@tiens.nguyen/gonext-local-worker 1.0.11 → 1.0.13

package/gonext-local-worker.mjs

@@ -6,9 +6,14 @@
  * - `gonext-local-worker` starts polling loop
  */
 import { mkdir, readFile, writeFile } from "node:fs/promises";
-import { homedir } from "node:os";
+import { execFile as execFileCallback } from "node:child_process";
+import { homedir, platform } from "node:os";
 import { join } from "node:path";
+import { promisify } from "node:util";
 import dotenv from "dotenv";
+
+/** Avoid `node:child_process/promises` — not available on some Node builds / older runtimes. */
+const execFile = promisify(execFileCallback);
 import OpenAI from "openai";
 
 const ENV_FILE = join(homedir(), ".gonext", "worker.env");
@@ -28,6 +33,9 @@ Usage:
 Examples:
   gonext-local-worker set abc123 --api-base https://hwohu56e8d.execute-api.ap-southeast-1.amazonaws.com
   gonext-local-worker
+
+Env (optional):
+  GONEXT_MLX_LM_PYTHON   Python executable for MLX LM native probe (default: python3)
 `);
 }
 
@@ -228,6 +236,50 @@ async function checkOpenAiModels(base, apiKey) {
   }
 }
 
+/** True MLX LM check: import mlx_lm in Python (macOS). Not the OpenAI HTTP surface. */
+async function checkMlxLmNativeImport() {
+  const preferred = (process.env.GONEXT_MLX_LM_PYTHON ?? "").trim() || "python3";
+  const code = [
+    "import sys",
+    "try:",
+    "    import mlx_lm",
+    "    v = getattr(mlx_lm, '__version__', None)",
+    "    print(v or 'ok')",
+    "except Exception:",
+    "    sys.exit(1)",
+  ].join("\n");
+
+  const candidates = [preferred];
+  if (preferred === "python3") candidates.push("python");
+
+  const tried = [];
+  for (const exe of [...new Set(candidates)]) {
+    tried.push(exe);
+    try {
+      const { stdout } = await execFile(exe, ["-c", code], {
+        timeout: 15000,
+        maxBuffer: 65536,
+        windowsHide: true,
+      });
+      const version = String(stdout ?? "").trim();
+      return {
+        available: true,
+        python: exe,
+        version: version || undefined,
+        method: "python_import_mlx_lm",
+      };
+    } catch {
+      /* try next */
+    }
+  }
+  return {
+    available: false,
+    python: preferred,
+    method: "python_import_mlx_lm",
+    error: `Could not import mlx_lm (tried: ${tried.join(", ")})`,
+  };
+}
+
 async function runLocalHealthJob(job) {
   const { jobId, payload } = job;
   const start = Date.now();
@@ -269,15 +321,66 @@ async function runLocalHealthJob(job) {
       }
     }
     const mlxRoot = normalizeOpenAiV1Root(payload?.mlxOpenAiBaseUrl);
-    let mlx = null;
+    let mlxHttp = null;
+    let mlxNative = null;
+
     if (mlxRoot) {
       const mlxStart = Date.now();
-      console.log(`[gonext-worker] local_health ${jobId} check mlx ${mlxRoot}`);
-      mlx = await checkOpenAiModels(mlxRoot, payload?.mlxApiKey ?? "");
+      console.log(`[gonext-worker] local_health ${jobId} check mlx HTTP ${mlxRoot}`);
+      mlxHttp = await checkOpenAiModels(mlxRoot, payload?.mlxApiKey ?? "");
       console.log(
-        `[gonext-worker] local_health ${jobId} mlx result ${mlxRoot} online=${mlx.online} models=${mlx.models.length} took=${((Date.now() - mlxStart) / 1000).toFixed(2)}s`
+        `[gonext-worker] local_health ${jobId} mlx HTTP online=${mlxHttp.online} models=${mlxHttp.models.length} took=${((Date.now() - mlxStart) / 1000).toFixed(2)}s`
       );
     }
+
+    const wantNativeFallback =
+      mlxRoot &&
+      payload?.mlxNativeFallback !== false &&
+      platform() === "darwin" &&
+      (!mlxHttp?.online || (mlxHttp?.models?.length ?? 0) === 0);
+
+    if (wantNativeFallback) {
+      const t0 = Date.now();
+      console.log(
+        `[gonext-worker] local_health ${jobId} mlx native probe (Python mlx_lm import)`
+      );
+      mlxNative = await checkMlxLmNativeImport();
+      console.log(
+        `[gonext-worker] local_health ${jobId} mlx native available=${mlxNative.available} took=${((Date.now() - t0) / 1000).toFixed(2)}s`
+      );
+    }
+
+    let mlx = null;
+    if (mlxRoot || mlxNative?.available) {
+      const httpOk = Boolean(mlxHttp?.online && (mlxHttp?.models?.length ?? 0) > 0);
+      const nativeOk = mlxNative?.available === true;
+      mlx = {
+        configured: httpOk || nativeOk,
+        online: httpOk || nativeOk,
+        models: httpOk
+          ? mlxHttp.models
+          : nativeOk
+            ? [
+                {
+                  id: "mlx_lm_native",
+                  name: mlxNative.version
+                    ? `MLX LM (${mlxNative.version})`
+                    : "MLX LM (Python import OK)",
+                  value: "mlx:mlx_lm_native",
+                },
+              ]
+            : [],
+        endpoint: mlxHttp?.endpoint,
+        http: mlxHttp
+          ? {
+              online: mlxHttp.online,
+              endpoint: mlxHttp.endpoint,
+              modelCount: mlxHttp.models.length,
+            }
+          : undefined,
+        native: mlxNative ?? undefined,
+      };
+    }
     const result = {
       ollama:
         ollamaBases.length > 0
@@ -289,14 +392,7 @@ async function runLocalHealthJob(job) {
               sources: ollamaSources,
             }
           : undefined,
-      mlx: mlx
-        ? {
-            configured: mlx.models.length > 0,
-            online: mlx.online,
-            models: mlx.models,
-            endpoint: mlx.endpoint,
-          }
-        : undefined,
+      mlx,
     };
     const totalTimeSeconds = (Date.now() - start) / 1000;
     await workerFetch(`/api/worker/jobs/${jobId}`, {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.11",
+  "version": "1.0.13",
   "description": "Polls GoNext cloud API for async local LLM jobs and runs them against Ollama/OpenAI-compatible servers on this Mac",
   "type": "module",
   "license": "MIT",