omnius 1.0.107 → 1.0.108

package/dist/index.js

@@ -578020,6 +578020,23 @@ async function needsTextToolMode(modelName, backendUrl2) {
   const hasTools = await checkToolSupport(modelName, backendUrl2);
   return !hasTools;
 }
+function detectUnifiedMemory() {
+  if (process.platform === "darwin" && process.arch === "arm64") return true;
+  if (process.platform === "linux") {
+    try {
+      if (existsSync91("/etc/nv_tegra_release")) return true;
+    } catch {
+    }
+    try {
+      if (existsSync91("/sys/devices/soc0/family")) {
+        const family = readFileSync74("/sys/devices/soc0/family", "utf8").trim().toLowerCase();
+        if (family.includes("tegra")) return true;
+      }
+    } catch {
+    }
+  }
+  return false;
+}
 function parseNvidiaSmi(stdout) {
   let total = 0;
   let free = 0;
@@ -578110,12 +578127,18 @@ function detectSystemSpecs() {
     } catch {
     }
   }
+  const unifiedMemory = detectUnifiedMemory();
+  if (unifiedMemory && totalRamGB > 0) {
+    const floorGB = totalRamGB * 0.8;
+    if (availableRamGB < floorGB) availableRamGB = floorGB;
+  }
   return {
     totalRamGB: Math.round(totalRamGB * 10) / 10,
     availableRamGB: Math.round(availableRamGB * 10) / 10,
     gpuVramGB: Math.round(gpuVramGB * 10) / 10,
     availableVramGB: Math.round(availableVramGB * 10) / 10,
-    gpuName
+    gpuName,
+    unifiedMemory
   };
 }
 async function detectSystemSpecsAsync() {
@@ -578168,12 +578191,18 @@ async function detectSystemSpecsAsync() {
     } catch {
     }
   }
+  const unifiedMemory = detectUnifiedMemory();
+  if (unifiedMemory && totalRamGB > 0) {
+    const floorGB = totalRamGB * 0.8;
+    if (availableRamGB < floorGB) availableRamGB = floorGB;
+  }
   return {
     totalRamGB: Math.round(totalRamGB * 10) / 10,
     availableRamGB: Math.round(availableRamGB * 10) / 10,
     gpuVramGB: Math.round(gpuVramGB * 10) / 10,
     availableVramGB: Math.round(availableVramGB * 10) / 10,
-    gpuName
+    gpuName,
+    unifiedMemory
   };
 }
 function recommendModel(specs) {
@@ -578198,7 +578227,16 @@ function calculateMemoryBoundedNumCtx(specs, modelSizeGB2, kvBytesPerToken, arch
     const maxTokens = Math.floor(usableGB * 1024 ** 3 / kvBytesPerToken);
     numCtx = Math.max(2048, Math.floor(maxTokens / 1024) * 1024);
   } else {
-    const kvEstimate = modelSizeGB2 <= 5 ? 524288 : modelSizeGB2 <= 20 ? 1048576 : 1572864;
+    const kvEstimate = modelSizeGB2 <= 5 ? 64 * 1024 : (
+      // 4B-class
+      modelSizeGB2 <= 12 ? 160 * 1024 : (
+        // 9B-class
+        modelSizeGB2 <= 25 ? 256 * 1024 : (
+          // 30B-class
+          384 * 1024
+        )
+      )
+    );
     const maxTokens = Math.floor(usableGB * 1024 ** 3 / kvEstimate);
     numCtx = Math.max(2048, Math.floor(maxTokens / 1024) * 1024);
   }
@@ -579999,6 +580037,17 @@ function modelSizeGB(models, modelName) {
   const known = QWEN_VARIANTS.find((v) => modelName.includes(v.tag.split(":")[1] ?? ""));
   return known?.sizeGB ?? 4;
 }
+function defaultLayersForArch(arch3) {
+  const a2 = arch3.toLowerCase();
+  if (a2.includes("qwen")) return 36;
+  if (a2.includes("llama")) return 32;
+  if (a2.includes("mistral")) return 32;
+  if (a2.includes("phi")) return 32;
+  if (a2.includes("gemma")) return 42;
+  if (a2.includes("granite")) return 40;
+  if (a2.includes("command")) return 40;
+  return 32;
+}
 async function queryModelKVInfo(backendUrl2, modelName) {
   try {
     const normalized = backendUrl2.replace(/\/+$/, "");
@@ -580014,13 +580063,17 @@ async function queryModelKVInfo(backendUrl2, modelName) {
     const info = data.model_info;
     const arch3 = info["general.architecture"];
     if (!arch3) return null;
-    const nLayers = info[`${arch3}.block_count`];
-    const nKVHeads = info[`${arch3}.attention.head_count_kv`] ?? info[`${arch3}.attention.head_count`];
-    const keyDim = info[`${arch3}.attention.key_length`];
-    const valDim = info[`${arch3}.attention.value_length`] ?? keyDim;
+    const nLayersRaw = info[`${arch3}.block_count`];
+    const nKVHeadsRaw = info[`${arch3}.attention.head_count_kv`] ?? info[`${arch3}.attention.head_count`];
+    const keyDimRaw = info[`${arch3}.attention.key_length`];
+    const valDimRaw = info[`${arch3}.attention.value_length`] ?? keyDimRaw;
     const archMax = info[`${arch3}.context_length`];
     if (!archMax) return null;
-    if (!nLayers || !nKVHeads || !keyDim || !valDim) return { archMax };
+    const keyDim = keyDimRaw ?? 128;
+    const valDim = valDimRaw ?? 128;
+    const nLayers = nLayersRaw ?? defaultLayersForArch(arch3);
+    const nKVHeads = nKVHeadsRaw ?? 32;
+    if (!nLayers) return { archMax };
     const kvBytesPerToken = nLayers * nKVHeads * (keyDim + valDim) * 2;
     return { kvBytesPerToken, archMax };
   } catch {

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.107",
+  "version": "1.0.108",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.107",
+      "version": "1.0.108",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.107",
+  "version": "1.0.108",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",