@tiens.nguyen/gonext-local-worker 1.0.10 → 1.0.12

package/gonext-local-worker.mjs

@@ -6,8 +6,9 @@
  * - `gonext-local-worker` starts polling loop
  */
 import { mkdir, readFile, writeFile } from "node:fs/promises";
-import { homedir } from "node:os";
+import { homedir, platform } from "node:os";
 import { join } from "node:path";
+import { execFile } from "node:child_process/promises";
 import dotenv from "dotenv";
 import OpenAI from "openai";
 
@@ -28,6 +29,9 @@ 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)
 `);
 }
 
@@ -179,15 +183,28 @@ function normalizeOpenAiV1Root(raw) {
   return /\/v1$/i.test(base) ? base : `${base}/v1`;
 }
 
+function sourceLabelFromBase(base) {
+  try {
+    return new URL(base).host || base;
+  } catch {
+    return base;
+  }
+}
+
 async function checkOllamaTags(base) {
   const endpoint = `${base}/api/tags`;
   try {
     const res = await fetch(endpoint, { method: "GET" });
     if (!res.ok) return { online: false, endpoint, models: [] };
     const j = await res.json();
+    const source = sourceLabelFromBase(base);
     const models = (j.models ?? []).map((m) => {
       const name = m.name ?? m.model ?? "model";
-      return { id: name, name, value: `ollama:${name}` };
+      return {
+        id: `${name}@@${source}`,
+        name: `${name} (${source})`,
+        value: `ollama:${name}@@${encodeURIComponent(base)}`,
+      };
     });
     return { online: true, endpoint, models };
   } catch {
@@ -215,9 +232,59 @@ 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();
+  const ollamaPayloadCount = Array.isArray(payload?.ollamaBaseUrls)
+    ? payload.ollamaBaseUrls.length
+    : 0;
+  console.log(
+    `[gonext-worker] local_health ${jobId} start (ollamaUrls=${ollamaPayloadCount}, mlx=${payload?.mlxOpenAiBaseUrl ? "yes" : "no"})`
+  );
   await workerFetch(`/api/worker/jobs/${jobId}`, {
     method: "PATCH",
     body: JSON.stringify({ jobStatus: "running" }),
@@ -229,18 +296,87 @@ async function runLocalHealthJob(job) {
     const dedup = new Map();
     let ollamaOnline = false;
     let ollamaEndpoint = "";
+    const ollamaSources = [];
     for (const base of ollamaBases) {
+      const baseStart = Date.now();
+      console.log(`[gonext-worker] local_health ${jobId} check ollama ${base}`);
       const r = await checkOllamaTags(base);
+      console.log(
+        `[gonext-worker] local_health ${jobId} ollama result ${base} online=${r.online} models=${r.models.length} took=${((Date.now() - baseStart) / 1000).toFixed(2)}s`
+      );
       ollamaOnline = ollamaOnline || r.online;
       if (!ollamaEndpoint) ollamaEndpoint = r.endpoint;
+       ollamaSources.push({
+        base,
+        label: sourceLabelFromBase(base),
+        endpoint: r.endpoint,
+        online: r.online,
+      });
       for (const m of r.models) {
         if (!dedup.has(m.value)) dedup.set(m.value, m);
       }
     }
     const mlxRoot = normalizeOpenAiV1Root(payload?.mlxOpenAiBaseUrl);
-    const mlx = mlxRoot
-      ? await checkOpenAiModels(mlxRoot, payload?.mlxApiKey ?? "")
-      : null;
+    let mlxHttp = null;
+    let mlxNative = null;
+
+    if (mlxRoot) {
+      const mlxStart = Date.now();
+      console.log(`[gonext-worker] local_health ${jobId} check mlx HTTP ${mlxRoot}`);
+      mlxHttp = await checkOpenAiModels(mlxRoot, payload?.mlxApiKey ?? "");
+      console.log(
+        `[gonext-worker] local_health ${jobId} mlx HTTP online=${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
@@ -249,16 +385,10 @@ async function runLocalHealthJob(job) {
               online: ollamaOnline,
               models: [...dedup.values()],
               endpoint: ollamaEndpoint,
+              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}`, {
@@ -270,8 +400,9 @@ async function runLocalHealthJob(job) {
         totalTimeSeconds,
       }),
     });
+    const onlineCount = ollamaSources.filter((s) => s.online).length;
     console.log(
-      `[gonext-worker] completed local_health ${jobId} (${totalTimeSeconds.toFixed(1)}s)`
+      `[gonext-worker] completed local_health ${jobId} (${totalTimeSeconds.toFixed(1)}s) summary: ollamaOnline=${onlineCount}/${ollamaSources.length}, mlx=${mlx ? (mlx.online ? "online" : "offline") : "n/a"}`
     );
   } catch (e) {
     const message = e instanceof Error ? e.message : String(e);
@@ -296,6 +427,9 @@ async function pollOnce() {
   }
   const job = await res.json();
   if (job?.jobId) {
+    console.log(
+      `[gonext-worker] claimed ${job.jobId} type=${job.jobType ?? "unknown"} modelKey=${job.modelKey ?? "unknown"}`
+    );
     const isLocalHealthByType = job.jobType === "local_health";
     const isLocalHealthByModelKey = job.modelKey === "local_health";
     const isLocalHealthByPayload =

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.10",
+  "version": "1.0.12",
   "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",