offgrid-ai 0.4.6 → 0.4.7

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "offgrid-ai",
-  "version": "0.4.6",
+  "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}`);