString(envString: string, target?: Record<string, string>): Record<string, string>;
+/** Never below the floor; honours a larger configured value; non-numbers read as 600. */
+export declare function modelLoadReadyTimeoutSecs(configuredTimeoutSecs: number | string): number;
+/** Readable text of any load error: message + details + `[code]`; never `[object Object]`. */
+export declare function formatLoadError(err: unknown): string;
+/** Codes that describe a recoverable user/environment condition, not a crash (kept out of Sentry). */
+export declare const RECOVERABLE_LOAD_ERROR_CODES: ReadonlySet<string>;
+export declare function isRecoverableLoadError(err: unknown): boolean;
+/** The extension's `codedLoadError`, as an `AtomicCoreError`. */
+export declare function codedLoadError(code: ErrorCode, message: string, details?: string): AtomicCoreError;