fix: enforce inference restriction in SageMaker MME invoke handler

[Vinya567]

What does the PR do?

Adds the missing --http-restricted-api=inference enforcement to the SageMaker MME /invoke handler. SagemakerAPIServer::SageMakerMMEHandleInfer was not calling the RETURN_AND_RESPOND_IF_RESTRICTED macro, so the four sibling SageMaker control endpoints (LIST/GET/LOAD/UNLOAD) and the regular HTTP /v2/<m>/infer path enforced the configured header while MME /invoke silently accepted requests without it. Adds the macro call at the head of the handler and a regression block in qa/L0_sagemaker/test.sh.

Checklist

• PR title reflects the change and is of format <type>: <description>
• Changes are described in the pull request.
• Related issues are referenced.
• Populated github labels field
• Added test plan and verified test passes.
• Verified that the PR passes existing CI.
• Verified copyright is correct on all changed files.
• Added succinct git squash message before merging
• All template sections are filled out.

Commit Type:

• fix

Related PRs:

N/A

Where should the reviewer start?

src/sagemaker_server.cc — one new line at the head of SageMakerMMEHandleInfer:

RETURN_AND_RESPOND_IF_RESTRICTED(req, RestrictedCategory::INFERENCE);

This matches the four existing call-sites in the same file (LIST/GET/LOAD/UNLOAD, all gated on RestrictedCategory::MODEL_REPOSITORY) and HTTPAPIServer::HandleInfer (gated on RestrictedCategory::INFERENCE).

qa/L0_sagemaker/test.sh — new block that launches Triton with --http-restricted-api=inference:X-Auth=secret and exercises POST /models/<m>/invoke and POST /invocations with and without the X-Auth header. Expects 403 without and 200 with.

Test plan:

Reproduced the BEFORE behavior on nvcr.io/nvidia/tritonserver:26.05-py3: POST /models/<m>/invoke without X-Auth returned 200 OK despite --http-restricted-api=inference:X-Auth=secret. After the fix the same request returns 403 Forbidden with {"error":"This API is restricted, expecting header 'X-Auth'"}.

The new regression block emitted the expected codes on internal CI:

POST /models/sm_mme_model/invoke   (no X-Auth)   ->  403
POST /models/sm_mme_model/invoke   (X-Auth)      ->  200
POST /invocations                  (no X-Auth)   ->  403
POST /invocations                  (X-Auth)      ->  200

• L0_sagemaker--base: passed (existing tests + new block)
• L0_http--base: passed (existing http_restricted_api_test.py confirms the RETURN_AND_RESPOND_IF_RESTRICTED macro is unchanged on non-SageMaker paths)
• CI Pipeline ID: 54470604

Caveats:

None.

Background

SagemakerAPIServer extends HTTPAPIServer and shares the RestrictedFeatures map via main.cc (which copies http_restricted_apis_ into the SagemakerAPIServer constructor). The four sibling SageMaker control endpoints (LIST/GET/LOAD/UNLOAD) gate on RestrictedCategory::MODEL_REPOSITORY; the regular HTTP /v2/<m>/infer path gates on RestrictedCategory::INFERENCE. The MME /invoke handler — which performs inference — was missing the gate. This PR adds the fifth call-site at the head of SageMakerMMEHandleInfer (before any model resolution or dispatch) so behavior is consistent across all HTTP inference paths.

Related Issues: (use Closes / Fixes / Resolves / Relates to)

N/A

[whoisj]

LGTM