omnius 1.0.105 → 1.0.107

package/dist/index.js

@@ -578020,10 +578020,51 @@ async function needsTextToolMode(modelName, backendUrl2) {
   const hasTools = await checkToolSupport(modelName, backendUrl2);
   return !hasTools;
 }
+function parseNvidiaSmi(stdout) {
+  let total = 0;
+  let free = 0;
+  let name10 = "";
+  const lines = stdout.trim().split("\n").filter(Boolean);
+  for (const line of lines) {
+    const parts = line.split(",").map((s2) => s2.trim());
+    const totMB = parseInt(parts[0] ?? "0", 10);
+    const freeMB = parseInt(parts[1] ?? "0", 10);
+    if (!isNaN(totMB)) total += totMB / 1024;
+    if (!isNaN(freeMB)) free += freeMB / 1024;
+    if (!name10 && parts[2]) name10 = parts[2];
+  }
+  return { total, free, name: name10 };
+}
+function parseFreeBytes(stdout) {
+  const memLine = stdout.split("\n").find((l2) => /^Mem:/i.test(l2)) ?? "";
+  const nums = memLine.match(/\d+/g);
+  if (!nums || nums.length === 0) return { total: 0, available: 0 };
+  const total = parseInt(nums[0], 10) || 0;
+  const available = nums.length >= 6 ? parseInt(nums[5], 10) || 0 : 0;
+  return { total, available };
+}
+function parseRocmSmi(stdout) {
+  let total = 0;
+  let used = 0;
+  let name10 = "";
+  const lines = stdout.trim().split("\n").filter(Boolean);
+  for (const line of lines) {
+    if (line.toLowerCase().startsWith("device") || line.startsWith("=")) continue;
+    const parts = line.split(",").map((s2) => s2.trim());
+    if (parts.length < 3) continue;
+    const usedBytes = parseInt(parts[1] ?? "0", 10);
+    const totBytes = parseInt(parts[2] ?? "0", 10);
+    if (!isNaN(usedBytes)) used += usedBytes / 1024 ** 3;
+    if (!isNaN(totBytes)) total += totBytes / 1024 ** 3;
+    if (!name10 && parts[0]) name10 = parts[0];
+  }
+  return { total, free: Math.max(0, total - used), name: name10 ? `AMD ${name10}` : "AMD GPU" };
+}
 function detectSystemSpecs() {
   let totalRamGB = 0;
   let availableRamGB = 0;
   let gpuVramGB = 0;
+  let availableVramGB = 0;
   let gpuName = "";
   try {
     const memInfo = execSync50("free -b 2>/dev/null || sysctl -n hw.memsize 2>/dev/null", {
@@ -578031,14 +578072,14 @@ function detectSystemSpecs() {
       timeout: 5e3
     });
     if (memInfo.includes("Mem:")) {
-      const match = memInfo.match(/^Mem:\s+(\d+)\s+\d+\s+\d+\s+\d+\s+\d+\s+(\d+)/m);
-      if (match) {
-        totalRamGB = parseInt(match[1], 10) / 1024 ** 3;
-        availableRamGB = parseInt(match[2], 10) / 1024 ** 3;
+      const { total, available } = parseFreeBytes(memInfo);
+      if (total > 0) {
+        totalRamGB = total / 1024 ** 3;
+        availableRamGB = available > 0 ? available / 1024 ** 3 : totalRamGB * 0.75;
       }
     } else {
       const bytes = parseInt(memInfo.trim(), 10);
-      if (!isNaN(bytes)) {
+      if (!isNaN(bytes) && bytes > 0) {
         totalRamGB = bytes / 1024 ** 3;
         availableRamGB = totalRamGB * 0.7;
       }
@@ -578047,24 +578088,33 @@ function detectSystemSpecs() {
   }
   try {
     const nvidiaSmi = execSync50(
-      "nvidia-smi --query-gpu=memory.total,name --format=csv,noheader,nounits 2>/dev/null",
+      "nvidia-smi --query-gpu=memory.total,memory.free,name --format=csv,noheader,nounits 2>/dev/null",
       { encoding: "utf8", timeout: 5e3 }
     );
-    const lines = nvidiaSmi.trim().split("\n");
-    if (lines.length > 0) {
-      for (const line of lines) {
-        const parts = line.split(",").map((s2) => s2.trim());
-        const vramMB = parseInt(parts[0] ?? "0", 10);
-        if (!isNaN(vramMB)) gpuVramGB += vramMB / 1024;
-        if (!gpuName && parts[1]) gpuName = parts[1];
-      }
-    }
+    const r2 = parseNvidiaSmi(nvidiaSmi);
+    gpuVramGB += r2.total;
+    availableVramGB += r2.free;
+    if (!gpuName && r2.name) gpuName = r2.name;
   } catch {
   }
+  if (gpuVramGB === 0) {
+    try {
+      const rocmSmi = execSync50(
+        "rocm-smi --showmeminfo vram --csv 2>/dev/null",
+        { encoding: "utf8", timeout: 5e3 }
+      );
+      const r2 = parseRocmSmi(rocmSmi);
+      gpuVramGB += r2.total;
+      availableVramGB += r2.free;
+      if (!gpuName && r2.name) gpuName = r2.name;
+    } catch {
+    }
+  }
   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
   };
 }
@@ -578072,6 +578122,7 @@ async function detectSystemSpecsAsync() {
   let totalRamGB = 0;
   let availableRamGB = 0;
   let gpuVramGB = 0;
+  let availableVramGB = 0;
   let gpuName = "";
   try {
     const { stdout: memInfo } = await execAsync2(
@@ -578079,14 +578130,14 @@ async function detectSystemSpecsAsync() {
       { timeout: 5e3 }
     );
     if (memInfo.includes("Mem:")) {
-      const match = memInfo.match(/^Mem:\s+(\d+)\s+\d+\s+\d+\s+\d+\s+\d+\s+(\d+)/m);
-      if (match) {
-        totalRamGB = parseInt(match[1], 10) / 1024 ** 3;
-        availableRamGB = parseInt(match[2], 10) / 1024 ** 3;
+      const { total, available } = parseFreeBytes(memInfo);
+      if (total > 0) {
+        totalRamGB = total / 1024 ** 3;
+        availableRamGB = available > 0 ? available / 1024 ** 3 : totalRamGB * 0.75;
       }
     } else {
       const bytes = parseInt(memInfo.trim(), 10);
-      if (!isNaN(bytes)) {
+      if (!isNaN(bytes) && bytes > 0) {
         totalRamGB = bytes / 1024 ** 3;
         availableRamGB = totalRamGB * 0.7;
       }
@@ -578095,24 +578146,33 @@ async function detectSystemSpecsAsync() {
   }
   try {
     const { stdout: nvidiaSmi } = await execAsync2(
-      "nvidia-smi --query-gpu=memory.total,name --format=csv,noheader,nounits 2>/dev/null",
+      "nvidia-smi --query-gpu=memory.total,memory.free,name --format=csv,noheader,nounits 2>/dev/null",
       { timeout: 5e3 }
     );
-    const lines = nvidiaSmi.trim().split("\n");
-    if (lines.length > 0) {
-      for (const line of lines) {
-        const parts = line.split(",").map((s2) => s2.trim());
-        const vramMB = parseInt(parts[0] ?? "0", 10);
-        if (!isNaN(vramMB)) gpuVramGB += vramMB / 1024;
-        if (!gpuName && parts[1]) gpuName = parts[1];
-      }
-    }
+    const r2 = parseNvidiaSmi(nvidiaSmi);
+    gpuVramGB += r2.total;
+    availableVramGB += r2.free;
+    if (!gpuName && r2.name) gpuName = r2.name;
   } catch {
   }
+  if (gpuVramGB === 0) {
+    try {
+      const { stdout: rocmSmi } = await execAsync2(
+        "rocm-smi --showmeminfo vram --csv 2>/dev/null",
+        { timeout: 5e3 }
+      );
+      const r2 = parseRocmSmi(rocmSmi);
+      gpuVramGB += r2.total;
+      availableVramGB += r2.free;
+      if (!gpuName && r2.name) gpuName = r2.name;
+    } catch {
+    }
+  }
   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
   };
 }
@@ -578127,9 +578187,10 @@ function recommendModel(specs) {
   }
   return QWEN_VARIANTS.find((v) => v.tag === "qwen3.5:cloud");
 }
-function calculateContextWindow(specs, modelSizeGB2, kvBytesPerToken, archMax) {
+function calculateMemoryBoundedNumCtx(specs, modelSizeGB2, kvBytesPerToken, archMax) {
   const ramBudget = specs.availableRamGB > 0 ? specs.availableRamGB : specs.totalRamGB;
-  const totalAvail = Math.max(specs.gpuVramGB, ramBudget);
+  const vramBudget = specs.availableVramGB > 0 ? specs.availableVramGB : specs.gpuVramGB;
+  const totalAvail = Math.max(vramBudget, ramBudget);
   const remaining = Math.max(0, totalAvail - modelSizeGB2);
   const usableGB = remaining * 0.85;
   let numCtx;
@@ -578141,7 +578202,6 @@ function calculateContextWindow(specs, modelSizeGB2, kvBytesPerToken, archMax) {
     const maxTokens = Math.floor(usableGB * 1024 ** 3 / kvEstimate);
     numCtx = Math.max(2048, Math.floor(maxTokens / 1024) * 1024);
   }
-  numCtx = Math.min(numCtx, 131072);
   if (archMax && archMax > 0) numCtx = Math.min(numCtx, archMax);
   if (kvBytesPerToken && kvBytesPerToken > 0 && modelSizeGB2 > 0) {
     const maxKVBytes = modelSizeGB2 * 4 * 1024 ** 3;
@@ -578149,6 +578209,11 @@ function calculateContextWindow(specs, modelSizeGB2, kvBytesPerToken, archMax) {
     const budgetCap = Math.max(2048, Math.floor(maxTokensFromBudget / 1024) * 1024);
     numCtx = Math.min(numCtx, budgetCap);
   }
+  return numCtx;
+}
+function calculateContextWindow(specs, modelSizeGB2, kvBytesPerToken, archMax) {
+  let numCtx = calculateMemoryBoundedNumCtx(specs, modelSizeGB2, kvBytesPerToken, archMax);
+  numCtx = Math.min(numCtx, 131072);
   const label = numCtx >= 1024 ? `${Math.floor(numCtx / 1024)}K` : String(numCtx);
   return { numCtx, label };
 }
@@ -578156,18 +578221,17 @@ function formatContextLabel(numCtx) {
   return numCtx >= 1024 ? `${Math.floor(numCtx / 1024)}K` : String(numCtx);
 }
 function calculateExpandedVariantContextWindow(specs, modelSizeGB2, kvBytesPerToken, archMax) {
-  const memoryBudget = calculateContextWindow(
+  const memoryFit = calculateMemoryBoundedNumCtx(
     specs,
     modelSizeGB2,
     kvBytesPerToken,
     archMax
   );
-  if (archMax && archMax > 0) {
-    const archCtx = Math.max(2048, Math.floor(archMax / 1024) * 1024);
-    const numCtx = Math.min(archCtx, memoryBudget.numCtx);
-    return { numCtx, label: formatContextLabel(numCtx) };
-  }
-  return memoryBudget;
+  const archCtx = archMax && archMax > 0 ? Math.max(2048, Math.floor(archMax / 1024) * 1024) : Number.POSITIVE_INFINITY;
+  const floor = Math.min(EXPANDED_VARIANT_MIN_NUM_CTX, archCtx);
+  const fits = Math.min(memoryFit, archCtx);
+  const numCtx = Math.max(floor, fits);
+  return { numCtx, label: formatContextLabel(numCtx) };
 }
 function ask(rl, question) {
   return new Promise((resolve52) => {
@@ -580359,7 +580423,7 @@ export PATH="${binDir}:$PATH"  # Added by omnius for nvim
   } catch {
   }
 }
-var execAsync2, OMNIUS_FIRST_RUN_BANNER, ANSI_RE, visibleLen2, QWEN_VARIANTS, _toolSupportCache, _cloudflaredInstallPromise;
+var execAsync2, OMNIUS_FIRST_RUN_BANNER, ANSI_RE, visibleLen2, QWEN_VARIANTS, _toolSupportCache, EXPANDED_VARIANT_MIN_NUM_CTX, _cloudflaredInstallPromise;
 var init_setup = __esm({
   "packages/cli/src/tui/setup.ts"() {
     "use strict";
@@ -580392,6 +580456,7 @@ var init_setup = __esm({
       { tag: "qwen3.5:397b-cloud", sizeGB: 0, label: "397B Cloud (Ollama Cloud)", cloud: true }
     ];
     _toolSupportCache = /* @__PURE__ */ new Map();
+    EXPANDED_VARIANT_MIN_NUM_CTX = 32768;
     _cloudflaredInstallPromise = null;
   }
 });

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.105",
+  "version": "1.0.107",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.105",
+      "version": "1.0.107",
       "bundleDependencies": [
         "image-to-ascii"
       ],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.105",
+  "version": "1.0.107",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",