offgrid-ai 0.3.16 → 0.3.17

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "offgrid-ai",
-  "version": "0.3.16",
+  "version": "0.3.17",
   "description": "Privacy-first CLI for running local LLMs — discover, configure, run, benchmark",
   "author": "Eeshan Srivastava (https://eeshans.com)",
   "type": "module",

package/src/autodetect.mjs

@@ -7,20 +7,26 @@ import { readGgufMetadata } from "./gguf.mjs";
 export function detectCapabilities(modelPath, mmprojPath) {
   const meta = safeReadGgufMetadata(modelPath);
   const name = basename(modelPath).toLowerCase();
+  const pathHints = String(modelPath).toLowerCase();
 
   // Architecture
   const architecture = meta["general.architecture"] ?? null;
 
   // Thinking / reasoning mode
   const hasThinkingKwargs = meta["chat_template_kwargs"] !== undefined;
-  const nameHintsThinking = /qwen3|gemma-4|gemma4|deepseek-r[12]/i.test(name);
+  const nameHintsThinking = /qwen3|qwen3\.\d|gemma-4|gemma4|deepseek-r[12]/i.test(pathHints);
   const thinking = hasThinkingKwargs || nameHintsThinking;
 
+  // Quantization-aware / imatrix quantization hints. These mostly affect
+  // display and defaults transparency; llama-server does not need a QAT flag.
+  const qat = /qat|imatrix|i-?matrix/i.test(pathHints) || Object.keys(meta).some((key) => key.startsWith("quantize.imatrix."));
+
   // Vision — mmproj present
   const vision = Boolean(mmprojPath && existsSync(mmprojPath));
 
-  // MTP (multi-token prediction) — detect speculative decoding
-  const mtp = /mtp/i.test(name) || architecture === "qwen3";
+  // MTP (multi-token prediction) — detect speculative decoding.
+  // Do not treat all Qwen models as MTP; require an explicit filename or metadata hint.
+  const mtp = /\bmtp\b|draft-mtp|multi-token/i.test(pathHints) || Object.keys(meta).some((key) => /mtp|draft|speculative/i.test(key));
 
   // Quantization
   const quant = name.match(/(Q\d_K_[A-Z]+|UD-[A-Z0-9_]+)/i)?.[1] ?? null;
@@ -31,7 +37,7 @@ export function detectCapabilities(modelPath, mmprojPath) {
     : undefined;
   const ctxSize = metaCtx ?? (thinking ? 80000 : 32768);
 
-  return { architecture, thinking, vision, mtp, quant, metaCtx, ctxSize, meta };
+  return { architecture, thinking, vision, mtp, qat, quant, metaCtx, ctxSize, meta };
 }
 
 // ── Compute llama-server flags from capabilities ───────────────────────────
@@ -42,7 +48,7 @@ export function computeFlags(capabilities, modelPath, mmprojPath, draftModelPath
 
   const flags = {
     host: "127.0.0.1",
-    port: 8080,
+    port: mtp ? 8081 : 8080,
     ctxSize: capabilities.ctxSize,
     flashAttention: "on",
     cacheTypeK: isLowMem ? "f16" : "bf16",

package/src/cli.mjs

@@ -14,6 +14,7 @@ import { checkForUpdate, currentPackageVersion, detectInvocation, updateCommand,
 import { removeInstallerPathEntries } from "./shell-path.mjs";
 import { configureLocalProfile } from "./profile-setup.mjs";
 import { buildPrettyCommand } from "./command.mjs";
+import { detectCapabilities } from "./autodetect.mjs";
 
 // ── Entry point ────────────────────────────────────────────────────────────
 
@@ -176,10 +177,10 @@ async function modelsCommand(argv) {
 
 async function modelCommandCenter(catalog) {
   const normalized = normalizeCatalog(catalog);
-  await printModelCatalog(normalized);
+  const items = modelCatalogItems(normalized);
+  await printModelCatalog(normalized, items);
   if (!process.stdin.isTTY) return;
 
-  const items = modelCatalogItems(normalized);
   if (items.length === 0) return;
 
   const prompt = createPrompt();
@@ -234,36 +235,54 @@ function normalizeCatalog(catalog) {
   return { profiles, ggufModels, managedModels, newModels, managedItems };
 }
 
-async function printModelCatalog({ profiles, newModels, managedModels }) {
-  if (profiles.length > 0) {
-    console.log(pc.bold("\nSaved profiles"));
+async function printModelCatalog({ profiles, newModels, managedItems }, items = modelCatalogItems({ profiles, newModels, managedItems })) {
+  const itemNumber = (predicate) => {
+    const index = items.findIndex(predicate);
+    return index === -1 ? "  " : String(index + 1).padStart(2, " ");
+  };
+
+  console.log(pc.bold("\nSaved profiles"));
+  if (profiles.length === 0) {
+    console.log(pc.dim("  None yet."));
+  } else {
     for (const profile of profiles) {
       const backend = backendFor(profile.backend);
       const colorMap = { "llama-cpp": pc.yellow, "llama-cpp-mtp": pc.blue, "ollama": pc.magenta, "omlx": pc.cyan };
       const running = await isProfileRunning(profile);
       const piConfigured = await hasPiModel(profile);
       const c = colorMap[profile.backend] ?? pc.magenta;
-      console.log(`  ${running ? pc.green("●") : pc.dim("○")} ${pc.bold(profile.label)} ${c(`[${backend.label}]`)} · ${pc.cyan(profile.modelAlias)} ${piConfigured ? pc.green("· Pi synced") : pc.yellow("· Pi not synced")}`);
+      const num = itemNumber((item) => item.type === "profile" && item.profile.id === profile.id);
+      console.log(`${num}. ${running ? pc.green("●") : pc.dim("○")} ${pc.bold(profile.label)} ${c(`[${backend.label}]`)} · ${pc.cyan(profile.modelAlias)} ${piConfigured ? pc.green("· Pi synced") : pc.yellow("· Pi not synced")}`);
     }
-  } else {
-    console.log(pc.bold("\nSaved profiles"));
-    console.log(pc.dim("  None yet."));
   }
 
-  if (newModels.length > 0) {
-    console.log(pc.bold("\nNew GGUF models"));
+  console.log("");
+  console.log(pc.bold("Downloaded models not set up yet"));
+  if (newModels.length === 0) {
+    console.log(pc.dim("  None. Every downloaded GGUF has a profile."));
+  } else {
     for (const model of newModels.slice(0, 20)) {
-      console.log(`  ${pc.cyan(model.label)} ${pc.dim(model.quant ?? "")} · ${pc.dim(formatBytes(model.sizeBytes))}`);
+      const caps = detectCapabilities(model.path, model.mmprojPath);
+      const num = itemNumber((item) => item.type === "new" && item.model.path === model.path);
+      console.log(`${num}. ${pc.cyan(model.label)} ${capabilityBadges(caps)} ${pc.dim(model.quant ?? "")}`);
+      console.log(`    alias:  ${pc.cyan(model.aliasSuggestion)}`);
+      console.log(`    size:   ${formatBytes(model.sizeBytes)}`);
     }
     if (newModels.length > 20) console.log(pc.dim(`  ... and ${newModels.length - 20} more`));
   }
 
-  for (const { backendId, models } of managedModels) {
-    if (models.length === 0) continue;
+  for (const backendId of ["ollama", "omlx"]) {
+    const backendItems = managedItems.filter((item) => item.backendId === backendId);
+    if (backendItems.length === 0) continue;
     const be = BACKENDS[backendId];
-    console.log(pc.bold(`\n${be.label} models`));
-    for (const model of models.slice(0, 10)) console.log(`  ${pc.cyan(model.label)}`);
-    if (models.length > 10) console.log(pc.dim(`  ... and ${models.length - 10} more`));
+    console.log("");
+    console.log(pc.bold(`${be.label} models`));
+    for (const { model } of backendItems.slice(0, 10)) {
+      const num = itemNumber((item) => item.type === "managed" && item.backendId === backendId && item.model.id === model.id);
+      console.log(`${num}. ${pc.cyan(model.label)} ${pc.dim(model.quant ?? "")}`);
+      console.log(`    id: ${pc.cyan(model.id)}`);
+    }
+    if (backendItems.length > 10) console.log(pc.dim(`  ... and ${backendItems.length - 10} more`));
   }
 }
 
@@ -276,17 +295,25 @@ function modelCatalogItems({ profiles, newModels, managedItems }) {
 }
 
 async function chooseCatalogItem(prompt, items, action) {
-  const allowed = action === "remove" ? items.filter((item) => item.type === "profile") : items;
-  if (allowed.length === 0) {
-    console.log(pc.yellow(action === "remove" ? "No saved profiles to remove." : "No models available."));
+  if (action === "remove" && !items.some((item) => item.type === "profile")) {
+    console.log(pc.yellow("No saved profiles to remove."));
     return null;
   }
-  const selected = await prompt.choice("Select", allowed.map((item, index) => ({
-    value: String(index),
-    label: item.label,
-    hint: item.hint,
-  })), "0");
-  return allowed[Number(selected)];
+
+  const input = await prompt.text("Select a number", "");
+  if (!input) return null;
+  const index = Number(input) - 1;
+  if (!Number.isInteger(index) || index < 0 || index >= items.length) {
+    console.log(pc.yellow(`No item ${input}.`));
+    return null;
+  }
+
+  const item = items[index];
+  if (action === "remove" && item.type !== "profile") {
+    console.log(pc.yellow("Only saved profiles can be removed."));
+    return null;
+  }
+  return item;
 }
 
 async function handleCatalogAction(prompt, action, item) {
@@ -337,6 +364,7 @@ async function printProfileDetails(profile) {
     ["ID", pc.cyan(profile.id)],
     ["Label", pc.bold(profile.label)],
     ["Backend", backend.label],
+    ...(profile.capabilities ? [["Detected", capabilitySummary(profile.capabilities)]] : []),
     ["Endpoint", pc.green(profile.baseUrl)],
     ...(!isManaged ? [
       ["Model", profile.modelPath ?? "unknown"],
@@ -354,8 +382,10 @@ async function printProfileDetails(profile) {
 }
 
 function printGgufModelDetails(model) {
+  const caps = detectCapabilities(model.path, model.mmprojPath);
   console.log("\n" + renderSection("GGUF model", renderRows([
     ["Label", pc.bold(model.label)],
+    ["Detected", capabilitySummary(caps)],
     ["Model", model.path],
     ["MMProj", model.mmprojPath ?? "none"],
     ["Quant", model.quant ?? "unknown"],
@@ -372,6 +402,26 @@ function printManagedModelDetails(model, backend) {
   ])));
 }
 
+function capabilitySummary(caps) {
+  const parts = [];
+  if (caps.architecture) parts.push(caps.architecture);
+  if (caps.quant) parts.push(caps.quant);
+  if (caps.mtp) parts.push("MTP");
+  if (caps.qat) parts.push("QAT/imatrix");
+  if (caps.thinking) parts.push("thinking");
+  if (caps.vision) parts.push("vision");
+  return parts.length > 0 ? parts.join(" · ") : "standard GGUF";
+}
+
+function capabilityBadges(caps) {
+  const badges = [];
+  if (caps.mtp) badges.push(pc.blue("[MTP]"));
+  if (caps.qat) badges.push(pc.green("[QAT]"));
+  if (caps.thinking) badges.push(pc.magenta("[thinking]"));
+  if (caps.vision) badges.push(pc.cyan("[vision]"));
+  return badges.join(" ");
+}
+
 function createManagedProfile(model, backendId) {
   return normalizeProfile({
     id: model.id.replace(/[^a-z0-9._-]+/gi, "-").toLowerCase(),

package/src/profile-setup.mjs

@@ -8,10 +8,27 @@ const CACHE_CHOICES = [
   { value: "q4_0", label: "q4_0", hint: "lowest memory, quality/speed tradeoff" },
 ];
 
+const GENERAL_DEFAULTS = {
+  topK: 20,
+  presencePenalty: 1.5,
+  repeatPenalty: 1.0,
+};
+
+const THINKING_DEFAULTS = {
+  topK: 64,
+  presencePenalty: 0,
+  repeatPenalty: 1.1,
+  chatTemplateKwargs: { enable_thinking: true },
+};
+
 export async function configureLocalProfile(prompt, profile) {
+  let configured = profile;
+  const caps = profile.capabilities ?? {};
+
   console.log("");
   console.log(renderSection("Model setup", renderRows([
     ["Model", pc.bold(profile.label)],
+    ["Detected", detectionSummary(caps)],
     ["Context", `${profile.flags.ctxSize.toLocaleString()} tokens`],
     ["KV cache", `${profile.flags.cacheTypeK}/${profile.flags.cacheTypeV}`],
     ["Sampling", samplingSummary(profile.flags)],
@@ -19,13 +36,36 @@ export async function configureLocalProfile(prompt, profile) {
   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 });
+  if (caps.mtp) {
+    console.log(renderSection("Detected MTP", renderRows([
+      ["Backend", "llama.cpp MTP"],
+      ["Port", "8081"],
+      ["Flags", "--spec-type draft-mtp --spec-draft-n-max 2"],
+    ])));
+    const useMtp = await prompt.yesNo("Use MTP speculative decoding flags?", true);
+    configured = useMtp ? applyMtpDefaults(configured) : removeMtpDefaults(configured);
+  }
+
+  if (caps.thinking || caps.qat) {
+    console.log("");
+    console.log(renderSection(caps.qat ? "Detected QAT / imatrix-style model" : "Detected thinking model", renderRows([
+      ["Defaults", "thinking / loop-safe"],
+      ["Flags", "--top-k 64 --presence-penalty 0 --repeat-penalty 1.1"],
+      ["Template", "--chat-template-kwargs { enable_thinking: true }"],
+    ])));
+    const useThinking = await prompt.yesNo("Use these thinking/QAT-safe defaults?", true);
+    configured = useThinking ? applyThinkingDefaults(configured) : removeThinkingDefaults(configured);
+  }
+
+  const ctxSize = await prompt.number("Context window tokens", configured.flags.ctxSize, 1024, 1048576);
+  const cacheTypeK = await prompt.choice("K cache precision", CACHE_CHOICES, configured.flags.cacheTypeK);
+  const cacheTypeV = await prompt.choice("V cache precision", CACHE_CHOICES, configured.flags.cacheTypeV);
+  configured = applyRuntimeFlagOverrides(configured, { ctxSize, cacheTypeK, cacheTypeV });
 
   console.log("");
   console.log(renderSection("Defaults", renderRows([
+    ["Backend", configured.backend],
+    ["Endpoint", configured.baseUrl],
     ["Temperature", configured.flags.temperature],
     ["Top-p", configured.flags.topP],
     ["Top-k", configured.flags.topK],
@@ -41,21 +81,63 @@ export async function configureLocalProfile(prompt, profile) {
 
 export function applyRuntimeFlagOverrides(profile, overrides) {
   const flags = { ...profile.flags, ...overrides };
-  return {
+  return applyProfileFlags(profile, flags);
+}
+
+function applyMtpDefaults(profile) {
+  const flags = { ...profile.flags, port: 8081 };
+  return applyProfileFlags({ ...profile, backend: "llama-cpp-mtp", providerId: "llama-cpp-mtp" }, flags, {
+    values: { "--spec-type": "draft-mtp", "--spec-draft-n-max": 2 },
+  });
+}
+
+function removeMtpDefaults(profile) {
+  const flags = { ...profile.flags, port: 8080 };
+  return applyProfileFlags({ ...profile, backend: "llama-cpp", providerId: "llama-cpp" }, flags, {
+    remove: ["--spec-type", "--spec-draft-n-max"],
+  });
+}
+
+function applyThinkingDefaults(profile) {
+  const flags = { ...profile.flags, ...THINKING_DEFAULTS };
+  return applyProfileFlags(profile, flags);
+}
+
+function removeThinkingDefaults(profile) {
+  const flags = { ...profile.flags, ...GENERAL_DEFAULTS };
+  delete flags.chatTemplateKwargs;
+  return applyProfileFlags(profile, flags, { remove: ["--chat-template-kwargs"] });
+}
+
+function applyProfileFlags(profile, flags, edits = {}) {
+  const next = {
     ...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,
-    }),
+    harnesses: {
+      ...(profile.harnesses ?? {}),
+      pi: { ...(profile.harnesses?.pi ?? {}), enabled: true, model: `${profile.providerId ?? profile.backend}/${profile.modelAlias ?? profile.id}` },
+    },
   };
+  next.commandArgv = updateArgv(profile.commandArgv ?? [], {
+    "--host": flags.host,
+    "--port": flags.port,
+    "--ctx-size": flags.ctxSize,
+    "--cache-type-k": flags.cacheTypeK,
+    "--cache-type-v": flags.cacheTypeV,
+    "--top-k": flags.topK,
+    "--presence-penalty": flags.presencePenalty,
+    "--repeat-penalty": flags.repeatPenalty,
+    ...(flags.chatTemplateKwargs ? { "--chat-template-kwargs": JSON.stringify(flags.chatTemplateKwargs) } : {}),
+  }, edits);
+  return next;
 }
 
-function updateArgv(argv, values) {
-  const next = [...argv];
-  for (const [flag, value] of Object.entries(values)) {
+function updateArgv(argv, values, edits = {}) {
+  let next = [...argv];
+  for (const flag of edits.remove ?? []) next = removeOption(next, flag);
+  for (const [flag, value] of Object.entries({ ...values, ...(edits.values ?? {}) })) {
+    if (value === undefined) continue;
     const index = next.indexOf(flag);
     if (index === -1) next.push(flag, String(value));
     else next[index + 1] = String(value);
@@ -63,6 +145,18 @@ function updateArgv(argv, values) {
   return next;
 }
 
+function removeOption(argv, flag) {
+  const next = [];
+  for (let i = 0; i < argv.length; i++) {
+    if (argv[i] === flag) {
+      if (argv[i + 1] && !argv[i + 1].startsWith("--")) i += 1;
+      continue;
+    }
+    next.push(argv[i]);
+  }
+  return next;
+}
+
 function renderMemoryEstimate(profile) {
   try {
     const est = estimateMemory(profile.modelPath, profile.mmprojPath, null, profile.flags);
@@ -77,6 +171,17 @@ function renderMemoryEstimate(profile) {
   }
 }
 
+function detectionSummary(caps) {
+  const parts = [];
+  if (caps.architecture) parts.push(caps.architecture);
+  if (caps.quant) parts.push(caps.quant);
+  if (caps.mtp) parts.push("MTP");
+  if (caps.qat) parts.push("QAT/imatrix");
+  if (caps.thinking) parts.push("thinking");
+  if (caps.vision) parts.push("vision");
+  return parts.length > 0 ? parts.join(" · ") : "standard GGUF";
+}
+
 function samplingSummary(flags) {
   return `temp ${flags.temperature}, top-p ${flags.topP}, top-k ${flags.topK}`;
 }

package/src/profiles.mjs

@@ -136,25 +136,41 @@ export function normalizeProfile(profile) {
 
 // ── Auto-create profile from a discovered model ────────────────────────────
 
-export async function createProfileFromModel(model, backendId = "llama-cpp") {
+export async function createProfileFromModel(model, backendId) {
   const { detectCapabilities } = await import("./autodetect.mjs");
   const caps = detectCapabilities(model.path, model.mmprojPath);
+  const backend = backendId ?? (caps.mtp ? "llama-cpp-mtp" : "llama-cpp");
   const id = slugFromLabel(model.label);
   const { flags, argv } = computeFlags(caps, model.path, model.mmprojPath, null);
 
   return normalizeProfile({
     id,
     label: model.label,
-    backend: backendId,
+    backend,
+    providerId: backend,
     modelAlias: model.aliasSuggestion,
     modelPath: model.path,
     mmprojPath: model.mmprojPath,
+    capabilities: summarizeCapabilities(caps),
     preset: null, // no presets — auto-detected
     flags,
     commandArgv: argv,
   });
 }
 
+function summarizeCapabilities(caps) {
+  return {
+    architecture: caps.architecture,
+    thinking: caps.thinking,
+    vision: caps.vision,
+    mtp: caps.mtp,
+    qat: caps.qat,
+    quant: caps.quant,
+    metaCtx: caps.metaCtx,
+    ctxSize: caps.ctxSize,
+  };
+}
+
 // ── State files (for running servers) ──────────────────────────────────────
 
 export async function readState(id) {