offgrid-ai 0.9.6 → 0.10.1

package/src/profile-setup.mjs

@@ -8,6 +8,7 @@ import { pc, formatBytes, renderRows, renderSection } from "./ui.mjs";
 import { detectCapabilities } from "./autodetect.mjs";
 import { matchDrafter } from "./scan.mjs";
 import { scanGgufModels } from "./scan.mjs";
+import { estimateMemoryMb } from "./mlx-flags.mjs";
 
 const execFileAsync = promisify(execFile);
 
@@ -297,3 +298,107 @@ function detectionSummary(caps) {
 function samplingSummary(flags) {
   return `temp ${flags.temperature}, top-p ${flags.topP}, top-k ${flags.topK}`;
 }
+
+// ── MLX profile configuration ─────────────────────────────────────────────
+
+/**
+ * Interactive configuration for an mlx-vlm profile.
+ */
+export async function configureMlxProfile(prompt, profile) {
+  let configured = profile;
+
+  console.log("");
+  console.log(renderSection("Model setup", renderRows([
+    ["Model", pc.bold(profile.label)],
+    ["Detected", mlxDetectionSummary(configured.capabilities)],
+    ["Context", String(configured.flags.ctxSize) + " tokens"],
+  ])));
+  console.log(pc.dim("Larger context windows use more memory. You can edit the profile later if needed.\n"));
+
+  if (configured.capabilities.vision) {
+    console.log(renderSection("Vision detected", renderRows([
+      ["Capability", "image / multimodal input"],
+      ["Note", "mlx-vlm loads vision from the model directory automatically."],
+    ])));
+  }
+
+  if (configured.capabilities.thinking) {
+    console.log("");
+    console.log(renderSection("Thinking mode", renderRows([
+      ["Flag", "--enable-thinking"],
+      ["Default", "on for Qwen 3 / Gemma 4 / DeepSeek-R class models"],
+    ])));
+    const useThinking = await prompt.yesNo("Enable thinking mode?", true);
+    configured = await applyMlxThinkingToggle(configured, useThinking);
+  }
+
+  const ctxSize = await prompt.number("Context window tokens", configured.flags.ctxSize, 1024, 1048576);
+  configured = applyMlxContextSize(configured, ctxSize);
+
+  console.log("\n" + renderMlxMemoryEstimate(configured));
+
+  console.log("");
+  console.log(renderSection("Defaults", renderRows([
+    ["Backend", configured.backend],
+    ["Endpoint", configured.baseUrl],
+    ["Context", String(configured.flags.ctxSize) + " tokens"],
+    ["Thinking", configured.capabilities.thinking && configured.commandArgv.includes("--enable-thinking") ? "on" : "off"],
+    ["Vision", configured.capabilities.vision ? "yes" : "no"],
+  ])));
+
+  if (!(await prompt.yesNo("Save profile with these settings?", true))) return null;
+  return configured;
+}
+
+async function applyMlxThinkingToggle(profile, enabled) {
+  if (!profile.capabilities.thinking) return profile;
+  const { computeMlxVlmFlags } = await import("./mlx-flags.mjs");
+  const { args } = computeMlxVlmFlags(profile.modelPath, {
+    port: profile.flags.port,
+    ctxSize: profile.flags.ctxSize,
+    thinkingEnabled: enabled,
+  });
+  return {
+    ...profile,
+    commandArgv: args,
+    capabilities: { ...profile.capabilities, thinkingEnabled: enabled },
+  };
+}
+
+function applyMlxContextSize(profile, ctxSize) {
+  const flags = { ...profile.flags, ctxSize };
+  const next = {
+    ...profile,
+    flags,
+    baseUrl: baseUrlForFlags(flags),
+  };
+  const idx = next.commandArgv.indexOf("--max-kv-size");
+  if (idx !== -1 && next.commandArgv[idx + 1] != null) {
+    next.commandArgv[idx + 1] = String(ctxSize);
+  } else if (ctxSize && ctxSize > 0) {
+    next.commandArgv.push("--max-kv-size", String(ctxSize));
+  }
+  return next;
+}
+
+function renderMlxMemoryEstimate(profile) {
+  const modelBytes = profile.modelSizeBytes || 0;
+  if (!modelBytes) {
+    return renderSection("Memory estimate", pc.dim("Model size unknown — save the profile to estimate."));
+  }
+  const totalMb = estimateMemoryMb(modelBytes);
+  const overheadBytes = Math.max(0, totalMb * 1024 * 1024 - modelBytes);
+  return renderSection("Memory estimate", renderRows([
+    ["Estimated total", pc.bold(`~${formatBytes(totalMb * 1024 * 1024)}`)],
+    ["Model", formatBytes(modelBytes)],
+    ["Overhead", `~${formatBytes(overheadBytes)} (KV cache, APC, runtime)`],
+  ]));
+}
+
+function mlxDetectionSummary(caps) {
+  const parts = [];
+  if (caps.architecture) parts.push(caps.architecture);
+  if (caps.thinking) parts.push("thinking");
+  if (caps.vision) parts.push("vision");
+  return parts.length > 0 ? parts.join(" · ") : "standard MLX";
+}

package/src/profiles.mjs

@@ -4,6 +4,8 @@ import { join } from "node:path";
 import { PROFILE_DIR, RUN_DIR, LOG_DIR } from "./config.mjs";
 import { backendFor, baseUrlForFlags, defaultFlagsForBackend } from "./backends.mjs";
 import { computeFlags } from "./autodetect.mjs";
+import { detectMlxCapabilities, defaultMlxContextLength } from "./mlx-discovery.mjs";
+import { detectHardware } from "./hardware.mjs";
 import { readJson, writeJson } from "./json.mjs";
 
 // ── Path helpers ───────────────────────────────────────────────────────────
@@ -161,6 +163,34 @@ export async function createProfileFromModel(model, backendId, drafterPath) {
   });
 }
 
+// ── Auto-create profile from a discovered MLX model ────────────────────────
+
+export async function createProfileFromMlxModel(model) {
+  const { computeMlxVlmFlags, DEFAULT_PORT } = await import("./mlx-flags.mjs");
+  const caps = await detectMlxCapabilities(model.filePath);
+  const ctxSize = defaultMlxContextLength(caps.contextLength, detectHardware().totalRamBytes / (1024 ** 3));
+  const { args } = computeMlxVlmFlags(model.filePath, {
+    port: DEFAULT_PORT,
+    ctxSize,
+    thinkingEnabled: caps.thinking,
+  });
+  return normalizeProfile({
+    id: slugFromLabel(model.label),
+    label: model.label,
+    backend: "mlx-vlm",
+    providerId: "mlx-vlm",
+    modelAlias: model.label,
+    source: model.source,
+    modelPath: model.filePath,
+    mmprojPath: null,
+    drafterPath: null,
+    modelSizeBytes: model.sizeBytes,
+    capabilities: caps,
+    flags: { host: "127.0.0.1", port: DEFAULT_PORT, ctxSize },
+    commandArgv: args,
+  });
+}
+
 function summarizeCapabilities(caps) {
   return {
     architecture: caps.architecture,

package/src/recommendations.mjs

@@ -1,17 +1,59 @@
-import { totalmem } from "node:os";
-
-const MODEL_TIERS = [
-  { maxGB: 8, lms: "google/gemma-4-e2b", ollama: "gemma4:e2b", label: "Gemma 4 E2B (2B effective)" },
-  { maxGB: 16, lms: "google/gemma-4-e4b", ollama: "gemma4:e4b", label: "Gemma 4 E4B (4B effective)" },
-  { maxGB: 32, lms: "qwen/qwen3.5-9b", ollama: "qwen3.5:9b-q4_K_M", label: "Qwen 3.5 9B" },
-  { maxGB: Infinity, lms: "qwen/qwen3.6-35b-a3b", ollama: "qwen3.6:35b-a3b", label: "Qwen 3.6 35B-A3B" },
-];
-
-export function recommendedModel() {
-  const gb = totalmem() / (1024 ** 3);
-  return MODEL_TIERS.find((tier) => gb <= tier.maxGB) ?? MODEL_TIERS[MODEL_TIERS.length - 1];
+import { readFileSync } from "node:fs";
+import { dirname, join } from "node:path";
+import { fileURLToPath } from "node:url";
+import { detectHardware } from "./hardware.mjs";
+
+const GB = 1024 ** 3;
+
+const RECOMMENDATIONS_PATH = join(dirname(fileURLToPath(import.meta.url)), "..", "resources", "recommendations.json");
+
+function loadRecommendations() {
+  try {
+    const raw = readFileSync(RECOMMENDATIONS_PATH, "utf8");
+    return JSON.parse(raw).models ?? [];
+  } catch {
+    return [];
+  }
+}
+
+/** All curated model entries. */
+export function getModelEntries() {
+  return loadRecommendations();
+}
+
+/** Recommend models that fit the detected hardware (max tier first). */
+export function recommendModels(hardware) {
+  const entries = loadRecommendations();
+  const fitting = entries.filter((e) => e.minRamGb * GB <= hardware.totalRamBytes);
+  if (fitting.length === 0) return [];
+  const maxTier = Math.max(...fitting.map((e) => e.minRamGb));
+  // All models at the top fitting tier are genuine alternatives; sort by label
+  // so the pick is deterministic regardless of JSON order.
+  return fitting
+    .filter((e) => e.minRamGb === maxTier)
+    .sort((a, b) => a.label.localeCompare(b.label));
+}
+
+/** Pick the best format for the platform. */
+export function selectFormat(entry, hardware) {
+  if (hardware.platform === "darwin" && hardware.arch === "arm64") {
+    if (entry.mlx) return "mlx";
+    if (entry.gguf) return "gguf";
+  } else {
+    if (entry.gguf) return "gguf";
+  }
+  return null;
+}
+
+/** Primary recommendation for this machine. */
+export function recommendedModel(hardware) {
+  const fitting = recommendModels(hardware ?? detectHardware());
+  return fitting[0] ?? null;
 }
 
-export function installedRamGB() {
-  return (totalmem() / (1024 ** 3)).toFixed(0);
+/** All models that fit, sorted best-first (tier desc, then label). */
+export function allFittingModels(hardware) {
+  const entries = loadRecommendations();
+  const fitting = entries.filter((e) => e.minRamGb * GB <= hardware.totalRamBytes);
+  return fitting.sort((a, b) => b.minRamGb - a.minRamGb || a.label.localeCompare(b.label));
 }

package/src/scan.mjs

@@ -1,9 +1,10 @@
 import { statSync } from "node:fs";
-import { readdir } from "node:fs/promises";
+import { readdir, stat } from "node:fs/promises";
 import { basename, dirname, join } from "node:path";
 import { getModelScanDirs } from "./config.mjs";
 import { readGgufMetadata } from "./gguf.mjs";
 import { parseModelName } from "./model-name.mjs";
+import { inferSourceLabel, MIN_MODEL_SIZE_BYTES, EMBEDDING_MODEL_TYPES } from "./discovery-shared.mjs";
 
 // ── Scan for GGUF models and MTP drafters ────────────────────────────────
 
@@ -13,7 +14,8 @@ export async function scanGgufModels(dirs) {
   const allDrafters = [];
 
   for (const root of scanDirs) {
-    const { models, drafters } = await scanOneDir(root);
+    const sourceLabel = inferSourceLabel(root);
+    const { models, drafters } = await scanOneDir(root, sourceLabel);
     allModels.push(...models);
     allDrafters.push(...drafters);
   }
@@ -36,7 +38,7 @@ export async function scanGgufModels(dirs) {
   return { models, drafters };
 }
 
-async function scanOneDir(root) {
+async function scanOneDir(root, sourceLabel = "local-gguf") {
   const files = await findFiles(root, (path) => path.toLowerCase().endsWith(".gguf"));
   const mmprojs = files.filter((path) => basename(path).toLowerCase().includes("mmproj"));
   const candidates = files.filter((path) => !basename(path).toLowerCase().includes("mmproj"));
@@ -49,11 +51,17 @@ async function scanOneDir(root) {
     const mmprojPath = mmprojs.find((candidate) => dirname(candidate) === dir) ?? null;
     const name = basename(path).replace(/\.gguf$/i, "");
     const sizeBytes = statSync(path).size;
+    if (sizeBytes < MIN_MODEL_SIZE_BYTES) continue;
     const parsed = parseModelName(name, "local-gguf");
 
-    // Read GGUF metadata to detect drafter architecture
+    // Read GGUF metadata to detect drafter architecture and embeddings
     const meta = safeReadGgufMetadata(path);
     const architecture = typeof meta["general.architecture"] === "string" ? meta["general.architecture"] : null;
+    const contextLength = architecture && typeof meta[`${architecture}.context_length`] === "number"
+      ? meta[`${architecture}.context_length`]
+      : null;
+
+    if (isEmbeddingArchitecture(architecture, name)) continue;
 
     if (architecture === "gemma4-assistant" || architecture === "gemma4_assistant") {
       // This is an MTP drafter model, not a main model
@@ -66,7 +74,7 @@ async function scanOneDir(root) {
         architecture,
         targetHint: drafterTargetHint(name),
         backend: "llama-cpp",
-        source: "local-gguf",
+        source: sourceLabel,
       });
     } else {
       models.push({
@@ -76,8 +84,9 @@ async function scanOneDir(root) {
         aliasSuggestion: parsed.id,
         quant: parsed.quant,
         sizeBytes,
+        contextLength,
         backend: "llama-cpp",
-        source: "local-gguf",
+        source: sourceLabel,
       });
     }
   }
@@ -85,6 +94,26 @@ async function scanOneDir(root) {
   return { models, drafters };
 }
 
+// ── Embedding model filtering ─────────────────────────────────────────────
+
+const EMBEDDING_FILENAME_PATTERNS = [
+  /(?:^|[-_])bge[-_]/i,
+  /(?:^|[-_])jina[-_]/i,
+  /(?:^|[-_])e5[-_]/i,
+  /(?:^|[-_])gte[-_]/i,
+  /(?:^|[-_])all[-_]minilm/i,
+  /(?:^|[-_])mpnet/i,
+  /(?:^|[-_])nomic[-_]embed/i,
+  /(?:^|[-_])embed/i,
+  /(?:^|[-_])rerank/i,
+];
+
+export function isEmbeddingArchitecture(architecture, filename = "") {
+  if (architecture && EMBEDDING_MODEL_TYPES.has(architecture.toLowerCase())) return true;
+  const lowerName = filename.toLowerCase();
+  return EMBEDDING_FILENAME_PATTERNS.some((pattern) => pattern.test(lowerName));
+}
+
 // ── Match drafters to target models ────────────────────────────────────
 
 // Map a drafter filename to a regex that matches its target model filenames.
@@ -137,8 +166,14 @@ async function findFiles(root, predicate) {
     }
     for (const entry of entries) {
       const path = join(dir, entry.name);
-      if (entry.isDirectory()) await walk(path);
-      else if (entry.isFile() && predicate(path)) result.push(path);
+      if (entry.isDirectory() || entry.isSymbolicLink()) {
+        // Follow symlinks (HF cache uses them) and avoid recursion loops.
+        const stats = await stat(path).catch(() => null);
+        if (stats?.isDirectory()) await walk(path);
+        else if (stats?.isFile() && predicate(path)) result.push(path);
+      } else if (entry.isFile() && predicate(path)) {
+        result.push(path);
+      }
     }
   }
   await walk(root);