fix(onboard): use Nemotron endpoint probe parameters

src/lib/inference/health.test.ts

@@ -186,6 +186,28 @@ describe("inference health", () => {
       expect(payload.model).toBe("meta/llama-3.3-70b-instruct");
     });
 
+    it.each(["nvidia-prod", "nvidia-nim"])(
+      "uses the NVIDIA Endpoints request shape for Nemotron 3 Super health through %s (#10880)",
+      (provider) => {
+        let capturedArgv: string[] = [];
+        const result = probeRemoteProviderHealth(provider, {
+          model: "nvidia/nemotron-3-super-120b-a12b",
+          getCredentialImpl: () => "nvapi-test",
+          runCurlProbeImpl: (argv) => {
+            capturedArgv = argv;
+            return httpOk();
+          },
+        });
+
+        expect(result?.ok).toBe(true);
+        expect(JSON.parse(capturedArgv[capturedArgv.indexOf("-d") + 1])).toMatchObject({
+          temperature: 1,
+          top_p: 0.95,
+          chat_template_kwargs: { enable_thinking: false },
+        });
+      },
+    );
+
     it("always resolves NVIDIA credentials from NVIDIA_INFERENCE_API_KEY, not the route's default credential env", () => {
       let resolvedEnvNames: string[] = [];
       const result = probeRemoteProviderHealth("nvidia-nim", {

src/lib/inference/health.ts

@@ -18,6 +18,7 @@ import type { LocalProviderHealthProbeOptions } from "./local";
 import { probeLocalProviderHealth } from "./local";
 import { MIN_PROBE_REPLY_TOKENS } from "./max-tokens-field";
 import { getChatCompletionsProbeCurlArgs } from "./onboard-probes";
+import { usesNvidiaEndpointProbePayload } from "./openai-probe-models";
 import { BUILD_ENDPOINT_URL } from "./provider-models";
 
 export interface ProviderHealthStatus {
@@ -55,7 +56,6 @@ export interface ProviderHealthProbeOptions {
 }
 
 const COMPATIBLE_PROVIDERS = new Set(["compatible-endpoint", "compatible-anthropic-endpoint"]);
-const NVIDIA_MANAGED_PROVIDERS = new Set(["nvidia-prod", "nvidia-nim"]);
 const NVIDIA_HEALTH_CREDENTIAL_ENV = "NVIDIA_INFERENCE_API_KEY";
 const HEALTH_PROBE_CONNECT_TIMEOUT_SECONDS = "3";
 const HEALTH_PROBE_MAX_TIME_SECONDS = "5";
@@ -125,6 +125,7 @@ function buildChatCompletionsStatusProbeCurlArgs(
   endpoint: string,
   authArgs: readonly string[],
   isWsl?: boolean,
+  useNvidiaEndpointProbePayload = false,
 ): string[] {
   const args = capStatusProbeOutput(
     useStatusProbeTiming(
@@ -133,6 +134,7 @@ function buildChatCompletionsStatusProbeCurlArgs(
         model,
         url: endpoint,
         isWsl,
+        useNvidiaEndpointProbePayload,
       }),
     ),
   );
@@ -500,6 +502,7 @@ function probeChatCompletionsProviderHealth(
   credentialEnv: string,
   endpoint: string,
   options: ProviderHealthProbeOptions,
+  useNvidiaEndpointProbePayload = false,
 ): ProviderHealthStatus {
   let apiKey = "";
   try {
@@ -523,7 +526,13 @@ function probeChatCompletionsProviderHealth(
   const rawResult = (() => {
     try {
       return runCurlProbeImpl(
-        buildChatCompletionsStatusProbeCurlArgs(model, endpoint, authConfig.args, options.isWsl),
+        buildChatCompletionsStatusProbeCurlArgs(
+          model,
+          endpoint,
+          authConfig.args,
+          options.isWsl,
+          useNvidiaEndpointProbePayload,
+        ),
         { trustedConfigFiles: authConfig.trustedConfigFiles },
       );
     } finally {
@@ -652,13 +661,14 @@ export function probeRemoteProviderHealth(
 
   if (!config?.model) return null;
 
-  if (NVIDIA_MANAGED_PROVIDERS.has(provider)) {
+  if (usesNvidiaEndpointProbePayload(provider)) {
     return probeChatCompletionsProviderHealth(
       providerLabel,
       config.model,
       NVIDIA_HEALTH_CREDENTIAL_ENV,
       `${BUILD_ENDPOINT_URL}/chat/completions`,
       options,
+      true,
     );
   }
 

src/lib/inference/onboard-probes.test.ts

@@ -309,8 +309,35 @@ describe("OpenAI-compatible inference probes", () => {
     });
   });
 
-  it("keeps the default chat-completions probe bounded for other models", () => {
-    expect(getChatCompletionsProbePayload("nvidia/nemotron-3-super-120b-a12b")).toEqual({
+  it("serializes the Nemotron 3 Super validation request parameters (#10880)", () => {
+    const args = getChatCompletionsProbeCurlArgs({
+      credentialArgs: FAKE_CREDENTIAL_ARGS,
+      model: "nvidia/nemotron-3-super-120b-a12b",
+      url: "https://integrate.api.nvidia.com/v1/chat/completions",
+      isWsl: false,
+      useNvidiaEndpointProbePayload: true,
+    });
+
+    expect(args).toContain("-d");
+    expect(JSON.parse(args[args.indexOf("-d") + 1])).toEqual({
+      model: "nvidia/nemotron-3-super-120b-a12b",
+      messages: [{ role: "user", content: "Reply with exactly: OK" }],
+      max_tokens: 16,
+      temperature: 1,
+      top_p: 0.95,
+      chat_template_kwargs: { enable_thinking: false },
+    });
+  });
+
+  it("keeps compatible endpoints on the generic request shape for the same Nemotron model (#10880)", () => {
+    const args = getChatCompletionsProbeCurlArgs({
+      credentialArgs: FAKE_CREDENTIAL_ARGS,
+      model: "nvidia/nemotron-3-super-120b-a12b",
+      url: "https://compatible.example.test/v1/chat/completions",
+      isWsl: false,
+    });
+
+    expect(JSON.parse(args[args.indexOf("-d") + 1])).toEqual({
       model: "nvidia/nemotron-3-super-120b-a12b",
       messages: [{ role: "user", content: "Reply with exactly: OK" }],
       max_tokens: 16,
@@ -905,7 +932,7 @@ exit 0
     });
 
     it("retries chat-completions when /responses errors then chat-completions times out", () => {
-      const script = `#!/usr/bin/env bash
+        const script = `#!/usr/bin/env bash
 outfile=""
 url=""
 while [ "$#" -gt 0 ]; do
@@ -954,7 +981,7 @@ exit 0
     });
 
     it("preserves query-param auth on doubled-timeout chat-completions retry", () => {
-        const script = `#!/usr/bin/env bash
+      const script = `#!/usr/bin/env bash
 outfile=""
 n=$(cat "${HARNESS_COUNTER}")
 n=$((n + 1))
@@ -1417,7 +1444,8 @@ exit 0
 });
 
 describe("onboard inference smoke abort cleanup", () => {
-  it("tears down the orphan managed gateway before exiting after a failed smoke", async () => {
+  it("uses the legacy NVIDIA Endpoints payload before cleaning up a failed smoke (#10880)", async () => {
+    const optimizedProbe = vi.fn().mockResolvedValue({ ok: false, message: "smoke failed" });
     const teardownOrphanManagedGatewayOnAbort = vi.fn();
     const exit = vi.spyOn(process, "exit").mockImplementation((() => undefined) as never);
     const error = vi.spyOn(console, "error").mockImplementation(() => undefined);
@@ -1428,18 +1456,18 @@ describe("onboard inference smoke abort cleanup", () => {
         {
           endpointUrl: "https://inference.example.com/v1",
           forceOpenAiLike: true,
-          model: "example/model",
-          provider: "example-provider",
+          model: "nvidia/nemotron-3-super-120b-a12b",
+          provider: "nvidia-nim",
         },
         {
-          probeOpenAiLikeEndpointOptimized: vi.fn().mockResolvedValue({
-            ok: false,
-            message: "smoke failed",
-          }),
+          probeOpenAiLikeEndpointOptimized: optimizedProbe,
           teardownOrphanManagedGatewayOnAbort,
         },
       );
 
+      expect(optimizedProbe.mock.calls[0]?.[3]).toMatchObject({
+        useNvidiaEndpointProbePayload: true,
+      });
       expect(teardownOrphanManagedGatewayOnAbort).toHaveBeenCalledOnce();
       expect(exit).toHaveBeenCalledWith(1);
       expect(teardownOrphanManagedGatewayOnAbort.mock.invocationCallOrder[0]).toBeLessThan(

src/lib/inference/onboard-probes.ts

@@ -35,7 +35,11 @@ const {
   getHostDockerInternalProbeFailure,
   isHijackedDockerInternalUrl,
 } = require("./onboard-host-docker-internal");
-const { isNvcfFunctionNotFoundForAccount, nvcfFunctionNotFoundMessage } = require("../validation");
+const {
+  isNvcfFunctionNotFoundForAccount,
+  nvcfFunctionNotFoundMessage,
+  shouldSkipResponsesProbe,
+} = require("../validation");
 const { isPrivateHostname, isPrivateIp, isLoopbackHostname } = require("../private-networks");
 const { buildResolvePinArgs, isOperatorTrustablePrivateIp } = require("./endpoint-ssrf-preflight");
 const {
@@ -57,6 +61,7 @@ const {
   STRICT_TOOL_PROBE_INITIAL_TOKENS,
   STRICT_TOOL_PROBE_RETRY_TOKEN_LADDER,
   strictToolProbeReasoningRetryMessage,
+  usesNvidiaEndpointProbePayload,
   vllmProbePolicyForModel,
 } = require("./openai-probe-models");
 const {
@@ -261,6 +266,16 @@ function getProbeAuthMode(_provider) {
   return undefined;
 }
 
+function getOpenAiSelectionProbeOptions(provider) {
+  return {
+    provider,
+    useNvidiaEndpointProbePayload: usesNvidiaEndpointProbePayload(provider),
+    requireResponsesToolCalling: shouldRequireResponsesToolCalling(provider),
+    skipResponsesProbe: shouldSkipResponsesProbe(provider),
+    authMode: getProbeAuthMode(provider),
+  };
+}
+
 export function getProbeExtraHeaders(provider) {
   if (provider === openrouter.OPENROUTER_PROVIDER_NAME) {
     return openrouter.getOpenRouterCurlHeaders();
@@ -535,6 +550,7 @@ export function getChatCompletionsProbeCurlArgs(opts: {
   isWsl?: boolean;
   pinnedAddresses?: readonly string[];
   validationTiming?: unknown;
+  useNvidiaEndpointProbePayload?: boolean;
 }) {
   const {
     credentialArgs,
@@ -544,6 +560,7 @@ export function getChatCompletionsProbeCurlArgs(opts: {
     isWsl: isWslOverride,
     pinnedAddresses,
     validationTiming,
+    useNvidiaEndpointProbePayload,
   } = opts;
   const platformOptions = getProbeTimingOptions({
     ...(typeof isWslOverride === "boolean" ? { isWsl: isWslOverride } : {}),
@@ -559,7 +576,7 @@ export function getChatCompletionsProbeCurlArgs(opts: {
     "Content-Type: application/json",
     ...credSlice,
     "-d",
-    JSON.stringify(getChatCompletionsProbePayload(model)),
+    JSON.stringify(getChatCompletionsProbePayload(model, { useNvidiaEndpointProbePayload })),
     url,
   ];
 }
@@ -573,6 +590,7 @@ function runChatCompletionsProbe({
   pinnedAddresses,
   trustedPrivateCapability,
   validationTiming,
+  useNvidiaEndpointProbePayload,
   spawnSyncImpl,
 }) {
   const args = getChatCompletionsProbeCurlArgs({
@@ -582,6 +600,7 @@ function runChatCompletionsProbe({
     isWsl: isWslOverride,
     pinnedAddresses,
     validationTiming,
+    useNvidiaEndpointProbePayload,
   });
   const probeOpts = {
     timeoutMs: getProbeProcessTimeoutMs(args),
@@ -621,7 +640,11 @@ function runDoubledTimeoutChatCompletionsRetry({
     "Content-Type: application/json",
     ...authConfig.args,
     "-d",
-    JSON.stringify(getChatCompletionsProbePayload(model)),
+    JSON.stringify(
+      getChatCompletionsProbePayload(model, {
+        useNvidiaEndpointProbePayload: options.useNvidiaEndpointProbePayload,
+      }),
+    ),
     `${baseUrl}/chat/completions`,
   ];
   const runRetryProbe = () =>
@@ -879,6 +902,7 @@ function probeOpenAiLikeEndpoint(endpointUrl, model, apiKey, options = {}) {
               pinnedAddresses,
               trustedPrivateCapability: options.trustedPrivateCapability,
               validationTiming,
+              useNvidiaEndpointProbePayload: options.useNvidiaEndpointProbePayload,
               spawnSyncImpl: options.spawnSyncImpl,
             }),
     };
@@ -1121,6 +1145,7 @@ module.exports = {
   hasChatCompletionsToolCallLeak,
   shouldRequireResponsesToolCalling,
   getProbeAuthMode,
+  getOpenAiSelectionProbeOptions,
   getProbeExtraHeaders,
   getValidationProbeCurlArgs,
   getDeepSeekV4ProValidationProbeCurlArgs,
@@ -1197,6 +1222,7 @@ export async function verifyOnboardInferenceSmoke(options: any, dependencies: an
     authMode: getProbeAuthMode(options.provider),
     extraHeaders: getProbeExtraHeaders(options.provider),
     skipResponsesProbe: true,
+    useNvidiaEndpointProbePayload: usesNvidiaEndpointProbePayload(options.provider),
     pinnedAddresses: options.pinnedAddresses,
     trustedPrivateCapability: options.trustedPrivateCapability,
   });

src/lib/inference/openai-probe-models.ts

@@ -9,6 +9,12 @@ export const STANDARD_NVIDIA_ENDPOINT_PROBE_POLICY =
 export const EXTENDED_NVIDIA_ENDPOINT_PROBE_POLICY =
   "nvidia.endpoint-validation.extended/v1";
 
+const NVIDIA_ENDPOINT_PROVIDERS = new Set(["nvidia-prod", "nvidia-nim"]);
+
+export function usesNvidiaEndpointProbePayload(provider: unknown): boolean {
+  return typeof provider === "string" && NVIDIA_ENDPOINT_PROVIDERS.has(provider);
+}
+
 export function vllmProbePolicyForModel(model: string): string {
   const normalized = model.trim().toLowerCase();
   const matches = loadManagedInferenceCatalog().models.filter(({ spec }) =>
@@ -47,7 +53,10 @@ export function isKimiK26Model(model: unknown): boolean {
   return String(model || "").toLowerCase() === "moonshotai/kimi-k2.6";
 }
 
-export function getChatCompletionsProbePayload(model: string): Record<string, unknown> {
+export function getChatCompletionsProbePayload(
+  model: string,
+  options: { useNvidiaEndpointProbePayload?: boolean } = {},
+): Record<string, unknown> {
   const maxTokensField = resolveMaxTokensField(model);
   const payload = {
     model,
@@ -74,6 +83,18 @@ export function getChatCompletionsProbePayload(model: string): Record<string, un
     };
   }
 
+  if (
+    options.useNvidiaEndpointProbePayload === true &&
+    model.toLowerCase() === "nvidia/nemotron-3-super-120b-a12b"
+  ) {
+    return {
+      ...payload,
+      temperature: 1,
+      top_p: 0.95,
+      chat_template_kwargs: { enable_thinking: false },
+    };
+  }
+
   return payload;
 }