@tuned-tensor/cli 0.13.0-beta.1 → 0.13.0

package/dist/local-runtime.js

@@ -66,7 +66,7 @@ import {
   withHuggingFaceCacheEnvironment,
   withOfflineHuggingFaceCacheEnvironment,
   writeJsonAtomic
-} from "./chunk-2VJ3VVZF.js";
+} from "./chunk-YJYCGSHJ.js";
 
 // src/local-runtime/index.ts
 import { readFileSync, realpathSync } from "fs";
@@ -221,6 +221,14 @@ async function runCommand(command, args, options = {}) {
 function firstLine(value) {
   return value.trim().split(/\r?\n/)[0] ?? "";
 }
+function nvidiaSmiSummary(stdout) {
+  const lines = stdout.split(/\r?\n/).map((line) => line.trim()).filter(Boolean);
+  const queried = lines.find(
+    (line) => !/^(?:mon|tue|wed|thu|fri|sat|sun)\b/i.test(line) && !line.startsWith("+---") && !line.startsWith("|")
+  );
+  if (queried) return queried;
+  return firstLine(stdout) || "nvidia-smi reported a GPU";
+}
 function commandText(command, args) {
   return [command, ...args].map((part) => /\s/.test(part) ? JSON.stringify(part) : part).join(" ");
 }
@@ -374,11 +382,15 @@ async function runDoctor(config, request) {
   }
   const device = config.evaluation.inference.device;
   if (!config.dryRun) {
-    const nvidiaSmi = await runCommand("nvidia-smi", []);
+    const queried = await runCommand("nvidia-smi", [
+      "--query-gpu=name,driver_version",
+      "--format=csv,noheader"
+    ]);
+    const nvidiaSmi = queried.code === 0 ? queried : await runCommand("nvidia-smi", []);
     checks.push({
       name: "nvidia-smi",
       ok: nvidiaSmi.code === 0,
-      message: nvidiaSmi.code === 0 ? firstLine(nvidiaSmi.stdout) : nvidiaSmi.error ?? (firstLine(nvidiaSmi.stderr) || "nvidia-smi not available")
+      message: nvidiaSmi.code === 0 ? nvidiaSmiSummary(nvidiaSmi.stdout) : nvidiaSmi.error ?? (firstLine(nvidiaSmi.stderr) || "nvidia-smi not available")
     });
   } else {
     checks.push({
@@ -672,7 +684,7 @@ var TT_LOCAL_VERSION = packageVersion();
 function getLocalRunnerInfo() {
   return {
     name: "tuned-tensor-local",
-    status: "local-runner-preview",
+    status: "local",
     description: "Local CUDA LoRA fine-tuning with held-out base-versus-tuned evaluation.",
     version: TT_LOCAL_VERSION
   };
@@ -1071,6 +1083,17 @@ async function verifyStoredModel(model, config) {
   };
 }
 async function verifyActivationEvidence(model) {
+  const generalBaseline = join3(model.artifact_dir, "general-baseline-eval.json");
+  const generalCandidate = join3(model.artifact_dir, "general-candidate-eval.json");
+  const [hasBaseline, hasCandidate] = await Promise.all([
+    stat(generalBaseline).then((metadata) => metadata.isFile(), () => false),
+    stat(generalCandidate).then((metadata) => metadata.isFile(), () => false)
+  ]);
+  if (!hasBaseline || !hasCandidate) {
+    throw new Error(
+      "This model cannot be activated because the run did not include a general-regression suite. Add evaluation.generalRegression to local-runner.json with a held-out dataset, re-run training, then activate. Until then, serve the adapter with `tt serve local-<run-id>`."
+    );
+  }
   await assertArtifactManifest(join3(model.artifact_dir, "artifact-manifest.json"), {
     requiredPaths: [
       "run-report.json",
@@ -1226,8 +1249,6 @@ async function serveStoredModelFromCli(args) {
     return;
   }
   process.stderr.write(`[tt-local] verified ${verified.integrity.checked} artifact file(s)
-`);
-  process.stderr.write(`[tt-local] model API: ${launch.url}
 `);
   await serveLocalModel(launch);
 }
@@ -1285,8 +1306,6 @@ async function serveBaseModelFromCli(args) {
     });
     return;
   }
-  process.stderr.write(`[tt-local] protected base model API: ${launch.url}
-`);
   await serveLocalModel(launch);
 }
 async function serveModelTargetFromCli(args) {
@@ -1296,7 +1315,11 @@ async function serveModelTargetFromCli(args) {
   if (args.target === "active") {
     const store = createLocalStore(args.config.storeRoot);
     const active = await getActiveModel(store);
-    if (!active.model) return serveBaseModelFromCli(args);
+    if (!active.model) {
+      throw new Error(
+        "No adapter is active. `tt serve active` would serve the protected base model. Activate a verified adapter first with `tt models activate <model-id>`, or pass `tt serve base` explicitly."
+      );
+    }
     return serveStoredModelFromCli({
       argv: args.argv,
       modelId: active.model.id,
@@ -1515,7 +1538,13 @@ async function main(argv) {
     if (subcommand === "get") {
       const id = cli.positionals[0];
       if (!id) throw new Error("models get requires <model-id>");
-      return printJson(await store.getModel(id));
+      const model = await store.getModel(id);
+      const verified = await verifyStoredModel(model, config);
+      return printJson({
+        ok: true,
+        model,
+        ...verified
+      });
     }
     if (subcommand === "verify") {
       const id = cli.positionals[0];