omnius 1.0.881 → 1.0.883

package/dist/omni-setup-worker.js

@@ -246907,12 +246907,16 @@ async function hostFacts(d) {
     {},
     false
   ).catch(() => ({ code: -1, stdout: "", stderr: "" }));
-  const discovery = discovered.code === 0 && Buffer.byteLength(discovered.stdout) <= 1024 * 1024 ? json(discovered.stdout) : void 0;
-  const installedBytes = await d.readHostFile("/usr/local/share/ollama-unify/gpu-negotiator.json").catch(() => Buffer.alloc(0));
+  const discovery = Buffer.byteLength(discovered.stdout) <= 1024 * 1024 ? json(discovered.stdout) : void 0;
+  let installationAbsent = false;
+  const installedBytes = await d.readHostFile("/usr/local/share/ollama-unify/gpu-negotiator.json").catch((error2) => {
+    installationAbsent = error2.code === "ENOENT";
+    return Buffer.alloc(0);
+  });
   const installed = installedBytes.length <= 1024 * 1024 ? json(installedBytes.toString()) : void 0;
-  const brokerDetected = discovered.code === 0 || installed?.protocol === BROKER_PROTOCOL;
-  const broker = discovery?.protocol === BROKER_PROTOCOL && discovery.available === true && typeof discovery.control_socket === "string" && discovery.control_socket.startsWith("/");
-  const ollamaHost = broker ? loopbackOllama(discovery?.public_ollama_api) ?? (installed?.protocol === BROKER_PROTOCOL ? loopbackOllama(installed.public_ollama_api) : void 0) : !brokerDetected && capableTegra ? "http://127.0.0.1:11434" : void 0;
+  const brokerDetected = discovered.code === 0 || discovery?.protocol === BROKER_PROTOCOL || !installationAbsent;
+  const broker = discovered.code === 0 && discovery?.protocol === BROKER_PROTOCOL && discovery.available === true && typeof discovery.control_socket === "string" && discovery.control_socket.startsWith("/");
+  const ollamaHost = broker ? loopbackOllama(discovery?.public_ollama_api) ?? (installed?.protocol === BROKER_PROTOCOL ? loopbackOllama(installed.public_ollama_api) : void 0) : !brokerDetected ? "http://127.0.0.1:11434" : void 0;
   const warnings = Array.isArray(discovery?.warnings) ? discovery.warnings.slice(0, 12).map((warning) => safeSummary(warning)) : [];
   const activeLeases = Array.isArray(discovery?.active_leases) ? discovery.active_leases : [];
   const summaries = activeLeases.slice(0, 12).map(
@@ -246947,7 +246951,7 @@ async function hostFacts(d) {
       );
       const capable = Number.isFinite(Number(memory)) && Number(memory) >= OMNI_RUNTIME_RESERVATIONS_MIB.ornith15 && Number.isFinite(Number(compute)) && Number(compute) >= 7;
       const eligible = Boolean(
-        policyReady && selected && capable && !held && !quarantined && !missing && !transitioning && ollamaHost
+        capable && !held && (!brokerDetected && !tegra || policyReady && selected && !quarantined && !missing && !transitioning) && ollamaHost
       );
       devices.push({
         uuid,
@@ -246955,7 +246959,7 @@ async function hostFacts(d) {
         eligible,
         ...Number.isFinite(Number(memory)) ? { memoryMiB: Number(memory) } : {},
         ...!eligible ? {
-          reason: held ? "Operator inference hold is active." : !broker ? "GPU broker discovery is unavailable or invalid." : !policyReady ? "GPU broker health or admission metadata does not permit new work." : !selected || missing ? "GPU is outside the healthy broker-selected scope." : quarantined ? "GPU has quarantined unregistered work." : transitioning ? "GPU has a pending or revoking lease." : !capable ? "Requires compute capability 7.0+ and at least 24,000 MiB VRAM; Qwen requires 45,000 MiB." : "A verified loopback Ollama service endpoint is required."
+          reason: !brokerDetected && tegra ? "Integrated Tegra deployments require the supported SoC admission path." : held ? "Operator inference hold is active." : brokerDetected && !broker ? "GPU broker discovery is unavailable or invalid." : brokerDetected && !policyReady ? "GPU broker health or admission metadata does not permit new work." : brokerDetected && (!selected || missing) ? "GPU is outside the healthy broker-selected scope." : brokerDetected && quarantined ? "GPU has quarantined unregistered work." : brokerDetected && transitioning ? "GPU has a pending or revoking lease." : !capable ? "Requires compute capability 7.0+ and at least 24,000 MiB VRAM; Qwen requires 45,000 MiB." : "A verified loopback Ollama service endpoint is required."
         } : {}
       });
     }
@@ -246991,6 +246995,7 @@ async function inspectOmniHost(input, options = {}) {
     `Derived speech GGUF cache: ${join13(p.runtime, "components")} (additional local disk usage, retained on cancellation).`,
     "GPU reservation: Ornith 24,000 MiB; Qwen 45,000 MiB. Download size is not a VRAM promise; Tegra uses upstream UMA admission.",
     `Verified local Ollama endpoint: ${facts.ollamaHost ?? "unavailable"}`,
+    facts.brokerDetected ? "GPU admission uses the installed broker; invalid discovery, holds and lease transitions block setup." : "No GPU broker installation was found. The upstream direct supervisor requires the chosen accelerator, current capacity and actual owned-runner residency proof; it does not coordinate external GPU workloads.",
     `Private operation diagnostics: ${join13(p.root, "setup-<operationId>.log")} (bounded, owner-only).`,
     ...facts.warnings.map((warning) => `Broker warning: ${warning}`),
     ...facts.summaries
@@ -247008,7 +247013,8 @@ async function inspectOmniHost(input, options = {}) {
     "ffmpeg",
     "ollama",
     "systemctl",
-    ...facts.broker ? ["docker", "jq", "curl", "openssl", "setsid", "ss", "nvidia-smi"] : []
+    ...!facts.tegra || facts.broker ? ["nvidia-smi"] : [],
+    ...facts.broker ? ["docker", "jq", "curl", "openssl", "setsid", "ss"] : []
   ]) {
     const found = await command(
       d,
@@ -247032,10 +247038,6 @@ async function inspectOmniHost(input, options = {}) {
   ).catch(() => ({ code: -1, stdout: "", stderr: "" }));
   if (!["running", "degraded"].includes(userBus.stdout.trim()))
     blockedReasons.push("A working systemd user manager is required.");
-  if (!facts.brokerDetected && !facts.tegra)
-    blockedReasons.push(
-      "Discrete CUDA deployment requires the ollama-unify GPU broker; direct or CPU fallback is disabled."
-    );
   if (facts.brokerDetected && !facts.broker)
     blockedReasons.push(
       "Installed GPU broker discovery is unavailable or invalid; direct fallback is disabled."
@@ -247243,7 +247245,7 @@ async function proof(d, p, model, gpuUuid, direct, startedAt, signal) {
     const state = json(
       (await optionalFile(join13(p.runtime, "state", "daemon-status.json")))?.toString() ?? ""
     );
-    if (state?.schema !== "robit.qwen-omni-daemon.status.v1" || state.state !== "ready" || state.model !== model || state.pid !== status.pid || state.updated_at * 1e3 < startedAt || state.startup_smoke !== false)
+    if (state?.schema !== "robit.qwen-omni-daemon.status.v1" || state.state !== "ready" || state.model !== model || state.pid !== status.pid || !Number.isFinite(state.updated_at) || state.updated_at * 1e3 < startedAt || state.startup_smoke !== false)
       return false;
     const child = Array.isArray(state.children) ? state.children.find((item) => item.name === "comprehension") : void 0;
     if (!child || !Number.isSafeInteger(child.pid) || !pids.includes(child.pid))
@@ -247253,14 +247255,17 @@ async function proof(d, p, model, gpuUuid, direct, startedAt, signal) {
       p.python,
       [
         "-c",
-        "import json,sys; from qwen_omni_adapters.accelerator import process_is_gpu_resident,residency_backend; print(json.dumps({'resident':process_is_gpu_resident(int(sys.argv[1])),'backend':residency_backend()}))",
-        String(child.pid)
+        "import json,sys; from qwen_omni_adapters.accelerator import process_is_gpu_resident,residency_backend; print(json.dumps({'resident':process_is_gpu_resident(int(sys.argv[1]),sys.argv[2]),'backend':residency_backend()}))",
+        String(child.pid),
+        gpuUuid === "tegra-integrated" ? "" : gpuUuid
       ],
       { signal, timeoutMs: 5e3 },
       false
     );
     const resident = json(result.stdout);
-    return resident?.resident === true && resident.backend === "tegra-device-handles";
+    if (result.code !== 0 || resident?.resident !== true) return false;
+    if (gpuUuid === "tegra-integrated") return resident.backend === "tegra-device-handles";
+    return resident.backend === "nvidia-smi-compute-apps" && await discreteResidency(d, p, pids, gpuUuid, signal, child.pid);
   }
   const ready = await command(
     d,
@@ -247297,6 +247302,9 @@ async function proof(d, p, model, gpuUuid, direct, startedAt, signal) {
     (lease) => lease.owner === p.ownerName && lease.state === "active" && Array.isArray(lease.gpu_uuids) && lease.gpu_uuids.length === 1 && lease.gpu_uuids[0] === gpuUuid
   ))
     return false;
+  return discreteResidency(d, p, pids, gpuUuid, signal);
+}
+async function discreteResidency(d, p, pids, gpuUuid, signal, requiredPid) {
   const processes = await command(
     d,
     "nvidia-smi",
@@ -247318,7 +247326,7 @@ async function proof(d, p, model, gpuUuid, direct, startedAt, signal) {
         "Owned CUDA work was observed outside the designated accelerator."
       );
     const cmd = (await d.readHostFile(`/proc/${pid}/cmdline`)).toString().split("\0"), env = (await d.readHostFile(`/proc/${pid}/environ`)).toString().split("\0");
-    if (cmd[0] === join13(
+    if ((requiredPid === void 0 || requiredPid === pid) && cmd[0] === join13(
       p.checkout,
       "vendor",
       "llama.cpp",
@@ -247330,6 +247338,13 @@ async function proof(d, p, model, gpuUuid, direct, startedAt, signal) {
   }
   return seen;
 }
+async function admitDirectDiscrete(d, gpuUuid, requiredMiB, signal) {
+  const result = await command(d, "nvidia-smi", ["--query-gpu=uuid,memory.free", "--format=csv,noheader,nounits"], { signal, timeoutMs: 5e3 }, false);
+  const rows = result.stdout.split("\n").map((line) => line.split(",").map((value) => value.trim())).filter((row) => row[0] === gpuUuid);
+  const freeMiB = rows.length === 1 && rows[0].length === 2 ? Number(rows[0][1]) : NaN;
+  if (result.code !== 0 || !Number.isFinite(freeMiB) || freeMiB < requiredMiB)
+    fail("accelerator_not_admitted", "The chosen GPU does not currently report sufficient free VRAM for the selected profile. No service was started; no external workload was stopped.");
+}
 async function verifySource(d, p, signal) {
   const revision = await command(
     d,
@@ -247571,7 +247586,8 @@ ${diagnostic.stderr.slice(-65536)}` : ""}`,
         "An existing healthy Omni stack must be attached explicitly; it will not be replaced."
       );
     const facts = await hostFacts(d);
-    direct = facts.tegra && !facts.brokerDetected;
+    direct = !facts.brokerDetected;
+    const directTegra = direct && facts.tegra;
     if (facts.held)
       return {
         state: "held-paused",
@@ -247586,7 +247602,7 @@ ${diagnostic.stderr.slice(-65536)}` : ""}`,
         "accelerator_required",
         "Choose an exact eligible accelerator from the current plan."
       );
-    if (direct && input.gpuUuid !== "tegra-integrated" || !direct && !GPU_UUID.test(input.gpuUuid))
+    if (directTegra && input.gpuUuid !== "tegra-integrated" || !directTegra && !GPU_UUID.test(input.gpuUuid))
       fail(
         "accelerator_required",
         "The selected accelerator does not match this host's deployment mode."
@@ -247594,7 +247610,7 @@ ${diagnostic.stderr.slice(-65536)}` : ""}`,
     const selectedDevice = facts.devices.find(
       (device) => device.uuid === input.gpuUuid
     );
-    if (!direct && (selectedDevice.memoryMiB ?? 0) < OMNI_RUNTIME_RESERVATIONS_MIB[input.profile])
+    if (!directTegra && (selectedDevice.memoryMiB ?? 0) < OMNI_RUNTIME_RESERVATIONS_MIB[input.profile])
       fail(
         "insufficient_vram",
         `The selected profile requires at least ${OMNI_RUNTIME_RESERVATIONS_MIB[input.profile].toLocaleString("en-US")} MiB total GPU memory before any build or model download.`
@@ -247747,7 +247763,7 @@ ${diagnostic.stderr.slice(-65536)}` : ""}`,
       OMNI_LANGUAGE_MODEL: profile.model,
       OMNI_PORTAL_RUNTIME_ROOT: p.runtime,
       OMNI_COMPREHENSION_GPU_UUID: input.gpuUuid,
-      CUDA_VISIBLE_DEVICES: direct ? "0" : input.gpuUuid,
+      CUDA_VISIBLE_DEVICES: directTegra ? "0" : input.gpuUuid,
       HIP_VISIBLE_DEVICES: "-1",
       ROCR_VISIBLE_DEVICES: "-1",
       OMNI_ENABLE_POINTING: "0",
@@ -247756,7 +247772,7 @@ ${diagnostic.stderr.slice(-65536)}` : ""}`,
       OMNI_STARTUP_SMOKE: "0",
       OMNI_PRINT_ACCESS_URL: "0",
       OMNI_KEEP_CACHE: "1",
-      OMNI_COMPREHENSION_CONTEXT_TOKENS: direct && input.profile === "qwen38" ? "16384" : "65536",
+      OMNI_COMPREHENSION_CONTEXT_TOKENS: directTegra && input.profile === "qwen38" ? "16384" : "65536",
       OMNI_COMPREHENSION_VRAM_MIB: String(
         OMNI_RUNTIME_RESERVATIONS_MIB[input.profile]
       ),
@@ -247797,13 +247813,13 @@ ${diagnostic.stderr.slice(-65536)}` : ""}`,
         message: "Runtime source and private configuration are installed; operator pause is preserved and no readiness is claimed."
       };
     if (fresh.broker !== facts.broker || fresh.brokerDetected !== facts.brokerDetected || fresh.ollamaHost !== facts.ollamaHost || !fresh.devices.some(
-      (device) => device.uuid === input.gpuUuid && device.eligible && (direct || (device.memoryMiB ?? 0) >= OMNI_RUNTIME_RESERVATIONS_MIB[input.profile])
+      (device) => device.uuid === input.gpuUuid && device.eligible && (directTegra || (device.memoryMiB ?? 0) >= OMNI_RUNTIME_RESERVATIONS_MIB[input.profile])
     ))
       fail(
         "accelerator_policy_changed",
         "Accelerator admission changed before service activation."
       );
-    if (direct) {
+    if (directTegra) {
       const admission = await command(
         d,
         p.python,
@@ -247821,7 +247837,7 @@ ${diagnostic.stderr.slice(-65536)}` : ""}`,
           "accelerator_not_admitted",
           "Tegra memory and utilization admission did not pass."
         );
-    }
+    } else if (direct) await admitDirectDiscrete(d, input.gpuUuid, OMNI_RUNTIME_RESERVATIONS_MIB[input.profile], options.signal);
     linked = true;
     await command(d, "systemctl", ["--user", "link", p.unitPath], {
       signal: options.signal
@@ -247859,6 +247875,7 @@ ${diagnostic.stderr.slice(-65536)}` : ""}`,
           adapterUrl: "http://127.0.0.1:8910",
           profile: input.profile,
           sourceRevision: OMNI_SETUP_SOURCE.revision,
+          accelerator: { mode: direct ? directTegra ? "direct-tegra" : "direct-discrete" : "broker", gpuUuid: input.gpuUuid },
           verified: {
             adapterContract: true,
             runtime: true,
@@ -300707,8 +300724,14 @@ function stored(root, id2) {
   if (!value) throw error("setup_job_missing", "Omni setup job was not found.", 404);
   if (value.version !== 1 || value.job?.id !== id2 || value.job.schema !== "omnius.omni-setup-job.v1" || !states.has(value.job.state) || !phases.has(value.job.phase) || !ID2.test(value.owner?.token ?? "") || !/^\d+$/.test(value.owner.start) || !Number.isInteger(value.owner.pid) || value.owner.pid <= 0 || !/^[0-9a-f]{64}$/.test(value.configurationRevision) || typeof value.gpuUuid !== "string" || !Number.isFinite(Date.parse(value.job.createdAt)) || !Number.isFinite(Date.parse(value.job.updatedAt)) || value.job.sourceRevision !== OMNI_SETUP_SOURCE.revision || !OMNI_SETUP_PROFILES.some((p) => p.id === value.job.profile && p.model === value.job.model)) throw error("invalid_setup_state", "Omni setup state has an invalid identity.");
   if (value.job.state === "ready" && (!value.job.receipt || value.job.receipt.model !== value.job.model || value.job.receipt.sourceRevision !== OMNI_SETUP_SOURCE.revision || value.job.receipt.action !== "deployed" || !value.job.receipt.verified?.adapterContract || !value.job.receipt.verified.runtime || !value.job.receipt.verified.accelerator)) throw error("invalid_setup_state", "Omni setup state is missing its readiness proof.");
+  if (value.job.receipt?.accelerator && !publicAccelerator(value.job.receipt.accelerator, value.gpuUuid)) throw error("invalid_setup_state", "Omni setup accelerator proof does not match the selected device.");
   return value;
 }
+function publicAccelerator(value, selected) {
+  if (!value || value.gpuUuid !== selected || !["broker", "direct-discrete", "direct-tegra"].includes(value.mode)) return;
+  if (value.mode === "direct-tegra" ? selected !== "tegra-integrated" : !/^GPU-[0-9a-f]{8}(?:-[0-9a-f]{4}){3}-[0-9a-f]{12}$/i.test(selected)) return;
+  return { mode: value.mode, gpuUuid: selected };
+}
 async function runOmniSetupJob(rootPath, id2) {
   const root = resolve11(rootPath), path3 = join14(root, "jobs", id2 + ".json");
   directory(root);
@@ -300787,6 +300810,7 @@ async function runOmniSetupJob(rootPath, id2) {
           throw new DOMException("Setup cancelled before commit", "AbortError");
         }
         if (receipt.model !== value.job.model || receipt.sourceRevision !== OMNI_SETUP_SOURCE.revision || receipt.action !== "deployed" || !receipt.verified.adapterContract || !receipt.verified.runtime || !receipt.verified.accelerator) throw error("setup_readiness_unverified", "Setup did not return complete runtime and accelerator proof.");
+        if (receipt.accelerator && !publicAccelerator(receipt.accelerator, value.gpuUuid)) throw error("setup_readiness_unverified", "Setup accelerator proof does not match the selected device.");
         const previous = readSavedOmniConfiguration(config.path);
         const revision = saveOmniConfiguration({ version: 1, model: receipt.model, adapterUrl: receipt.adapterUrl }, value.configurationRevision, config.path);
         try {

package/dist/postinstall-daemon.cjs

@@ -1093,6 +1093,14 @@ function repairBrokenWrappers() {
 function main() {
   if (!ensureCoreCommandAvailable()) return fatalExit(1);
 
+  // The coordinated worker already owns service handoff and verifies the
+  // replacement package/runtime. Ancillary model/native/companion setup must
+  // not hold this package transaction open or mutate unrelated runtime state.
+  if (process.env.OMNIUS_UPDATE_COORDINATED === "1") {
+    log("Coordinated update owns daemon restart and exact-version verification.");
+    return safeExit(0);
+  }
+
   // Always do the nexus cleanup first, regardless of opt-out.
   progress(1, INSTALL_STEPS, "Cleaning stale Nexus runtime state");
   cleanNexus();
@@ -1113,11 +1121,6 @@ function main() {
   progress(5, INSTALL_STEPS, "Ensuring local REST key");
   ensureBootstrapApiKeyForUser(effectiveUser());
 
-  if (process.env.OMNIUS_UPDATE_COORDINATED === "1") {
-    log("Coordinated update owns daemon restart and exact-version verification.");
-    return safeExit(0);
-  }
-
   if (process.env.OMNIUS_SKIP_DAEMON_INSTALL === "1") {
     log("OMNIUS_SKIP_DAEMON_INSTALL=1 — skipping daemon service install.");
     return safeExit(0);