GPU Lab app; CPU, NPU, GPU, RAM and ROM wording

README.md

@@ -27,7 +27,7 @@ tags:
 
 > **"For 50 years, the operating system lived inside the computer. Now the operating system lives inside the AI."**
 
-ViKernel wires **ROM, RAM, and an xPU chip circuit** directly into the weights of an LLM, then runs our own **Secure OS** on top.
+ViKernel wires **CPU, NPU, GPU, RAM and ROM circuits** directly into the weights of an LLM, then runs our own **Secure OS** on top.
 Inside this OS, you can install and run your own approved apps.
 
 | | |

gpulab.js

@@ -0,0 +1,77 @@
+/* GPU Lab: run a small GPU kernel on the GPU unit wired into the weights of Meta Muse Glimmer 30B (option 3 backend).
+   The result is checked against the reference executor and plain Python. Honors the sleep/wake of the GPU backend. */
+(() => {
+'use strict';
+const K = {
+  vadd: {label:'Vector add', make:n => ({a:rnd(n, 50), b:rnd(n, 50)}), py:d => ({c:d.a.map((v, i) => v + d.b[i])})},
+  dot: {label:'Dot product', make:n => ({a:rnd(n, 12), b:rnd(n, 12)}), py:d => ({out:[d.a.reduce((s, v, i) => s + v * d.b[i], 0)]})},
+  reduce: {label:'Reduce (sum)', make:n => ({x:rnd(n, 99)}), py:d => ({out:[d.x.reduce((s, v) => s + v, 0)]})},
+  histogram: {label:'Histogram (16 bins)', make:n => ({x:rnd(n, 15)}), py:d => ({bins:Array.from({length:16}, (_, i) => d.x.filter(v => v === i).length)})},
+};
+function rnd(n, m) { return Array.from({length:n}, () => Math.floor(Math.random() * (m + 1))); }
+async function gpu(path, body, ms) {
+  const ctl = new AbortController(); const to = setTimeout(() => ctl.abort(), ms || 300000);
+  try {
+    const r = await fetch('api/agx/' + path, body === undefined ? {cache:'no-store', signal:ctl.signal} : {method:'POST', headers:{'Content-Type':'application/json'}, body:JSON.stringify(body), signal:ctl.signal});
+    if (r.status === 429) throw new Error('rate');
+    if (!r.ok) throw new Error('http ' + r.status);
+    return await r.json();
+  } finally { clearTimeout(to); }
+}
+const css = document.createElement('style');
+css.textContent = `.gl{padding:14px;font-family:system-ui,'Segoe UI',sans-serif;display:flex;flex-direction:column;gap:12px;height:100%;overflow:auto}
+.gl h3{margin:0;font-size:18px}.gl .mut{color:var(--muted);font-size:13px}.gl .row{display:flex;gap:8px;flex-wrap:wrap;align-items:center}
+.gl .card{border:1px solid var(--border);border-radius:12px;padding:12px;background:var(--card)}.gl pre{margin:6px 0 0;white-space:pre-wrap;word-break:break-word;font:12px ui-monospace,Consolas,monospace;background:var(--bg);border-radius:8px;padding:8px}
+.gl .ok{color:#16A34A;font-weight:800}.gl .bad{color:#E11D48;font-weight:800}.gl .tag{font:800 11px system-ui;padding:3px 8px;border-radius:999px;background:#EEF2FF;color:#4338CA}
+.gl select,.gl input[type=range]{height:32px}.gl .grid{display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:8px}.gl .kv b{display:block;font-size:20px}`;
+document.head.append(css);
+window.VK_EXTRA_APPS = window.VK_EXTRA_APPS || [];
+window.VK_EXTRA_APPS.push(h => ({id:'gpulab', name:() => 'GPU Lab', color:'#0F766E',
+  icon:'<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><rect x="4" y="4" width="16" height="16" rx="2"/><rect x="8" y="8" width="8" height="8" rx="1"/><path d="M9 1v3M15 1v3M9 20v3M15 20v3M1 9h3M1 15h3M20 9h3M20 15h3"/></svg>',
+  w:820, h:640, render(body) {
+    body.style.padding = '0';
+    const root = document.createElement('div'); root.className = 'gl'; body.append(root);
+    root.innerHTML = `<div><h3>GPU Lab</h3><div class="mut">A GPU unit wired into the weights of Meta Muse Glimmer 30B runs a small kernel inside the model's own forward pass. The result is checked token by token against the reference executor and against plain Python.</div></div>
+      <div class="card row"><label>Kernel <select data-k="k">${Object.entries(K).map(([k, v]) => `<option value="${k}">${v.label}</option>`).join('')}</select></label>
+        <label>Length n <input type="range" min="4" max="16" value="8" data-k="n"> <b data-k="nv">8</b></label>
+        <button class="btn" data-k="rand">Random inputs</button><button class="btn pri" data-k="run">Run on the GPU in the weights</button><span class="tag" data-k="state">checking backend...</span></div>
+      <div class="card"><b>Inputs</b><pre data-k="in"></pre></div>
+      <div class="card" data-k="out"><span class="mut">Pick a kernel and press Run. A run takes from a few seconds to about a minute, because every GPU step is computed by the model.</span></div>`;
+    const q = k => root.querySelector(`[data-k="${k}"]`);
+    let data = null, waking = false;
+    const fresh = () => { data = K[q('k').value].make(+q('n').value); q('in').textContent = JSON.stringify(data); };
+    q('k').onchange = fresh; q('n').oninput = () => { q('nv').textContent = q('n').value; fresh(); }; q('rand').onclick = fresh; fresh();
+    async function health() {
+      try { const r = await gpu('health', undefined, 10000); return !!(r && r.live); } catch (e) { return false; }
+    }
+    async function ensureAwake() {
+      if (await health()) { q('state').textContent = 'GPU backend ready'; return true; }
+      q('state').textContent = 'waking the in-weights model (about 4 min)...';
+      if (waking) return false; waking = true;
+      for (let i = 0; i < 40; i++) { await new Promise(r => setTimeout(r, 15000)); if (!root.isConnected) return false; if (await health()) { waking = false; q('state').textContent = 'GPU backend ready'; h.toast('GPU Lab: the in-weights model is awake'); return true; } }
+      waking = false; q('state').textContent = 'still waking; try again shortly'; return false;
+    }
+    ensureAwake();
+    q('run').onclick = async () => {
+      const out = q('out'), k = q('k').value, n = +q('n').value;
+      if (!(await health())) { out.innerHTML = '<span class="mut">The in-weights model is asleep to save cost. It is waking now (about 4 min). Agent Guard option 1 stays live meanwhile.</span>'; ensureAwake(); return; }
+      q('run').disabled = true; out.innerHTML = '<span class="mut">Running inside the model...</span>';
+      const t0 = performance.now();
+      try {
+        const r = await gpu('gpu', {kernel:k, data, n});
+        if (r.error) throw new Error(r.error);
+        const py = K[k].py(data), wall = (performance.now() - t0) / 1000;
+        const okPy = JSON.stringify(r.outputs) === JSON.stringify(py);
+        out.innerHTML = `<div class="grid"><div class="kv"><span class="mut">exact vs reference</span><b class="${r.equals_reference ? 'ok' : 'bad'}">${r.equals_reference ? '&#10003; exact' : '&#10007; differs'}</b></div>
+          <div class="kv"><span class="mut">token-exact vs executor</span><b class="${r.token_exact ? 'ok' : 'bad'}">${r.token_exact ? '&#10003;' : '&#10007;'}</b></div>
+          <div class="kv"><span class="mut">matches plain Python</span><b class="${okPy ? 'ok' : 'bad'}">${okPy ? '&#10003;' : '&#10007;'}</b></div>
+          <div class="kv"><span class="mut">model tokens</span><b>${(r.tokens || 0).toLocaleString('en-US')}</b></div>
+          <div class="kv"><span class="mut">GPU time in the model</span><b>${r.sec} s</b></div><div class="kv"><span class="mut">round trip</span><b>${wall.toFixed(1)} s</b></div></div>
+          <b style="display:block;margin-top:10px">Output</b><pre>${JSON.stringify(r.outputs)}</pre><b>Plain Python</b><pre>${JSON.stringify(py)}</pre>`;
+      } catch (e) {
+        out.innerHTML = `<span class="mut">${e.message === 'rate' ? 'Too many requests from your network; try again in a few minutes.' : 'The in-weights model is busy or waking. Please try again in a minute.'}</span>`;
+        ensureAwake();
+      } finally { q('run').disabled = false; }
+    };
+  }}));
+})();

i18n.js

@@ -5,7 +5,7 @@ window.VK_I18N = {
     quote: "“For 50 years, the operating system lived inside the computer. Now the operating system lives inside the AI.”",
     tagline: "An OS that boots, decides, remembers and stops inside the weights of an LLM.",
     fullcopy: "Full statement",
-    intro: "ViKernel wires ROM, RAM, and an xPU chip circuit directly into the weights of an LLM, then runs our own Secure OS on top. Inside this OS, you can install and run your own approved apps.",
+    intro: "ViKernel wires CPU, NPU, GPU, RAM and ROM circuits directly into the weights of an LLM, then runs our own Secure OS on top. Inside this OS, you can install and run your own approved apps.",
     p1a: "Boots in the weights.", p1b: " The OS wakes up the moment the model starts. Change a single byte, and it locks itself.",
     p2a: "Decides in the weights.", p2b: " Every AI action is allowed or blocked by rule. Persuasion does not change the verdict.",
     p3a: "Remembers in the weights.", p3b: " Every decision is sealed in a chain that cannot be erased.",
@@ -99,7 +99,7 @@ window.VK_I18N = {
     quote: "“50년 동안 운영체제는 컴퓨터 안에 있었다. 이제 운영체제가 AI 안으로 들어갔다.”",
     tagline: "LLM 가중치 안에서 부팅하고, 판단하고, 기억하고, 멈추는 운영체제.",
     fullcopy: "공식 문안 전체",
-    intro: "ViKernel은 LLM의 가중치 안에 ROM, RAM, xPU 칩 회로도를 직접 배선하고, 그 위에 직접 개발한 시큐어 OS를 올렸습니다. 이 OS 안에서는 승인받은 나만의 앱을 설치해 운영할 수 있습니다.",
+    intro: "ViKernel은 LLM의 가중치 안에 CPU, NPU, GPU, RAM, ROM 회로를 직접 배선하고, 그 위에 직접 개발한 시큐어 OS를 올렸습니다. 이 OS 안에서는 승인받은 나만의 앱을 설치해 운영할 수 있습니다.",
     p1a: "가중치로 부팅한다.", p1b: " 모델을 켜면 OS가 스스로 깨어납니다. 1바이트라도 바뀌면 스스로 잠깁니다.",
     p2a: "가중치로 판단한다.", p2b: " AI의 모든 행동을 규칙에 따라 허용하거나 막습니다. 말로 설득해도 판단은 바뀌지 않습니다.",
     p3a: "가중치로 기억한다.", p3b: " 모든 판단이 지울 수 없는 사슬로 남습니다.",

index.html

@@ -65,7 +65,7 @@
       <div class="atitle">Inside one model file</div>
       <div class="al a-apps">Apps <small>approved only</small></div>
       <div class="al a-os">Secure OS <small>rules · audit chain · 🛑 kill switch</small></div>
-      <div class="al a-hw">ROM · RAM · xPU chip circuit</div>
+      <div class="al a-hw">CPU · NPU · GPU · RAM · ROM circuits</div>
       <div class="al a-w">LLM / VLM weights <small>original weights untouched</small></div>
       <div class="abar"><span>Boots</span><span>Decides</span><span>Remembers</span><span>Stops</span><b>in the weights</b></div>
     </div>
@@ -89,6 +89,7 @@
 <script src="browser.js"></script>
 <script src="settings.js"></script>
 <script src="rsi.js"></script>
+<script src="gpulab.js"></script>
 <script src="store.js"></script>
 <script src="ag.js"></script>
 <script src="app.js"></script>

server.py

@@ -10,8 +10,8 @@ BK_TOKEN = os.environ.get("BACKEND_TOKEN", "")
 RATE = int(os.environ.get("RATE_PER_10MIN", "20"))
 TIMEOUT = float(os.environ.get("BACKEND_TIMEOUT", "300"))
 GET_OK = {"health", "status", "audit", "store/catalog", "browser/img", "browser/res"}
-POST_OK = {"chat", "scenario", "kill", "resume", "tamper", "reset", "store/submit", "store/publish", "store/app", "store/install", "gate/decide", "vmask", "game/hint", "game/cheat", "browser/open", "browser/agent", "browser/snap", "browser/policy", "settings/get", "settings/revoke", "settings/grant", "settings/egress", "settings/policy", "settings/locked", "rsi/state", "rsi/propose", "rsi/apply", "rsi/rollback", "rsi/reset", "ztc/status", "ztc/judge", "ztc/run", "backdoor/run", "backdoor/scan", "backdoor/wake"}
-STATIC = {"index.html", "app.js", "ag.js", "store.js", "common.js", "display.js", "terminal.js", "mines.js", "browser.js", "settings.js", "live_apps.js", "game.js", "rsi.js", "ztc.js", "backdoor.js", "backdoor.css", "backdoor_metrics.json", "downloads/vdrpack.py", "demo/poisoned-page", "samples/weather-plus.vdr", "samples/weather-plus-tampered.vdr", "i18n.js", "demo_data.js", "styles.css", "ag_replay.json"}
+POST_OK = {"chat", "scenario", "kill", "resume", "tamper", "reset", "store/submit", "store/publish", "store/app", "store/install", "gate/decide", "vmask", "game/hint", "game/cheat", "browser/open", "browser/agent", "browser/snap", "browser/policy", "settings/get", "settings/revoke", "settings/grant", "settings/egress", "settings/policy", "settings/locked", "rsi/state", "rsi/propose", "rsi/apply", "rsi/rollback", "rsi/reset", "ztc/status", "ztc/judge", "ztc/run", "backdoor/run", "backdoor/scan", "backdoor/wake", "gpu"}
+STATIC = {"index.html", "app.js", "ag.js", "store.js", "common.js", "display.js", "terminal.js", "mines.js", "browser.js", "settings.js", "live_apps.js", "game.js", "rsi.js", "gpulab.js", "ztc.js", "backdoor.js", "backdoor.css", "backdoor_metrics.json", "downloads/vdrpack.py", "demo/poisoned-page", "samples/weather-plus.vdr", "samples/weather-plus-tampered.vdr", "i18n.js", "demo_data.js", "styles.css", "ag_replay.json"}
 HITS = {}