offgrid-ai 0.9.5 → 0.10.0

package/src/mlx-flags.mjs

@@ -0,0 +1,93 @@
+// mlx-vlm server flag computation — pure functions, no side effects.
+// Ported from deprecated-offgrid-desktop/src/main/server-flags.ts (MLX subset).
+//
+// Benchmark-informed decisions (see sidequests/mlx-backend-benchmark/RESULTS.md):
+// - mlx-vlm requires APC_ENABLED=1 env var (86x TTFT improvement) — set at spawn
+//   time in process.mjs, NOT here (this module only computes args).
+// - mlx-vlm uses a strict=False wrapper script for shared-KV architectures
+//   (Gemma 4-class). Safe for all models — strict=False is a no-op for models
+//   that load fine with strict=True.
+// - mlx-vlm uses --enable-thinking for thinking-mode control.
+// - mlx-vlm uses --max-kv-size for the KV cache / context window.
+//
+// Only the mlx-vlm-relevant logic is ported here. offgrid-ai's existing GGUF
+// flag logic (autodetect.mjs / profile-setup.mjs / estimate.mjs) is unchanged.
+
+import { fileURLToPath } from "node:url";
+import { dirname, join } from "node:path";
+
+const MB = 1024 ** 2;
+
+/** Default port for the local model server. Matches the desktop's DEFAULT_PORT. */
+export const DEFAULT_PORT = 18080;
+
+/** Resolved path to the bundled strict=False wrapper script (sibling of src/). */
+export const MLX_VLM_WRAPPER = join(dirname(fileURLToPath(import.meta.url)), "..", "resources", "mlxvlm-server-wrapper.py");
+
+/** Overhead multiplier for mlx-vlm: weights × 1.5 (covers KV cache, activations, APC cache; benchmark-validated). */
+const MLX_VLM_OVERHEAD_MULTIPLIER = 1.5;
+
+/** Server process overhead in MB. */
+const PROCESS_OVERHEAD_MB = 200;
+
+/**
+ * Estimate mlx-vlm memory usage (MB): model weights × 1.5 + process overhead.
+ *
+ * The 1.5 multiplier covers KV cache, activations, and APC cache overhead
+ * (benchmark-validated; see sidequests/mlx-backend-benchmark/RESULTS.md).
+ * GGUF/llama-server estimation uses the detailed path in estimate.mjs.
+ *
+ * @param {number} fileSizeBytes - model size on disk (sum of MLX safetensors).
+ * @returns {number} estimated memory in MB.
+ */
+export function estimateMemoryMb(fileSizeBytes) {
+  return Math.round((fileSizeBytes / MB) * MLX_VLM_OVERHEAD_MULTIPLIER + PROCESS_OVERHEAD_MB);
+}
+
+/**
+ * Compute mlx-vlm server arguments.
+ *
+ * mlx-vlm is the MLX-native server (benchmark-validated best throughput + memory
+ * efficiency on Apple Silicon). Invoked via the strict=False wrapper script for
+ * compatibility with shared-KV architectures (Gemma 4-class).
+ *
+ * The APC_ENABLED=1 env var is MANDATORY but is set at spawn time in
+ * process.mjs, not in args.
+ *
+ * The wrapper script (resources/mlxvlm-server-wrapper.py) applies strict=False
+ * model loading + the BatchRotatingKVCache.merge() fix, both required for
+ * shared-KV architectures (Gemma 4-class). It is resolved to a real path via
+ * MLX_VLM_WRAPPER; there is intentionally no raw-mlx_vlm.server path.
+ *
+ * @param {string} modelPath - path to the MLX model directory.
+ * @param {object} [options]
+ * @param {number} [options.port] - port (default DEFAULT_PORT).
+ * @param {number} [options.ctxSize] - context window (passed as --max-kv-size).
+ * @param {boolean} [options.thinkingEnabled=true] - whether to enable thinking.
+ * @returns {{ args: string[], port: number }}
+ */
+export function computeMlxVlmFlags(modelPath, options = {}) {
+  const port = options.port ?? DEFAULT_PORT;
+  const ctxSize = options.ctxSize;
+  const thinkingEnabled = options.thinkingEnabled ?? true;
+
+  // The binary is "python3" (resolved by backendBinaryFor in backends.mjs); the
+  // wrapper path is the first arg.
+  const args = [
+    MLX_VLM_WRAPPER,
+    "--model", modelPath,
+    "--host", "127.0.0.1",
+    "--port", String(port),
+  ];
+
+  if (thinkingEnabled) {
+    args.push("--enable-thinking");
+  }
+
+  // Context size: mlx-vlm uses --max-kv-size for the KV cache / context window.
+  if (ctxSize && ctxSize > 0) {
+    args.push("--max-kv-size", String(ctxSize));
+  }
+
+  return { args, port };
+}

package/src/model-catalog.mjs

@@ -1,29 +1,34 @@
 import { scanGgufModels, matchDrafter } from "./scan.mjs";
 import { loadProfiles, normalizeProfile, sanitizeProfileId } from "./profiles.mjs";
 import { scanManagedModels } from "./managed.mjs";
+import { scanMlxModels } from "./mlx-discovery.mjs";
 import { isProfileFileMissing } from "./model-summary.mjs";
 
 export async function loadModelCatalog() {
-  const [profiles, { models: ggufModels, drafters }, managedModels] = await Promise.all([
+  const [profiles, { models: ggufModels, drafters }, managedModels, mlxModels] = await Promise.all([
     loadProfiles(),
     scanGgufModels(),
     scanManagedModels(),
+    scanMlxModels(),
   ]);
-  return normalizeCatalog({ profiles, ggufModels, drafters, managedModels });
+  return normalizeCatalog({ profiles, ggufModels, drafters, managedModels, mlxModels });
 }
 
 export function normalizeCatalog(catalog) {
   if (catalog.newModels && catalog.managedItems) return catalog;
-  const { profiles, ggufModels, drafters, managedModels } = catalog;
+  const { profiles, ggufModels, drafters, managedModels, mlxModels = [] } = catalog;
   const profiledPaths = new Set(profiles.map((profile) => profile.modelPath).filter(Boolean));
-  const newModels = ggufModels.filter((model) => !profiledPaths.has(model.path));
+  const newModels = [
+    ...ggufModels.filter((model) => !profiledPaths.has(model.path)),
+    ...mlxModels.filter((model) => !profiledPaths.has(model.path)),
+  ];
   const managedItems = [];
   for (const { backendId, models, status } of managedModels) {
     if (status === "unavailable") continue;
     const profiledAliases = new Set(
       profiles
         .filter((profile) => profile.backend === backendId)
-        .map((profile) => backendId === "ollama" ? `ollama:${profile.ollamaModel ?? profile.modelAlias}` : `omlx:${profile.omlxModel ?? profile.modelAlias}`),
+        .map((profile) => `omlx:${profile.omlxModel ?? profile.modelAlias}`),
     );
     for (const model of models) {
       if (!profiledAliases.has(`${backendId}:${model.id}`)) managedItems.push({ model, backendId });
@@ -66,8 +71,9 @@ export function createManagedProfile(model, backendId) {
     id: `${backendId}-${sanitizeProfileId(model.id)}`,
     label: model.label,
     backend: backendId,
+    source: backendId,
     modelAlias: model.aliasSuggestion,
-    ...(backendId === "ollama" ? { ollamaModel: model.id } : {}),
+    modelSizeBytes: model.sizeBytes || 0,
     ...(backendId === "omlx" ? { omlxModel: model.id } : {}),
   });
 }

package/src/model-name.mjs

@@ -68,9 +68,8 @@ const TAG_TOKENS = [
 /**
  * Parse a raw model identifier into a structured display name.
  *
- * @param {string} rawId  The raw identifier: GGUF filename (no .gguf),
- *                        Ollama model name, or oMLX model id.
- * @param {"local-gguf"|"ollama"|"omlx"} source  Where this name came from.
+ * @param {string} rawId  The raw identifier: GGUF filename (no .gguf) or oMLX model id.
+ * @param {"local-gguf"|"omlx"} source  Where this name came from.
  * @returns {{ publisher: string|null, model: string, params: string|null,
  *             quant: string|null, tags: string[], display: string,
  *             sort: string, id: string }}
@@ -87,18 +86,7 @@ export function parseModelName(rawId, source) {
     name = rawId.slice(slashIdx + 1);
   }
 
-  // 2. For Ollama, split on : to separate model from tag (e.g. "gemma3:4b")
-  //    The tag after : is a model size/variant identifier — not a GGUF quant.
-  let ollamaTag = null;
-  if (source === "ollama") {
-    const colonIdx = name.lastIndexOf(":");
-    if (colonIdx !== -1) {
-      ollamaTag = name.slice(colonIdx + 1);
-      name = name.slice(0, colonIdx);
-    }
-  }
-
-  // 3. Extract quant (GGUF quantization suffix)
+  // 2. Extract quant (GGUF quantization suffix)
   let quant = null;
   for (const pattern of QUANT_PATTERNS) {
     const match = name.match(pattern);
@@ -125,13 +113,7 @@ export function parseModelName(rawId, source) {
   // Clean up leftover separators
   name = name.replace(/[-_]{2,}/g, "-").replace(/^[-_]+|[-_]+$/g, "");
 
-  // 5. For Ollama, re-attach the tag as part of the model name
-  //    (Ollama tags like "4b" or "30b-a3b" are size variants, not quants)
-  if (ollamaTag) {
-    name = name + "-" + ollamaTag;
-  }
-
-  // 6. Title-case the remaining model name
+  // 5. Title-case the remaining model name
   let model = titleCaseModel(name);
 
   // If nothing is left after parsing, fall back to the raw name
@@ -139,13 +121,13 @@ export function parseModelName(rawId, source) {
     model = rawId.includes("/") ? rawId : rawId.replace(/[-_]/g, " ");
   }
 
-  // 7. Extract params (size like 30B, 12B) for sort/filter convenience
+  // 6. Extract params (size like 30B, 12B) for sort/filter convenience
   const params = extractParams(model);
 
-  // 8. Build display string
+  // 7. Build display string
   const display = buildDisplay(publisher, model, tags, quant);
 
-  // 9. Build sort key (lowercase, no publisher, for alphabetical ordering)
+  // 8. Build sort key (lowercase, no publisher, for alphabetical ordering)
   const sort = model.toLowerCase().replace(/[-_]/g, " ");
 
   return { publisher, model, params, quant, tags, display, sort, id };

package/src/model-presenters.mjs

@@ -1,7 +1,8 @@
 import { existsSync, statSync } from "node:fs";
-import { BACKENDS, backendFor } from "./backends.mjs";
+import { basename, dirname } from "node:path";
+import { backendFor } from "./backends.mjs";
 import { readCommandArgv } from "./profiles.mjs";
-import { isProfileRunning, isProfileServerUp } from "./process.mjs";
+import { isProfileRunning } from "./process.mjs";
 import { buildPrettyCommand } from "./command.mjs";
 import { pc, formatBytes, renderRows, renderSection } from "./ui.mjs";
 import { capabilitySummary, ggufDetailParts, isProfileFileMissing, profileDetailParts } from "./model-summary.mjs";
@@ -11,6 +12,7 @@ import { findBenchmarkRepo } from "./benchmark.mjs";
 
 const OPTION_SEPARATOR = pc.dim("  │  ");
 const OPTION_STATUS_WIDTH = 10;
+const OPTION_BACKEND_WIDTH = 14;
 const OPTION_SOURCE_WIDTH = 14;
 const OPTION_CTX_WIDTH = 5;
 
@@ -25,7 +27,7 @@ function optionPad(text, color, width) {
 function optionStatusTag(kind) {
   const statuses = {
     running: ["RUNNING", pc.green],
-    serverup: ["SERVER UP", pc.yellow],
+    serverup: ["READY", pc.blue],
     ready: ["READY", pc.blue],
     missing: ["MISSING", pc.red],
     setup: ["SETUP", pc.yellow],
@@ -34,17 +36,70 @@ function optionStatusTag(kind) {
   return optionPad(text, color, OPTION_STATUS_WIDTH);
 }
 
-function optionSourceTag(sourceId, label) {
+function optionSourceTag(sourceId) {
+  const label = formatSourceLabel(sourceId);
   const colors = {
-    "llama-cpp": pc.cyan,
-    "llama-cpp-mtp": pc.blue,
-    ollama: pc.green,
+    huggingface: pc.cyan,
+    lmstudio: pc.blue,
     omlx: pc.magenta,
+    "llama.cpp": pc.cyan,
     gguf: pc.cyan,
+    mlx: pc.yellow,
+    "mlx-vlm": pc.yellow,
   };
   return optionPad(label, colors[sourceId] ?? pc.dim, OPTION_SOURCE_WIDTH);
 }
 
+function optionBackendTag(backendId) {
+  const backend = backendId ? backendFor(backendId) : null;
+  const label = backend?.label ?? backendId ?? "unknown";
+  const colors = {
+    "llama-cpp": pc.cyan,
+    "llama-cpp-mtp": pc.blue,
+    omlx: pc.magenta,
+    "mlx-vlm": pc.yellow,
+  };
+  return optionPad(label, colors[backendId] ?? pc.dim, OPTION_BACKEND_WIDTH);
+}
+
+function formatSourceLabel(sourceId) {
+  if (!sourceId) return "unknown";
+  const map = {
+    huggingface: "HuggingFace",
+    lmstudio: "LM Studio",
+    omlx: "oMLX",
+    "llama.cpp": "llama.cpp",
+    gguf: "GGUF file",
+    mlx: "MLX",
+    "mlx-vlm": "MLX",
+  };
+  return map[sourceId] ?? String(sourceId);
+}
+
+function inferSourceFromPath(modelPath) {
+  if (!modelPath) return null;
+  const normalized = modelPath.toLowerCase().replace(/\\/g, "/");
+  if (normalized.includes("/.omlx/models")) return "omlx";
+  if (normalized.includes("/.lmstudio/models")) return "lmstudio";
+  if (normalized.includes("/.cache/huggingface")) return "huggingface";
+  if (normalized.includes("/.cache/llama.cpp")) return "llama.cpp";
+  const parent = basename(dirname(modelPath));
+  if (parent && parent !== ".") return parent.replace(/^\./, "");
+  return null;
+}
+
+function discoverySourceForProfile(profile) {
+  const backend = backendFor(profile.backend);
+  if (backend.type === "managed-server") return backend.id;
+  if (profile.source && profile.source !== "local-gguf") return profile.source;
+  return inferSourceFromPath(profile.modelPath);
+}
+
+function discoverySourceForItem(item) {
+  if (item.type === "profile") return discoverySourceForProfile(item.profile);
+  return item.model?.source ?? null;
+}
+
 function optionCtxLabel(item) {
   if (item.type === "profile" && item.profile.flags?.ctxSize) {
     return optionPad(`${(item.profile.flags.ctxSize / 1000).toFixed(0)}k`, null, OPTION_CTX_WIDTH);
@@ -52,12 +107,18 @@ function optionCtxLabel(item) {
   return optionPad("—", null, OPTION_CTX_WIDTH);
 }
 
-function optionSizeLabel(item) {
+function optionSizeLabel(item, managedModels) {
   if (item.type === "profile") {
     if (item.fileMissing) return "—";
+    if (item.profile.modelSizeBytes) return formatBytes(item.profile.modelSizeBytes);
     if (item.profile.modelPath && existsSync(item.profile.modelPath)) {
-      return formatBytes(statSync(item.profile.modelPath).size);
+      const s = statSync(item.profile.modelPath);
+      // Only stat regular files — a modelPath that is a directory (MLX)
+      // reports the dir entry size, not the model size.
+      if (s.isFile()) return formatBytes(s.size);
     }
+    const managedSize = managedProfileSizeBytes(item.profile, managedModels);
+    if (managedSize) return formatBytes(managedSize);
     return "—";
   }
   if (item.type === "new") {
@@ -76,26 +137,34 @@ export function modelNameWidth(items) {
   return Math.max(20, maxName + 2);
 }
 
-function optionLabel({ status, source, name, ctx, size, nameWidth }) {
-  return [status, source, pc.bold(optionPad(name, null, nameWidth)), ctx, pc.dim(size)].join(OPTION_SEPARATOR);
+function optionLabel({ status, backend, source, name, ctx, size, nameWidth }) {
+  return [status, backend, source, pc.bold(optionPad(name, null, nameWidth)), ctx, pc.dim(size)].join(OPTION_SEPARATOR);
 }
 
-export function modelSelectOption(item, { runningProfilesNow, serverUpIds, nameWidth }) {
+export function modelSelectOption(item, { runningProfilesNow, modelMissingIds, nameWidth, managedModels }) {
+  const sourceId = discoverySourceForItem(item) ?? "unknown";
+  const backendId = inferBackendId(item);
   if (item.type === "profile") {
     const backend = backendFor(item.profile.backend);
     const running = runningProfilesNow.some((profile) => profile.id === item.profile.id);
-    const serverUp = !running && !item.fileMissing && serverUpIds?.has(item.profile.id);
-    const status = item.fileMissing ? "missing" : running ? "running" : serverUp ? "serverup" : "ready";
+    const modelMissing = !item.fileMissing && modelMissingIds?.has(item.profile.id);
+    const status = item.fileMissing || modelMissing ? "missing" : running ? "running" : "ready";
+    const drafterMissing = Boolean(item.profile.drafterPath) && !existsSync(item.profile.drafterPath);
+    const hint = drafterMissing ? "MTP drafter missing — reconfigure"
+      : modelMissing ? `${backend.label} model no longer available`
+      : undefined;
     return {
       value: itemKey(item),
       label: optionLabel({
         status: optionStatusTag(status),
-        source: optionSourceTag(item.profile.backend, backend.label),
+        backend: optionBackendTag(backendId),
+        source: optionSourceTag(sourceId),
         name: item.profile.label,
         nameWidth,
         ctx: optionCtxLabel(item),
-        size: optionSizeLabel(item),
+        size: optionSizeLabel(item, managedModels),
       }),
+      ...(hint ? { hint: pc.red(hint) } : {}),
     };
   }
   if (item.type === "new") {
@@ -103,41 +172,60 @@ export function modelSelectOption(item, { runningProfilesNow, serverUpIds, nameW
       value: itemKey(item),
       label: optionLabel({
         status: optionStatusTag("setup"),
-        source: optionSourceTag("gguf", "GGUF file"),
+        backend: optionBackendTag(backendId),
+        source: optionSourceTag(sourceId),
         name: item.model.label,
         nameWidth,
         ctx: optionCtxLabel(item),
-        size: optionSizeLabel(item),
+        size: optionSizeLabel(item, managedModels),
       }),
     };
   }
-  const backend = BACKENDS[item.backendId];
   return {
     value: itemKey(item),
     label: optionLabel({
       status: optionStatusTag("setup"),
-      source: optionSourceTag(item.backendId, backend.label),
+      backend: optionBackendTag(backendId),
+      source: optionSourceTag(sourceId),
       name: item.model.label,
       nameWidth,
       ctx: optionCtxLabel(item),
-      size: optionSizeLabel(item),
+      size: optionSizeLabel(item, managedModels),
     }),
   };
 }
 
-export function printWorkspaceHeader(normalized, runningProfilesNow, serverUpIds = new Set()) {
+function managedProfileSizeBytes(profile, managedModels) {
+  if (!managedModels || !Array.isArray(managedModels)) return null;
+  const backend = backendFor(profile.backend);
+  if (backend.type !== "managed-server") return null;
+  const backendModels = managedModels.find((m) => m.backendId === profile.backend)?.models ?? [];
+  const modelId = profile.omlxModel ?? null;
+  if (!modelId) return null;
+  const model = backendModels.find((m) => m.id === modelId);
+  return model?.sizeBytes || null;
+}
+
+function inferBackendId(item) {
+  if (item.type === "profile") return item.profile.backend;
+  if (item.type === "managed") return item.backendId;
+  // new model: derive from format
+  if (item.model?.format === "mlx") return "mlx-vlm";
+  if (item.model?.backend) return item.model.backend;
+  return "llama-cpp";
+}
+
+export function printWorkspaceHeader(normalized, runningProfilesNow, modelMissingIds = new Set()) {
   const profiles = normalized.profiles;
   const isRunning = (p) => runningProfilesNow.some((r) => r.id === p.id);
-  const isMissing = (p) => isProfileFileMissing(p);
-  const readyCount = profiles.filter((p) => !isMissing(p) && !isRunning(p) && !serverUpIds.has(p.id)).length;
+  const isMissing = (p) => isProfileFileMissing(p) || modelMissingIds.has(p.id);
+  const readyCount = profiles.filter((p) => !isMissing(p) && !isRunning(p)).length;
   const runningCount = runningProfilesNow.length;
-  const serverUpCount = profiles.filter((p) => !isMissing(p) && serverUpIds.has(p.id) && !isRunning(p)).length;
   const missingCount = profiles.filter(isMissing).length;
   const setupCount = normalized.newModels.length + normalized.managedItems.length;
 
   const countParts = [];
   if (runningCount > 0) countParts.push(pc.green(`${runningCount} running`));
-  if (serverUpCount > 0) countParts.push(pc.yellow(`${serverUpCount} server up, model not loaded`));
   if (readyCount > 0) countParts.push(pc.blue(`${readyCount} model${readyCount === 1 ? "" : "s"} ready`));
   if (missingCount > 0) countParts.push(pc.red(`${missingCount} model${missingCount === 1 ? "" : "s"} missing`));
   if (setupCount > 0) countParts.push(pc.yellow(`${setupCount} model${setupCount === 1 ? "" : "s"} need${setupCount === 1 ? "s" : ""} setup`));
@@ -160,11 +248,10 @@ export async function printProfileDetails(profile) {
   const backend = backendFor(profile.backend);
   const isManaged = backend.type === "managed-server";
   const running = await isProfileRunning(profile);
-  const serverUp = !running && isManaged && await isProfileServerUp(profile);
   const fileMissing = !isManaged && isProfileFileMissing(profile);
   console.log("\n" + renderSection("Model overview", renderRows([
     ["Name", pc.bold(profile.label)],
-    ["Status", fileMissing ? pc.red("File missing") : running ? pc.green("Running now") : serverUp ? pc.yellow("Server up, model not loaded") : pc.blue("Ready")],
+    ["Status", fileMissing ? pc.red("File missing") : running ? pc.green("Running now") : pc.blue("Ready")],
     ["Details", profileDetailParts(profile, { fileMissing }).join(pc.dim(" · "))],
     ["Server", fileMissing ? pc.red(profile.baseUrl) : profile.baseUrl],
   ])));
@@ -213,6 +300,29 @@ export function printGgufModelDetails(model, drafter) {
   console.log("\n" + renderSection("Model details", renderRows(detailRows), { columns: 110 }));
 }
 
+export async function printMlxModelDetails(model) {
+  const { detectMlxCapabilities } = await import("./mlx-discovery.mjs");
+  const caps = await detectMlxCapabilities(model.filePath ?? model.path);
+  const parts = [];
+  if (caps.architecture) parts.push(caps.architecture);
+  if (caps.thinking) parts.push("thinking");
+  if (caps.vision) parts.push("vision");
+  const summary = parts.length > 0 ? parts.join(pc.dim(" · ")) : "standard MLX";
+  console.log("\n" + renderSection("Downloaded model", renderRows([
+    ["Name", pc.bold(model.label)],
+    ["Status", pc.yellow("Needs one-time setup")],
+    ["Details", summary],
+  ])));
+  console.log("\n" + renderSection("Model details", renderRows([
+    ["Model dir", model.path],
+    ["Backend", "mlx-vlm"],
+    ["Source", formatSourceLabel(model.source)],
+    ["Detected", summary],
+    ["Size", formatBytes(model.sizeBytes)],
+    ["Context", caps.contextLength ? `${caps.contextLength.toLocaleString()} trained` : "unknown"],
+  ]), { columns: 110 }));
+}
+
 export function printManagedModelDetails(model, backend) {
   console.log("\n" + renderSection(`${backend.label} model`, renderRows([
     ["Name", pc.bold(model.label)],