dsh-plugin-local-infer 0.1.3 → 0.1.4

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/engines/{ninfer.d.ts → 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/engines/{ninfer.js → 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',
@@ -212,6 +222,7 @@ 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',
@@ -222,6 +233,7 @@ export const NINFER_PARAMETERS = [
         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/{ninferAdapter.d.ts → ninfer/ninferAdapter.d.ts}

@@ -9,10 +9,10 @@
  * 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/ninferAdapter
+ * @module dsh-plugin-local-infer/providers/engines/ninfer/ninferAdapter
  */
-import { EngineAdapter } from '../../../core/lib/index.js';
-import type { Config, DeployOptions, DeployResult, EngineLaunchSignature, EnginePlatformConfig, HealthCheckOptions, HealthResult, HardwareProfile, LaunchPlan, LaunchProgress, LifecycleHost, ParameterSpec } from '../../../core/lib/index.js';
+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
@@ -63,10 +63,9 @@ export declare class NInferAdapter extends EngineAdapter {
     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): Promise<void>;
+    stop(plan: LaunchPlan, log?: CommandLogger): Promise<void>;
     private buildMode;
     private port;
-    private planKey;
     private downloadModel;
     /**
      * resolveChatTemplate rewrites a --chat-template flag that carries an HF repo
@@ -79,4 +78,4 @@ export declare class NInferAdapter extends EngineAdapter {
      */
     private resolveChatTemplate;
 }
-//# sourceMappingURL=ninferAdapter.d.ts.map
+//# sourceMappingURL=ninferAdapter.d.ts.map

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

@@ -9,13 +9,28 @@
  * 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/ninferAdapter
+ * @module dsh-plugin-local-infer/providers/engines/ninfer/ninferAdapter
  */
 import { basename, dirname, join } from 'node:path';
-import { EngineAdapter } from '../../../core/lib/index.js';
-import { customArgFlags, flagsFromSpecs, nodeLifecycleHost, pollHealth, trackChild, } from "../adapters.js";
-import { syncAndRebuildEngine } from "../acquire.js";
+import { CHAT_TEMPLATE_FILE, EngineAdapter, autoRebuildEnabled, chatTemplateArtifactPath, chatTemplatePresent, hfChatTemplateCommand, hfDownloadCommand, hfHubCacheDirectory, hubCacheSnapshotFilePath, modelArtifactPresent, modelDownloadTarget, } from '../../../../core/lib/index.js';
+import { auditRun, customArgFlags, flagsFromSpecs, nodeLifecycleHost, pollHealth, resolveLaunchArgv, stopProcessPlan, trackChild, } from "../../adapters.js";
+import { syncEngineForDeploy } from "../../acquire.js";
 import { NINFER_PARAMETERS } from "./ninfer.js";
+import { ninferLinuxBinary } from "./ninferLinux.js";
+import { ninferMacOSBinary } from "./ninferMacOS.js";
+import { ninferWindowsBinary } from "./ninferWindows.js";
+/**
+ * The NInfer engine binary is OS-specific (the .exe on Windows, where the
+ * vcpkg build-native route writes the Release binary), so the launch binary
+ * comes from the host OS's binary function. The adapter's renderLaunch reads
+ * this map by the hardware profile's os, keeping the per-OS name in its own OS
+ * file.
+ */
+const NINFER_BINARY = {
+    linux: ninferLinuxBinary,
+    macos: ninferMacOSBinary,
+    windows: ninferWindowsBinary,
+};
 /**
  * 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
@@ -149,15 +164,25 @@ export class NInferAdapter extends EngineAdapter {
                 ...(modelDirectory === '' ? {} : { modelDirectory }),
                 port,
                 healthUrl: `http://127.0.0.1:${port}/health`,
+                os: _hw.os,
             };
         }
-        const modelPath = modelDirectory === '' ? file : join(modelDirectory, file);
+        // The hub-cache layout (the model directory is the plugin default)
+        // resolves the concrete snapshot file the engine reads; a user override
+        // keeps the flat model-directory path. A not-yet-downloaded model falls
+        // back to the bare file name, the same as the flat layout's missing-file
+        // case.
+        const modelPath = modelDirectory === ''
+            ? (hubCacheSnapshotFilePath(hfHubCacheDirectory(), artifact.repo, file) ?? file)
+            : join(modelDirectory, file);
         const customArgs = customArgFlags(config);
         return {
             engineId: this.id,
             kind: 'process',
             argv: [
-                'ninfer-serve',
+                // The engine binary is OS-specific (the .exe on Windows), so the name
+                // comes from the host OS's binary function.
+                NINFER_BINARY[_hw.os](),
                 modelPath,
                 '--host',
                 '127.0.0.1',
@@ -173,6 +198,7 @@ export class NInferAdapter extends EngineAdapter {
             ...(modelDirectory === '' ? {} : { modelDirectory }),
             port,
             healthUrl: `http://127.0.0.1:${port}/health`,
+            os: _hw.os,
         };
     }
     async deploy(plan, platform, onProgress, options) {
@@ -181,41 +207,30 @@ export class NInferAdapter extends EngineAdapter {
             let buildLogs = '';
             // Auto Rebuild off (the card's checkbox): the git sync (fetch, rebase,
             // rebuild) is skipped and the checkout launches as-is.
-            if (engineDir !== undefined && options?.autoRebuild !== false) {
-                // The git sync is the checkout download: the card's button reads
-                // "Downloading..." for it, not "Building...".
-                onProgress?.('download');
+            if (engineDir !== undefined && autoRebuildEnabled(options?.autoRebuild)) {
                 // The conditional rebuild fetches the checkout, prices it against the
                 // remote branch, and rebuilds only when it is behind (or a fresh
                 // clone). The decision, the build command, and its output are captured
                 // into buildLogs and threaded to the card's Logs section.
-                const sync = await syncAndRebuildEngine(this.host, {
+                const sync = await syncEngineForDeploy(this.host, {
                     engineId: this.id,
                     directory: engineDir,
-                    ...(platform?.engineFork !== undefined ? { fork: platform.engineFork } : {}),
-                    ...(plan.argv?.[0] !== undefined ? { binary: plan.argv[0] } : {}),
-                    ...(options?.allowRunDeps !== undefined ? { allowRunDeps: options.allowRunDeps } : {}),
-                    ...(platform?.buildCommands !== undefined
-                        ? { buildCommands: platform.buildCommands }
-                        : {}),
-                    // Report the build phase the moment the build steps start, so the
-                    // card's button reads "Building..." while the build is in flight.
-                    onBuildStart: () => onProgress?.('build'),
-                }, () => { });
-                buildLogs = sync.logs;
+                    ...(platform !== undefined ? { platform } : {}),
+                    ...(options !== undefined ? { options } : {}),
+                    // The host OS the build steps pick for: a Windows host reads the
+                    // engine's Windows build steps (the vcpkg route) when the user
+                    // carries no build-command override.
+                    ...(plan.os !== undefined ? { os: plan.os } : {}),
+                    ...(onProgress !== undefined ? { onProgress } : {}),
+                });
+                buildLogs = sync.buildLogs;
                 if (!sync.ok) {
                     return { ok: false, detail: sync.detail ?? 'unknown', logs: buildLogs };
                 }
-                // The build output reads as a fresh Logs section: a "Building..."
-                // separator leads the build decision and output, the same line the
-                // card writes when the build step event lands.
-                if (sync.rebuilt && buildLogs !== '') {
-                    buildLogs = `--- Building...\n${buildLogs}`;
-                }
             }
             const withLogs = (result) => buildLogs !== '' ? { ...result, logs: buildLogs } : result;
             onProgress?.('download');
-            const download = await this.downloadModel(plan);
+            const download = await this.downloadModel(plan, options?.onLog);
             if (!download.ok)
                 return withLogs(download);
             // A chat template the user picked as an HF repo reference (the combo
@@ -223,20 +238,36 @@ export class NInferAdapter extends EngineAdapter {
             // deploy step downloads it into the model directory and rewrites the
             // --chat-template flag to the local file before the spawn.
             onProgress?.('download');
-            const templateResult = await this.resolveChatTemplate(plan);
+            const templateResult = await this.resolveChatTemplate(plan, options?.onLog);
             if (!templateResult.ok)
                 return withLogs(templateResult);
-            const handle = this.host.spawn(plan.argv ?? [], plan.env ?? {}, engineDir);
+            // A relative launch path (the NInfer Windows build tree,
+            // build-windows\apps\Release\ninfer-serve.exe) is joined against the
+            // engine directory the spawn runs in, so the spawn resolves an absolute
+            // path. A bare name (the Linux and macOS binaries, resolved against the
+            // path) is left as-is.
+            const resolvedArgv = resolveLaunchArgv(engineDir, plan.argv);
+            // The plan is the launch record's argv source (launchRecordFromPlan reads
+            // it), and the re-link rule substring-matches every record argv token
+            // against the live command line, so the record must name the spawn, not
+            // the build-tree relative binary.
+            plan.argv = resolvedArgv;
+            const handle = this.host.spawn(resolvedArgv, plan.env ?? {}, engineDir);
             trackChild(this.children, this.planKey(plan), handle);
             return withLogs({ ok: true, handle });
         }
         if (plan.kind === 'docker') {
-            // The image pull is the docker build step: the card's button reads
-            // "Building..." for it.
-            onProgress?.('build');
-            const pull = await this.host.run(`docker pull ${NINFER_DOCKER_IMAGE}`);
-            if (!pull.ok) {
-                return { ok: false, detail: `docker pull failed: ${pull.detail ?? 'unknown'}` };
+            // Auto Rebuild off (the card's checkbox): the image pull is skipped and
+            // the already-present image starts. The image pull is the docker build
+            // step: the card's button reads "Building..." for it, and its
+            // `$ <command>` line reaches the launch run's log sink before the child
+            // spawns.
+            if (autoRebuildEnabled(options?.autoRebuild)) {
+                onProgress?.('build');
+                const pull = await auditRun(this.host, `docker pull ${NINFER_DOCKER_IMAGE}`, options?.onLog);
+                if (!pull.ok) {
+                    return { ok: false, detail: `docker pull failed: ${pull.detail ?? 'unknown'}` };
+                }
             }
             const handle = this.host.spawn(plan.argv ?? [], plan.env ?? {});
             trackChild(this.children, this.planKey(plan), handle);
@@ -246,33 +277,56 @@ export class NInferAdapter extends EngineAdapter {
         return { ok: true };
     }
     async healthCheck(plan, options) {
-        return pollHealth(this.host, plan, options);
+        // The tracked handle rides into the poll, so a child that exits (a
+        // crash) ends the poll early with a failed result.
+        return pollHealth(this.host, plan, options, this.children.get(this.planKey(plan)));
     }
-    async stop(plan) {
+    async stop(plan, log) {
         if (plan.kind === 'docker' && plan.containerName !== undefined) {
-            await this.host.run(`docker stop ${plan.containerName}`);
+            await auditRun(this.host, `docker stop ${plan.containerName}`, log);
         }
         const handle = this.children.get(this.planKey(plan));
         if (handle !== undefined) {
             handle.kill();
             this.children.delete(this.planKey(plan));
+            return;
         }
+        // No tracked child (an engine a DSH restart re-adopted): signal the pid
+        // the stored launch record stamped on the plan, so the stop reaches a
+        // process this adapter instance did not spawn.
+        await stopProcessPlan(this.host, plan, log);
     }
     buildMode(config) {
         return config.buildMode === 'docker' ? 'docker' : 'local';
     }
     port(config) {
+        // 18020 is the port the ninfer-serve CLI binds by default, so a bare
+        // config launches on it.
         return typeof config.port === 'number' ? config.port : 18020;
     }
-    planKey(plan) {
-        return `${plan.engineId}:${plan.port}`;
-    }
-    async downloadModel(plan) {
+    async downloadModel(plan, log) {
         const artifact = plan.modelId !== undefined ? NINFER_MODEL_ARTIFACTS[plan.modelId] : undefined;
         if (artifact === undefined)
             return { ok: true };
-        const target = plan.modelDirectory ?? '.';
-        const download = await this.host.run(`hf download ${artifact.repo} ${artifact.file} --local-dir ${target}`);
+        // The hub-cache layout (no --local-dir) applies when the model directory is
+        // the plugin default; a user override keeps the flat --local-dir target.
+        const { baseDir, useHubCache } = modelDownloadTarget(plan.modelDirectory);
+        const entry = {
+            engineId: this.id,
+            modelId: plan.modelId ?? '',
+            label: artifact.modelId,
+            url: `https://huggingface.co/${artifact.repo}`,
+            file: artifact.file,
+        };
+        // A model the shared Get-steps routine already downloaded is on disk, so
+        // the deploy step skips the re-fetch and a launch after a Get does no
+        // duplicate download. Presence is the deterministic snapshot the resolved
+        // layout holds (modelArtifactPresent), never a bare blobs folder: a cache
+        // whose snapshot links are missing re-runs the cheap hf download, which
+        // reuses the blobs and writes the links.
+        if (modelArtifactPresent(entry, baseDir, useHubCache))
+            return { ok: true };
+        const download = await auditRun(this.host, hfDownloadCommand(entry, baseDir, useHubCache), log);
         if (!download.ok) {
             return { ok: false, detail: `model download failed: ${download.detail ?? 'unknown'}` };
         }
@@ -287,7 +341,7 @@ export class NInferAdapter extends EngineAdapter {
      * disk is used without a re-fetch, and a missing hf CLI leaves the flag
      * untouched so the engine reports the missing template itself.
      */
-    async resolveChatTemplate(plan) {
+    async resolveChatTemplate(plan, log) {
         const argv = plan.argv;
         if (argv === undefined)
             return { ok: true };
@@ -301,24 +355,38 @@ export class NInferAdapter extends EngineAdapter {
         // is a bare file name the engine resolves against its own directory.
         if (value.startsWith('/') || !value.includes('/'))
             return { ok: true };
-        // An HF repo reference: download the template file into the model
-        // directory and point the flag at the local copy.
-        const target = plan.modelDirectory ?? '.';
-        const localPath = `${target.endsWith('/') ? '' : `${target}/`}chat_template.jinja`;
+        // An HF repo reference: download the template file and point the flag at
+        // the local copy. The hub-cache layout (the model directory is the plugin
+        // default) resolves the concrete snapshot file path the hf CLI writes; a
+        // user override keeps the flat model-directory path. The flat path is
+        // built with join (the host's path module normalizes the separators), so
+        // the flag the engine reads resolves on Windows, Linux, and macOS alike.
+        const { baseDir, useHubCache } = modelDownloadTarget(plan.modelDirectory);
+        // The hub-cache branch resolves the deterministic snapshot file; the flat
+        // branch asks the shared presence predicate for the template join.
+        const resolvePath = () => useHubCache
+            ? hubCacheSnapshotFilePath(baseDir, value, CHAT_TEMPLATE_FILE)
+            : chatTemplatePresent(value, baseDir)
+                ? chatTemplateArtifactPath(baseDir)
+                : null;
+        let localPath = resolvePath();
         // A template already on disk skips the re-fetch; the hf CLI is optional,
         // so a download failure is reported but does not abort the launch.
-        const probe = await this.host.run(`[ -f ${localPath} ]`);
-        if (!probe.ok) {
-            const download = await this.host.run(`hf download ${value} chat_template.jinja --local-dir ${target}`);
+        if (localPath === null) {
+            const download = await auditRun(this.host, hfChatTemplateCommand(value, baseDir, useHubCache), log);
             if (!download.ok) {
                 return {
                     ok: false,
                     detail: `chat template download failed: ${download.detail ?? 'unknown'}`,
                 };
             }
+            localPath = resolvePath();
+            if (localPath === null) {
+                return { ok: false, detail: 'chat template download failed' };
+            }
         }
         argv[index + 1] = localPath;
         return { ok: true };
     }
 }
-//# sourceMappingURL=ninferAdapter.js.map
+//# sourceMappingURL=ninferAdapter.js.map

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

@@ -0,0 +1,9 @@
+/**
+ * The NInfer Arch Linux build profile. Arch builds NInfer through the shared
+ * Linux route (cmake with the Ninja generator), so this profile reuses the
+ * shared Linux profile and carries no build steps of its own.
+ * @module dsh-plugin-local-infer/providers/engines/ninfer/ninferArch
+ */
+import type { EngineOsProfile } from '../engineOs.ts';
+export declare const NINFER_ARCH: EngineOsProfile;
+//# sourceMappingURL=ninferArch.d.ts.map

package/packages/providers/lib/engines/ninfer/ninferArch.js

@@ -0,0 +1,9 @@
+/**
+ * The NInfer Arch Linux build profile. Arch builds NInfer through the shared
+ * Linux route (cmake with the Ninja generator), so this profile reuses the
+ * shared Linux profile and carries no build steps of its own.
+ * @module dsh-plugin-local-infer/providers/engines/ninfer/ninferArch
+ */
+import { NINFER_LINUX } from "./ninferLinux.js";
+export const NINFER_ARCH = NINFER_LINUX;
+//# sourceMappingURL=ninferArch.js.map

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

@@ -0,0 +1,9 @@
+/**
+ * The NInfer Debian build profile. Debian builds NInfer through the shared
+ * Linux route (cmake with the Ninja generator), so this profile reuses the
+ * shared Linux profile and carries no build steps of its own.
+ * @module dsh-plugin-local-infer/providers/engines/ninfer/ninferDebian
+ */
+import type { EngineOsProfile } from '../engineOs.ts';
+export declare const NINFER_DEBIAN: EngineOsProfile;
+//# sourceMappingURL=ninferDebian.d.ts.map

package/packages/providers/lib/engines/ninfer/ninferDebian.js

@@ -0,0 +1,9 @@
+/**
+ * The NInfer Debian build profile. Debian builds NInfer through the shared
+ * Linux route (cmake with the Ninja generator), so this profile reuses the
+ * shared Linux profile and carries no build steps of its own.
+ * @module dsh-plugin-local-infer/providers/engines/ninfer/ninferDebian
+ */
+import { NINFER_LINUX } from "./ninferLinux.js";
+export const NINFER_DEBIAN = NINFER_LINUX;
+//# sourceMappingURL=ninferDebian.js.map

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

@@ -0,0 +1,9 @@
+/**
+ * The NInfer Fedora build profile. Fedora builds NInfer through the shared
+ * Linux route (cmake with the Ninja generator), so this profile reuses the
+ * shared Linux profile and carries no build steps of its own.
+ * @module dsh-plugin-local-infer/providers/engines/ninfer/ninferFedora
+ */
+import type { EngineOsProfile } from '../engineOs.ts';
+export declare const NINFER_FEDORA: EngineOsProfile;
+//# sourceMappingURL=ninferFedora.d.ts.map

package/packages/providers/lib/engines/ninfer/ninferFedora.js

@@ -0,0 +1,9 @@
+/**
+ * The NInfer Fedora build profile. Fedora builds NInfer through the shared
+ * Linux route (cmake with the Ninja generator), so this profile reuses the
+ * shared Linux profile and carries no build steps of its own.
+ * @module dsh-plugin-local-infer/providers/engines/ninfer/ninferFedora
+ */
+import { NINFER_LINUX } from "./ninferLinux.js";
+export const NINFER_FEDORA = NINFER_LINUX;
+//# sourceMappingURL=ninferFedora.js.map

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

@@ -0,0 +1,17 @@
+/**
+ * The NInfer shared Linux build profile. Every Linux distro (Arch, Debian,
+ * Fedora) builds NInfer through this one route: cmake with the Ninja generator
+ * into the `build` directory, which is where the `ninfer-serve` 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/ninfer/ninferLinux
+ */
+import type { EngineOsProfile } from '../engineOs.ts';
+export declare const NINFER_LINUX: EngineOsProfile;
+/**
+ * ninferLinuxBinary returns the NInfer engine binary name on Linux: the plain
+ * `ninfer-serve` (no .exe). The name comes from the shared Linux profile, so
+ * the spelling has one home.
+ */
+export declare function ninferLinuxBinary(): string;
+//# sourceMappingURL=ninferLinux.d.ts.map