Forward HTTP headers to /generate endpoint (Fixes #8790)

[92subrata]

What

--http-header-forward-pattern forwarded matching HTTP headers to the
backend via request.parameters() on the /infer endpoint, but
HandleGenerate() never called ForwardHeaders(). As a result the flag
had no effect on the /generate endpoint (including the
ensemble → Python backend path) and headers were silently dropped.

Fix

Add the missing ForwardHeaders(req, irequest) call in
HandleGenerate(), matching the placement used by /infer, so header
forwarding behaves consistently across both endpoints.

Test

Adds test_http_header_forward_pattern_generate in qa/L0_parameters,
POSTing to /generate on both the parameter model and the ensemble
(identity → Python backend), asserting the forwarded headers come back
as request parameters. Wired into test.sh with
--http-header-forward-pattern MY_HEADER.*.

Existing test methods are untouched — the change is purely additive.

Fixes #8790

[92subrata]

Hi maintainers — this is a small, additive fix + regression test for #8790 (--http-header-forward-pattern was silently dropped on the /generate endpoint, including the ensemble → Python backend path). Existing tests are untouched. Would really appreciate a review whenever you have a moment. Happy to address any feedback. Thanks!

[greptile-apps]

Greptile Summary

The PR makes configured HTTP header forwarding apply to the /generate endpoint and adds integration coverage for direct and ensemble-backed generation.

• Calls ForwardHeaders while constructing generate inference requests.
• Adds raw HTTP /generate checks for matching header names, values, and associations.
• Runs the new regression test with --http-header-forward-pattern.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
src/http_server.cc	Applies the existing configured header-forwarding behavior to generated inference requests before dispatch.
qa/L0_parameters/parameters_test.py	Adds direct and ensemble /generate regression coverage and correctly verifies header-to-value associations.
qa/L0_parameters/test.sh	Registers the new test and starts its server with the required HTTP header-forwarding pattern.

_{Reviews (5): Last reviewed commit: "Merge branch 'main' into fix/generate-he..." | Re-trigger Greptile}

[92subrata]

@yinggeh @Vinya567 — could you please review this PR when you get a chance? Thanks!

[Vinya567]

@yinggeh @Vinya567 — could you please review this PR when you get a chance? Thanks!

Hi @92subrata , thanks for the PR.

The change looks good: HandleGenerate was missing the ForwardHeaders call that /infer already has, and the L0_parameters regression covers the /generate path (including ensemble → Python).

One prerequisite : Triton requires all contributors (or their corporate entity) to have a signed Contributor License Agreement on file with triton-cla@nvidia.com.