fix(docker): relax vLLM pull stall watchdog

src/lib/adapters/docker/pull.test.ts

@@ -107,6 +107,65 @@ describe("docker pull progress watchdog", () => {
     });
   });
 
+  it("allows quiet pull finalization within the 15 minute stall timeout", async () => {
+    vi.useFakeTimers();
+    const child = new FakeChild();
+    const pull = dockerPullWithProgressWatchdog("example/image:latest", {
+      suppressOutput: true,
+      spawnImpl: () => child,
+    });
+
+    child.stderr.emit("data", Buffer.from("abc123def: Pull complete\n"));
+    await vi.advanceTimersByTimeAsync(15 * 60_000 - 10_000);
+    expect(child.kill).not.toHaveBeenCalled();
+
+    child.emit("close", 0, null);
+    await expect(pull).resolves.toMatchObject({
+      status: 0,
+      timedOut: false,
+      timeoutKind: null,
+    });
+  });
+
+  it("kills a quiet pull after the default 15 minute stall timeout", async () => {
+    vi.useFakeTimers();
+    const child = new FakeChild();
+    const pull = dockerPullWithProgressWatchdog("example/image:latest", {
+      suppressOutput: true,
+      spawnImpl: () => child,
+    });
+
+    child.stderr.emit("data", Buffer.from("abc123def: Pull complete\n"));
+    await vi.advanceTimersByTimeAsync(16 * 60_000);
+
+    await expect(pull).resolves.toMatchObject({
+      status: 124,
+      timedOut: true,
+      timeoutKind: "stall",
+    });
+    expect(child.kill).toHaveBeenCalledWith("SIGTERM");
+  });
+
+  it("respects a custom stall timeout", async () => {
+    vi.useFakeTimers();
+    const child = new FakeChild();
+    const pull = dockerPullWithProgressWatchdog("example/image:latest", {
+      suppressOutput: true,
+      stallTimeoutMs: 5 * 60_000,
+      spawnImpl: () => child,
+    });
+
+    child.stderr.emit("data", Buffer.from("abc123def: Pull complete\n"));
+    await vi.advanceTimersByTimeAsync(5 * 60_000);
+
+    await expect(pull).resolves.toMatchObject({
+      status: 124,
+      timedOut: true,
+      timeoutKind: "stall",
+    });
+    expect(child.kill).toHaveBeenCalledWith("SIGTERM");
+  });
+
   it("kills a pull when output repeats without forward progress", async () => {
     vi.useFakeTimers();
     const child = new FakeChild();

src/lib/adapters/docker/pull.ts

@@ -12,7 +12,11 @@ export function dockerPull(imageRef: string, opts: DockerRunOptions = {}): Docke
   return dockerRun(["pull", imageRef], opts);
 }
 
-export const DEFAULT_DOCKER_PULL_STALL_TIMEOUT_MS = 120 * 1000;
+// DGX Spark vLLM pulls can spend several minutes quiet while Docker finalizes
+// large NGC layers. Keep the stall watchdog above the reported ~5.5 minute
+// successful control pull, while the separate max timeout still bounds total
+// wall-clock runtime.
+const DEFAULT_DOCKER_PULL_STALL_TIMEOUT_MS = 15 * 60 * 1000;
 export const DEFAULT_DOCKER_PULL_MAX_TIMEOUT_MS = 12 * 60 * 60 * 1000;
 const DOCKER_PULL_OUTPUT_TAIL_LINES = 200;
 const DOCKER_PULL_PROGRESS_STATE_LIMIT = 512;

src/lib/inference/vllm.ts

@@ -227,6 +227,9 @@ function dockerPrereqsOk(): { ok: boolean; reason?: string } {
 
 export async function pullImage(profile: VllmProfile): Promise<{ ok: boolean; reason?: string }> {
   emit(`Pulling vLLM image: ${profile.image}`);
+  // Docker can be quiet while finalizing large layers on every supported vLLM
+  // profile, so all profiles intentionally share the 15-minute stall default.
+  // The profile-specific maximum still bounds the complete pull operation.
   const result = await dockerPullWithProgressWatchdog(profile.image, {
     maxTimeoutMs: profile.pullTimeoutSec * 1000,
     logLine: emit,