Description

Qwen3.5-VL uses string chat content, but its registration did not opt into the serving path that preserves ordered OpenAI content parts. Requests containing alternating text and images therefore bulk-prepended all vision placeholders, changing prompt tokenization and multimodal positions.

Enable the existing interleaved-placeholder path for the shared dense and MoE metadata. Requests without ordered content parts keep the existing BEFORE_TEXT fallback.

This addresses the placeholder-ordering part of #16792. The separate residual generation parity case described there remains out of scope.

Test Coverage

• Extended the dense and MoE placeholder-registration tests to cover the interleaving contract.
• pre-commit run --files tensorrt_llm/_torch/models/modeling_qwen3_5.py tests/unittest/_torch/modeling/test_modeling_qwen3_5_vl.py tests/unittest/_torch/modeling/test_modeling_qwen3_5_vl_moe.py
• python3 -m compileall -q tensorrt_llm/_torch/models/modeling_qwen3_5.py tests/unittest/_torch/modeling/test_modeling_qwen3_5_vl.py tests/unittest/_torch/modeling/test_modeling_qwen3_5_vl_moe.py

PR Checklist

• Reviewed the checklist in the pull request template as applicable.

Dev Engineer Review

• Enabled interleaved placeholder registration for shared Qwen3.5-VL dense and MoE metadata.
• Configuration is consistent with the intended OpenAI content-part ordering behavior and preserves the existing fallback path.
• Added metadata assertions for both variants. No apparent API, performance, or error-handling regressions.
• Residual generation parity remains out of scope.

QA Engineer Review

• Modified placeholder-metadata tests for qwen3_5 and qwen3_5_moe, adding assertions that interleave_placeholders is enabled.
• No corresponding test-db/ or qa/ coverage entries were identified.
• Verdict: needs follow-up — CI/manual coverage data is unavailable.