[FLINK-40684][model] Fix tensor name case in Triton models.

[damjad]

What is the purpose of the change

Triton Inference Server's KServe V2 protocol treats tensor names as case-sensitive. A model
configured with input / output tensors in its config.pbtxt rejects requests that send
INPUT / OUTPUT, returning HTTP 400 unexpected inference output 'OUTPUT'.

TritonInferenceModelFunction.buildInferenceRequest() unconditionally uppercased both tensor names:

// Before (buggy):
inputNode.put("name", inputName.toUpperCase());
outputNode.put("name", outputName.toUpperCase());

This caused every request to a model with lowercase tensor names to fail at the network layer.

Brief change log

• Remove .toUpperCase() from the input tensor name assignment in buildInferenceRequest()
• Remove .toUpperCase() from the output tensor name assignment in buildInferenceRequest()
• Add TritonTensorNameCasingTest — two MockWebServer-backed integration tests that capture the outgoing HTTP request and assert the name fields in inputs and outputs match the column names declared in the model schema without any case transformation

Verifying this change

This change added tests and can be verified as follows:

• TritonTensorNameCasingTest#testLowercaseTensorNamesArePreserved — columns named input/output produce "name":"input" / "name":"output" in the JSON body
• TritonTensorNameCasingTest#testMixedCaseTensorNamesArePreserved — columns named myInput/myOutput are also sent unchanged
• Tests fail when .toUpperCase() is re-introduced (expected: "input" but was: "INPUT")
• Manually verified against a live Triton endpoint (torch-model with lowercase tensors): HTTP 400 → HTTP 200

Does this pull request potentially affect one of the following parts:

• Dependencies (does it add or upgrade a dependency): no
• The public API, i.e., is any changed class annotated with @public(Evolving): no
• The serializers: no
• The runtime per-record code paths (performance sensitive): no
• Anything that affects how connectors interact with Flink: no
• The SQL / Table API: no
• Core ML / Model inference (flink-model-triton): yes

Documentation

No public API or configuration changes. The fix aligns behaviour with the KServe V2 specification,
which defines tensor names as opaque strings whose casing is determined by the model's
config.pbtxt

---

Was generative AI tooling used to co-author this PR?

• Yes (please specify the tool below)

Generated-by: Claude Sonnet 4.6 noreply@anthropic.com

[flinkbot]

CI report:

• eabadbc Azure: SUCCESS

Bot commands

The @flinkbot bot supports the following commands:

• @flinkbot run azure re-run the last Azure build

[spuru9]

@damjad One of the commits has Coauthor-by: AI tag can you squash the commits and add disclosure as per AI guidelines.

[damjad]

@spuru9 updated.

[damjad]

@featzhang tagging you for visibility.