offgrid-ai 0.4.5 → 0.4.7

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "offgrid-ai",
-  "version": "0.4.5",
+  "version": "0.4.7",
   "description": "Privacy-first CLI for running local LLMs — discover, configure, run, benchmark",
   "author": "Eeshan Srivastava (https://eeshans.com)",
   "type": "module",

package/src/cli.mjs

@@ -637,12 +637,15 @@ async function runProfile(profile, options = {}) {
   // Show memory estimate for local models
   if (!isManaged && profile.modelPath && existsSync(profile.modelPath)) {
     try {
-      const est = estimateMemory(profile.modelPath, profile.mmprojPath, null, profile.flags);
-      console.log(renderSection("Memory estimate", renderRows([
+      const est = estimateMemory(profile.modelPath, profile.mmprojPath, profile.drafterPath, profile.flags);
+      const rows = [
         ["Estimated total", pc.bold(`~${formatBytes(est.totalBytes)}`)],
         ["Model file", formatBytes(est.modelBytes)],
-        ["Conversation memory", est.kvBytes ? `~${formatBytes(est.kvBytes)}` : "unknown"],
-      ])));
+      ];
+      if (est.draftBytes) rows.push(["Drafter", formatBytes(est.draftBytes)]);
+      if (est.mmprojBytes) rows.push(["Vision projector", formatBytes(est.mmprojBytes)]);
+      rows.push(["Conversation memory", est.kvBytes ? `~${formatBytes(est.kvBytes)}` : "unknown"]);
+      console.log(renderSection("Memory estimate", renderRows(rows)));
     } catch { /* estimate failed, skip */ }
   }
 

package/src/logs.mjs

@@ -32,6 +32,13 @@ export function tailFriendly(rawLogPath, friendlyLogPath) {
 function friendlyLine(line) {
   const lower = line.toLowerCase();
   const trimmed = line.trim();
+  // Known-harmless errors during MTP memory estimation — llama.cpp tries to measure
+  // the draft model's memory usage before allocating the shared KV cache, which
+  // fails because the assistant architecture needs the trunk context. This is
+  // expected and the server proceeds to load the draft model correctly afterward.
+  if (lower.includes("ctx_other") || lower.includes("failed to measure draft model memory")) {
+    return pc.dim(`[mtp] ${trimmed}`);
+  }
   if (lower.includes("error") || lower.includes("failed")) return pc.red(`[error] ${trimmed}`);
   if (lower.includes("listening") || lower.includes("http server")) return pc.green(`[server] ${trimmed}`);
   if (lower.includes("llm_load") || lower.includes("load_model") || lower.includes("loading model")) return pc.cyan(`[load] ${trimmed}`);

package/src/profile-setup.mjs

@@ -4,6 +4,9 @@ import { estimateMemory } from "./estimate.mjs";
 import { findLlamaServer } from "./config.mjs";
 import { baseUrlForFlags, LLAMA_CPP_PORT, LLAMA_CPP_MTP_PORT } from "./backends.mjs";
 import { pc, formatBytes, renderRows, renderSection } from "./ui.mjs";
+import { detectCapabilities } from "./autodetect.mjs";
+import { matchDrafter } from "./scan.mjs";
+import { scanGgufModels } from "./scan.mjs";
 
 const execFileAsync = promisify(execFile);
 
@@ -29,7 +32,28 @@ const THINKING_DEFAULTS = {
 
 export async function configureLocalProfile(prompt, profile) {
   let configured = profile;
-  const caps = profile.capabilities ?? {};
+  // Re-detect capabilities from the model file and check for drafters
+  // so that re-setup can pick up MTP availability, vision changes, etc.
+  const freshCaps = detectCapabilities(profile.modelPath, profile.mmprojPath);
+  let drafterPath = profile.drafterPath ?? null;
+  if (!drafterPath) {
+    const { drafters } = await scanGgufModels();
+    const drafter = matchDrafter(profile.modelPath, drafters);
+    if (drafter) drafterPath = drafter.path;
+  }
+  const hasMtp = freshCaps.mtp || Boolean(drafterPath);
+  const caps = { ...freshCaps, mtp: hasMtp };
+  // If MTP is newly available, switch backend and add drafter path
+  if (hasMtp && configured.backend !== "llama-cpp-mtp") {
+    configured = { ...configured, backend: "llama-cpp-mtp", providerId: "llama-cpp-mtp", drafterPath, capabilities: { ...configured.capabilities, mtp: true } };
+  }
+  if (drafterPath && !configured.drafterPath) {
+    configured = { ...configured, drafterPath };
+  }
+  // If vision was previously disabled but mmproj is back, re-enable
+  if (configured.disabledMmprojPath && configured.mmprojPath === null && freshCaps.vision) {
+    configured = { ...configured, mmprojPath: configured.disabledMmprojPath, disabledMmprojPath: undefined, capabilities: { ...configured.capabilities, vision: true, visionDisabledReason: undefined } };
+  }
 
   console.log("");
   console.log(renderSection("Model setup", renderRows([