dsh-plugin-local-infer 0.1.2 → 0.1.4

package/packages/providers/lib/engines/llamacpp/llamacppLinux.d.ts

@@ -0,0 +1,17 @@
+/**
+ * The llama.cpp shared Linux build profile. Every Linux distro (Arch, Debian,
+ * Fedora) builds llama.cpp through this one route: cmake with the Ninja
+ * generator into the `build` directory, which is where the `llama-server`
+ * binary lands. The distro profiles reuse this profile, so a distro carries no
+ * build steps of its own.
+ * @module dsh-plugin-local-infer/providers/engines/llamacpp/llamacppLinux
+ */
+import type { EngineOsProfile } from '../engineOs.ts';
+export declare const LLAMACPP_LINUX: EngineOsProfile;
+/**
+ * llamacppLinuxBinary returns the llama.cpp engine binary name on Linux: the
+ * plain `llama-server` (no .exe). The name comes from the shared Linux
+ * profile, so the spelling has one home.
+ */
+export declare function llamacppLinuxBinary(): string;
+//# sourceMappingURL=llamacppLinux.d.ts.map

package/packages/providers/lib/engines/llamacpp/llamacppLinux.js

@@ -0,0 +1,28 @@
+/**
+ * The llama.cpp shared Linux build profile. Every Linux distro (Arch, Debian,
+ * Fedora) builds llama.cpp through this one route: cmake with the Ninja
+ * generator into the `build` directory, which is where the `llama-server`
+ * binary lands. The distro profiles reuse this profile, so a distro carries no
+ * build steps of its own.
+ * @module dsh-plugin-local-infer/providers/engines/llamacpp/llamacppLinux
+ */
+import { SYSTEM_DEPS } from "../../systemDeps.js";
+export const LLAMACPP_LINUX = {
+    buildSteps: [
+        'cmake -S {dir} -B {dir}/build -G Ninja -DCMAKE_BUILD_TYPE=Release && cmake --build {dir}/build -j',
+    ],
+    buildDir: 'build',
+    binary: 'llama-server',
+    // The llama.cpp local build is a thin cmake route: the compiler and the hf
+    // CLI to download the model.
+    deps: [SYSTEM_DEPS.cmake, SYSTEM_DEPS.hfCli],
+};
+/**
+ * llamacppLinuxBinary returns the llama.cpp engine binary name on Linux: the
+ * plain `llama-server` (no .exe). The name comes from the shared Linux
+ * profile, so the spelling has one home.
+ */
+export function llamacppLinuxBinary() {
+    return LLAMACPP_LINUX.binary ?? 'llama-server';
+}
+//# sourceMappingURL=llamacppLinux.js.map

package/packages/providers/lib/engines/llamacpp/llamacppMacOS.d.ts

@@ -0,0 +1,15 @@
+/**
+ * The llama.cpp macOS build profile. macOS builds llama.cpp through the same
+ * cmake route as Linux, so this profile reuses the shared Linux build steps
+ * and build directory; the binary keeps its plain name (no .exe on macOS).
+ * @module dsh-plugin-local-infer/providers/engines/llamacpp/llamacppMacOS
+ */
+import type { EngineOsProfile } from '../engineOs.ts';
+export declare const LLAMACPP_MACOS: EngineOsProfile;
+/**
+ * llamacppMacOSBinary returns the llama.cpp engine binary name on macOS: the
+ * plain `llama-server` (no .exe on macOS). The name comes from the shared
+ * Linux profile the macOS profile reuses, so the spelling has one home.
+ */
+export declare function llamacppMacOSBinary(): string;
+//# sourceMappingURL=llamacppMacOS.d.ts.map

package/packages/providers/lib/engines/llamacpp/llamacppMacOS.js

@@ -0,0 +1,17 @@
+/**
+ * The llama.cpp macOS build profile. macOS builds llama.cpp through the same
+ * cmake route as Linux, so this profile reuses the shared Linux build steps
+ * and build directory; the binary keeps its plain name (no .exe on macOS).
+ * @module dsh-plugin-local-infer/providers/engines/llamacpp/llamacppMacOS
+ */
+import { LLAMACPP_LINUX } from "./llamacppLinux.js";
+export const LLAMACPP_MACOS = LLAMACPP_LINUX;
+/**
+ * llamacppMacOSBinary returns the llama.cpp engine binary name on macOS: the
+ * plain `llama-server` (no .exe on macOS). The name comes from the shared
+ * Linux profile the macOS profile reuses, so the spelling has one home.
+ */
+export function llamacppMacOSBinary() {
+    return LLAMACPP_MACOS.binary ?? 'llama-server';
+}
+//# sourceMappingURL=llamacppMacOS.js.map

package/packages/providers/lib/engines/llamacpp/llamacppWindows.d.ts

@@ -0,0 +1,15 @@
+/**
+ * The llama.cpp Windows build profile. llama.cpp carries no dedicated Windows
+ * build route, so a Windows host reads the shared Linux cmake steps and build
+ * directory; the binary is the .exe.
+ * @module dsh-plugin-local-infer/providers/engines/llamacpp/llamacppWindows
+ */
+import type { EngineOsProfile } from '../engineOs.ts';
+export declare const LLAMACPP_WINDOWS: EngineOsProfile;
+/**
+ * llamacppWindowsBinary returns the llama.cpp engine binary name on Windows:
+ * the `llama-server.exe` the Windows build leaves in the build tree. The name
+ * comes from the Windows profile, so the .exe spelling has one home.
+ */
+export declare function llamacppWindowsBinary(): string;
+//# sourceMappingURL=llamacppWindows.d.ts.map

package/packages/providers/lib/engines/llamacpp/llamacppWindows.js

@@ -0,0 +1,26 @@
+/**
+ * The llama.cpp Windows build profile. llama.cpp carries no dedicated Windows
+ * build route, so a Windows host reads the shared Linux cmake steps and build
+ * directory; the binary is the .exe.
+ * @module dsh-plugin-local-infer/providers/engines/llamacpp/llamacppWindows
+ */
+import { LLAMACPP_LINUX } from "./llamacppLinux.js";
+import { SYSTEM_DEPS } from "../../systemDeps.js";
+export const LLAMACPP_WINDOWS = {
+    buildSteps: LLAMACPP_LINUX.buildSteps,
+    buildDir: 'build',
+    binary: 'llama-server.exe',
+    // llama.cpp carries no dedicated Windows build route, so a Windows host
+    // reads the shared Linux cmake steps; the local build deps are the same thin
+    // cmake-plus-hf list.
+    deps: [SYSTEM_DEPS.cmake, SYSTEM_DEPS.hfCli],
+};
+/**
+ * llamacppWindowsBinary returns the llama.cpp engine binary name on Windows:
+ * the `llama-server.exe` the Windows build leaves in the build tree. The name
+ * comes from the Windows profile, so the .exe spelling has one home.
+ */
+export function llamacppWindowsBinary() {
+    return LLAMACPP_WINDOWS.binary ?? 'llama-server.exe';
+}
+//# sourceMappingURL=llamacppWindows.js.map

package/packages/providers/lib/{ninfer.d.ts → engines/ninfer/ninfer.d.ts}

@@ -7,7 +7,7 @@
  *
  * The data is the surface; the `NInferAdapter` (Phase 5) composes it into a
  * `LaunchPlan`. Adding a flag here is a data edit, never a core rewire.
- * @module dsh-plugin-local-infer/providers/ninfer
+ * @module dsh-plugin-local-infer/providers/engines/ninfer/ninfer
  */
 import type { ComboOption, EnumOption, ParameterSpec } from 'dsh-plugin-local-infer/core';
 /**
@@ -53,4 +53,4 @@ export declare const LM_HEAD_DRAFT_GATE_MESSAGE = "LM-head draft is only availab
  * the field is disabled.
  */
 export declare const NINFER_PARAMETERS: readonly ParameterSpec[];
-//# sourceMappingURL=ninfer.d.ts.map
+//# sourceMappingURL=ninfer.d.ts.map

package/packages/providers/lib/{ninfer.js → engines/ninfer/ninfer.js}

@@ -7,7 +7,7 @@
  *
  * The data is the surface; the `NInferAdapter` (Phase 5) composes it into a
  * `LaunchPlan`. Adding a flag here is a data edit, never a core rewire.
- * @module dsh-plugin-local-infer/providers/ninfer
+ * @module dsh-plugin-local-infer/providers/engines/ninfer/ninfer
  */
 /**
  * The KV cache quantization formats NInfer serves, in the order the form
@@ -94,6 +94,7 @@ export const NINFER_PARAMETERS = [
         // The port contributes no flag of its own: renderLaunch spells the local
         // `--port` and the docker host-port mapping from this value, and the
         // health URL and container name derive from it too.
+        deploy: { target: 'cli' },
     },
     {
         key: 'kvDtype',
@@ -115,6 +116,7 @@ export const NINFER_PARAMETERS = [
             return { ok: true };
         },
         toArgv: (value) => ['--kv-dtype', String(value)],
+        deploy: { target: 'cli' },
     },
     {
         key: 'maxContext',
@@ -124,6 +126,7 @@ export const NINFER_PARAMETERS = [
         defaultValue: 32768,
         group: 'Memory',
         toArgv: (value) => ['--max-context', String(value)],
+        deploy: { target: 'cli' },
     },
     {
         key: 'kvCapacity',
@@ -142,6 +145,7 @@ export const NINFER_PARAMETERS = [
         },
         group: 'Memory',
         toArgv: (value) => ['--kv-capacity', String(value)],
+        deploy: { target: 'cli' },
     },
     {
         key: 'hostKvMib',
@@ -159,6 +163,7 @@ export const NINFER_PARAMETERS = [
         },
         group: 'Memory',
         toArgv: (value) => ['--host-kv-mib', String(value)],
+        deploy: { target: 'cli' },
     },
     {
         key: 'hostStateSlots',
@@ -167,6 +172,7 @@ export const NINFER_PARAMETERS = [
         form: { kind: 'number', min: 0, step: 1, unit: 'slots' },
         group: 'Memory',
         toArgv: (value) => ['--host-state-slots', String(value)],
+        deploy: { target: 'cli' },
     },
     {
         key: 'mediaLiveMib',
@@ -175,6 +181,7 @@ export const NINFER_PARAMETERS = [
         form: { kind: 'number', min: 0, step: 1024, unit: 'MiB' },
         group: 'Memory',
         toArgv: (value) => ['--media-live-mib', String(value)],
+        deploy: { target: 'cli' },
     },
     {
         key: 'gdnStateFp16',
@@ -184,6 +191,7 @@ export const NINFER_PARAMETERS = [
         defaultValue: false,
         group: 'Memory',
         toArgv: (value) => (value === true ? ['--gdn-state-fp16'] : []),
+        deploy: { target: 'cli' },
     },
     {
         key: 'maxConcurrency',
@@ -193,6 +201,7 @@ export const NINFER_PARAMETERS = [
         defaultValue: 1,
         group: 'Concurrency',
         toArgv: (value) => ['--max-concurrency', String(value)],
+        deploy: { target: 'cli' },
     },
     {
         key: 'spec',
@@ -202,6 +211,7 @@ export const NINFER_PARAMETERS = [
         defaultValue: 'none',
         group: 'Speculation',
         toArgv: (value) => (value === 'none' ? [] : ['--spec', String(value)]),
+        deploy: { target: 'cli' },
     },
     {
         key: 'draftTokens',
@@ -212,16 +222,18 @@ export const NINFER_PARAMETERS = [
         group: 'Speculation',
         requires: (config) => config.spec && config.spec !== 'none' ? null : SPEC_GATE_MESSAGE,
         toArgv: (value) => ['--draft-tokens', String(value)],
+        deploy: { target: 'cli' },
     },
     {
         key: 'lmHeadDraft',
         label: 'LM-head draft',
-        description: 'Drafts from the small proposal head instead of the full output head, cutting the per-draft projection cost. An MTP-backend feature, so the field unlocks only while MTP is selected.',
+        description: 'Drafts from the small proposal head instead of the full output head, cutting the per-draft projection cost.',
         form: { kind: 'boolean' },
         defaultValue: false,
         group: 'Speculation',
         requires: (config) => config.spec === 'mtp' ? null : LM_HEAD_DRAFT_GATE_MESSAGE,
         toArgv: (value) => (value === true ? ['--lm-head-draft'] : []),
+        deploy: { target: 'cli' },
     },
     {
         key: 'temperature',
@@ -231,6 +243,7 @@ export const NINFER_PARAMETERS = [
         defaultValue: 0.9,
         group: 'Sampling',
         toArgv: (value) => ['--temperature', String(value)],
+        deploy: { target: 'cli' },
     },
     {
         key: 'minP',
@@ -240,6 +253,7 @@ export const NINFER_PARAMETERS = [
         defaultValue: 0.05,
         group: 'Sampling',
         toArgv: (value) => ['--min-p', String(value)],
+        deploy: { target: 'cli' },
     },
     {
         key: 'defaultMaxTokens',
@@ -264,6 +278,7 @@ export const NINFER_PARAMETERS = [
         },
         group: 'Sampling',
         toArgv: (value) => ['--default-max-tokens', String(value)],
+        deploy: { target: 'cli' },
     },
     {
         key: 'defaultThinkingBudget',
@@ -272,6 +287,7 @@ export const NINFER_PARAMETERS = [
         form: { kind: 'number', min: 0, step: 1, unit: 'tokens' },
         group: 'Sampling',
         toArgv: (value) => ['--default-thinking-budget', String(value)],
+        deploy: { target: 'cli' },
     },
     {
         key: 'preserveThinking',
@@ -281,6 +297,7 @@ export const NINFER_PARAMETERS = [
         defaultValue: false,
         group: 'Sampling',
         toArgv: (value) => (value === true ? ['--preserve-thinking'] : []),
+        deploy: { target: 'cli' },
     },
     {
         key: 'prefillChunk',
@@ -310,6 +327,7 @@ export const NINFER_PARAMETERS = [
         },
         group: 'Engine',
         toArgv: (value) => ['--prefill-chunk', String(value)],
+        deploy: { target: 'cli' },
     },
     {
         key: 'vision',
@@ -319,6 +337,7 @@ export const NINFER_PARAMETERS = [
         defaultValue: false,
         group: 'Engine',
         toArgv: (value) => (value === true ? ['--vision'] : []),
+        deploy: { target: 'cli' },
     },
     {
         key: 'structuredOutput',
@@ -328,6 +347,7 @@ export const NINFER_PARAMETERS = [
         defaultValue: false,
         group: 'Engine',
         toArgv: (value) => (value === true ? ['--structured-output'] : []),
+        deploy: { target: 'cli' },
     },
     {
         key: 'deviceProfile',
@@ -337,6 +357,7 @@ export const NINFER_PARAMETERS = [
         defaultValue: 'auto',
         group: 'Engine',
         toArgv: (value) => (value === 'auto' ? [] : ['--device-profile', String(value)]),
+        deploy: { target: 'cli' },
     },
     {
         key: 'chatTemplate',
@@ -345,6 +366,7 @@ export const NINFER_PARAMETERS = [
         form: { kind: 'combo', options: [...CHAT_TEMPLATE_OPTIONS] },
         group: 'Engine',
         toArgv: (value) => (value ? ['--chat-template', String(value)] : []),
+        deploy: { target: 'cli' },
     },
 ];
-//# sourceMappingURL=ninfer.js.map
+//# sourceMappingURL=ninfer.js.map

package/packages/providers/lib/engines/ninfer/ninferAdapter.d.ts

@@ -0,0 +1,81 @@
+/**
+ * The NInfer engine adapter. NInferModelArtifact is the per-model deploy data:
+ * the `.ninfer` artifact file, the Hugging Face repo the model download pulls
+ * it from, and the `--model-id` alias the server advertises. The data is the
+ * surface; adding a model is a data edit here, never an adapter rewire.
+ * NInferAdapter is the canonical both-capable engine. `renderLaunch` renders a
+ * local `ninfer-serve` process or a `docker run` of the ghcr image from the
+ * same spec surface; `deploy` builds the engine in the engine directory (or
+ * pulls the image), downloads the model into the model directory, and starts
+ * the engine; `healthCheck` polls the health URL; `stop` signals the tracked
+ * process and stops the named container.
+ * @module dsh-plugin-local-infer/providers/engines/ninfer/ninferAdapter
+ */
+import { EngineAdapter } from '../../../../core/lib/index.js';
+import type { CommandLogger, Config, DeployOptions, DeployResult, EngineLaunchSignature, EnginePlatformConfig, HealthCheckOptions, HealthResult, HardwareProfile, LaunchPlan, LaunchProgress, LifecycleHost, ParameterSpec } from '../../../../core/lib/index.js';
+/**
+ * NInferModelArtifact is the per-model deploy data: the `.ninfer` artifact
+ * file, the Hugging Face repo the model download pulls it from, and the
+ * `--model-id` alias the server advertises. The data is the surface; adding a
+ * model is a data edit here, never an adapter rewire.
+ */
+export interface NInferModelArtifact {
+    file: string;
+    repo: string;
+    modelId: string;
+}
+/**
+ * The NInfer model artifacts, keyed by the model registry id. The 27B entry
+ * is the default the DEFAULT_MODELS list names; the 35B-A3B entries are the
+ * Qwen3.8 measured artifact (`qwen3-35b`) and the Qwen3.6 default
+ * (`qwen36-35b`).
+ */
+export declare const NINFER_MODEL_ARTIFACTS: Record<string, NInferModelArtifact>;
+/**
+ * The NInfer docker image the adapter pulls and runs. The engineFork platform
+ * config overrides the git source for a local build; the image stays the
+ * upstream ghcr artifact.
+ */
+export declare const NINFER_DOCKER_IMAGE = "ghcr.io/iamwavecut/ninfer-all";
+/**
+ * NInferAdapter is the canonical both-capable engine. `renderLaunch` renders a
+ * local `ninfer-serve` process or a `docker run` of the ghcr image from the
+ * same spec surface; `deploy` builds the engine in the engine directory (or
+ * pulls the image), downloads the model into the model directory, and starts
+ * the engine; `healthCheck` polls the health URL; `stop` signals the tracked
+ * process and stops the named container.
+ */
+export declare class NInferAdapter extends EngineAdapter {
+    readonly id = "ninfer";
+    readonly supports: {
+        docker: boolean;
+        local: boolean;
+    };
+    private readonly host;
+    private readonly children;
+    constructor(host?: LifecycleHost);
+    parameters(): readonly ParameterSpec[];
+    modelRef(config: Config): {
+        id: string;
+        name: string;
+    };
+    launchSignature(config?: Config): EngineLaunchSignature;
+    renderLaunch(config: Config, _hw: HardwareProfile): LaunchPlan;
+    deploy(plan: LaunchPlan, platform?: EnginePlatformConfig, onProgress?: LaunchProgress, options?: DeployOptions): Promise<DeployResult>;
+    healthCheck(plan: LaunchPlan, options?: HealthCheckOptions): Promise<HealthResult>;
+    stop(plan: LaunchPlan, log?: CommandLogger): Promise<void>;
+    private buildMode;
+    private port;
+    private downloadModel;
+    /**
+     * resolveChatTemplate rewrites a --chat-template flag that carries an HF repo
+     * reference into the local file the engine reads. A value with a slash that
+     * is not an absolute path (the combo control's known choices) is a repo the
+     * deploy step downloads into the model directory; an already-local path is
+     * left as-is. The download is best-effort: a template that is already on
+     * disk is used without a re-fetch, and a missing hf CLI leaves the flag
+     * untouched so the engine reports the missing template itself.
+     */
+    private resolveChatTemplate;
+}
+//# sourceMappingURL=ninferAdapter.d.ts.map