offgrid-ai 0.3.11 → 0.3.12

package/README.md

@@ -51,7 +51,7 @@ curl -fsSL https://raw.githubusercontent.com/eeshansrivastava89/offgrid-ai/main/
 
 1. **Auto-detect everything.** Scans for GGUF models in LM Studio, HuggingFace, and Ollama directories. Reads model metadata (quantization, context size, vision, thinking mode) directly from the GGUF file. No presets, no manual configuration.
 
-2. **One command to run.** `offgrid-ai` → pick a model → it figures out the flags, starts llama-server, syncs Pi config, and launches Pi.
+2. **One command to run.** `offgrid-ai` → pick a model → confirm context/KV memory settings on first setup → it starts llama-server, syncs Pi config, and launches Pi.
 
 3. **One model at a time.** Laptops have limited RAM. One server, one model, no confusion.
 
@@ -77,7 +77,7 @@ When you run `offgrid-ai` for the first time on a fresh machine:
    - **oMLX** — Apple Silicon optimized
 4. **Models** — If no models found, tells you where to get them.
 
-Subsequent runs skip everything that's already installed.
+Subsequent runs skip everything that's already installed. When a GGUF model is set up for the first time, offgrid-ai asks only for the memory-impacting choices: context window and KV cache precision. Sampling defaults are shown but not forced into a tuning wizard.
 
 ## Data directory
 

package/package.json

@@ -1,11 +1,11 @@
 {
   "name": "offgrid-ai",
-  "version": "0.3.11",
+  "version": "0.3.12",
   "description": "Privacy-first CLI for running local LLMs — discover, configure, run, benchmark",
   "author": "Eeshan Srivastava (https://eeshans.com)",
   "type": "module",
   "bin": {
-    "offgrid-ai": "./bin/offgrid-ai.mjs"
+    "offgrid-ai": "bin/offgrid-ai.mjs"
   },
   "files": [
     "bin/*.mjs",

package/src/cli.mjs

@@ -12,6 +12,7 @@ import { estimateMemory } from "./estimate.mjs";
 import { pc, formatBytes, renderRows, renderSection, startInteractive, createPrompt, parseOptions } from "./ui.mjs";
 import { checkForUpdate, currentPackageVersion, detectInvocation, updateCommand, runUpdateCommand } from "./updates.mjs";
 import { removeInstallerPathEntries } from "./shell-path.mjs";
+import { configureLocalProfile } from "./profile-setup.mjs";
 
 // ── Entry point ────────────────────────────────────────────────────────────
 
@@ -270,10 +271,12 @@ async function pickAndRun(prompt, profiles, newModels, managedItems) {
     const model = newModels.find((m) => m.path === modelPath);
     if (!model) throw new Error("Model not found.");
     const profile = await createProfileFromModel(model);
-    await saveProfile(profile);
-    console.log(pc.green(`Auto-configured: ${profile.label}`));
-    await syncPiConfig(profile);
-    return await runProfile(profile);
+    const configured = await configureLocalProfile(prompt, profile);
+    if (!configured) return;
+    await saveProfile(configured);
+    console.log(pc.green(`Saved profile: ${configured.label}`));
+    await syncPiConfig(configured);
+    return await runProfile(configured);
   }
 
   if (selected.startsWith("managed:")) {

package/src/profile-setup.mjs

@@ -0,0 +1,82 @@
+import { estimateMemory } from "./estimate.mjs";
+import { pc, formatBytes, renderRows, renderSection } from "./ui.mjs";
+
+const CACHE_CHOICES = [
+  { value: "bf16", label: "bf16", hint: "default: stable, good quality" },
+  { value: "f16", label: "f16", hint: "stable fallback, similar memory to bf16" },
+  { value: "q8_0", label: "q8_0", hint: "lower memory, usually safe" },
+  { value: "q4_0", label: "q4_0", hint: "lowest memory, quality/speed tradeoff" },
+];
+
+export async function configureLocalProfile(prompt, profile) {
+  console.log("");
+  console.log(renderSection("Model setup", renderRows([
+    ["Model", pc.bold(profile.label)],
+    ["Context", `${profile.flags.ctxSize.toLocaleString()} tokens`],
+    ["KV cache", `${profile.flags.cacheTypeK}/${profile.flags.cacheTypeV}`],
+    ["Sampling", samplingSummary(profile.flags)],
+  ])));
+  console.log(pc.dim("Larger context windows use more memory. KV cache precision controls memory used by attention history."));
+  console.log(pc.dim("Sampling defaults are shown for transparency; you can edit command.json later if needed.\n"));
+
+  const ctxSize = await prompt.number("Context window tokens", profile.flags.ctxSize, 1024, 1048576);
+  const cacheTypeK = await prompt.choice("K cache precision", CACHE_CHOICES, profile.flags.cacheTypeK);
+  const cacheTypeV = await prompt.choice("V cache precision", CACHE_CHOICES, profile.flags.cacheTypeV);
+  const configured = applyRuntimeFlagOverrides(profile, { ctxSize, cacheTypeK, cacheTypeV });
+
+  console.log("");
+  console.log(renderSection("Defaults", renderRows([
+    ["Temperature", configured.flags.temperature],
+    ["Top-p", configured.flags.topP],
+    ["Top-k", configured.flags.topK],
+    ["Min-p", configured.flags.minP],
+    ["Presence penalty", configured.flags.presencePenalty],
+    ["Repeat penalty", configured.flags.repeatPenalty],
+  ])));
+
+  console.log("\n" + renderMemoryEstimate(configured));
+  if (!(await prompt.yesNo("Save profile with these settings?", true))) return null;
+  return configured;
+}
+
+export function applyRuntimeFlagOverrides(profile, overrides) {
+  const flags = { ...profile.flags, ...overrides };
+  return {
+    ...profile,
+    flags,
+    baseUrl: `http://${flags.host}:${flags.port}/v1`,
+    commandArgv: updateArgv(profile.commandArgv ?? [], {
+      "--ctx-size": flags.ctxSize,
+      "--cache-type-k": flags.cacheTypeK,
+      "--cache-type-v": flags.cacheTypeV,
+    }),
+  };
+}
+
+function updateArgv(argv, values) {
+  const next = [...argv];
+  for (const [flag, value] of Object.entries(values)) {
+    const index = next.indexOf(flag);
+    if (index === -1) next.push(flag, String(value));
+    else next[index + 1] = String(value);
+  }
+  return next;
+}
+
+function renderMemoryEstimate(profile) {
+  try {
+    const est = estimateMemory(profile.modelPath, profile.mmprojPath, null, profile.flags);
+    return renderSection("Memory", renderRows([
+      ["Estimated total", pc.bold(`~${formatBytes(est.totalBytes)}`)],
+      ["Model", formatBytes(est.modelBytes)],
+      ["KV cache", est.kvBytes ? `~${formatBytes(est.kvBytes)} (${profile.flags.ctxSize.toLocaleString()} ctx, ${profile.flags.cacheTypeK}/${profile.flags.cacheTypeV})` : "unknown"],
+      ...(est.note ? [["Note", pc.yellow(est.note)]] : []),
+    ]));
+  } catch {
+    return renderSection("Memory", pc.dim("Estimate unavailable for this model."));
+  }
+}
+
+function samplingSummary(flags) {
+  return `temp ${flags.temperature}, top-p ${flags.topP}, top-k ${flags.topK}`;
+}