[https://nvbugs/6485390][fix] Fix Qwen disagg throughput regression by restoring MTP forward hook

[trtllm-agent]

Summary

• Root cause: The MTPEagleDynamicTreeWorker.forward method was overriding the parent MTPEagleWorker.forward, which bypassed the standard entry-point wrapping (NVTX/profiling and dispatch hooks) used by the base worker path. This caused the dynamic-tree MTP worker to take a slower code path on Qwen disaggregated runs, producing a measurable Total_Token_Throughput regression versus 1.3.0rc15. Additionally, the disaggregated GPT-OSS perf-sanity YAML pinned transceiver_runtime: PYTHON, forcing the slower Python NIXL transceiver instead of the optimized default.
• Fix: Rename the override to _forward_impl so the parent class's forward remains the single entry point and correctly dispatches into the dynamic-tree implementation, restoring the fast path used prior to rc15. The two perf-sanity YAMLs were updated to drop the transceiver_runtime: PYTHON override so the default (faster) transceiver runtime is used, aligning benchmark configuration with production defaults.
• Perf metric: total_token_throughput (higher-is-better)
• Bad perf: 9.517e+04
• Good perf: 1.052e+05
• Perf after fix: 1.043e+05
• Automated fix generated by repair-bot

Test plan

• Verify fix on the same GPU type as the original failure
• Check for regressions in related tests

Links

• Bug: https://nvbugs/6485390

Dev Engineer Review

• Renamed MTPEagleDynamicTreeWorker.forward to _forward_impl without changing its signature or implementation, preserving the parent worker’s standard forward entry point and associated hooks.
• Restores the optimized dynamic-tree MTP execution path while retaining existing verification, KV relocation, metadata restoration, and result construction behavior.
• Removed explicit transceiver_runtime: PYTHON overrides from the GPT-OSS performance-sanity configurations so they use the default optimized runtime.
• Changes are scoped and consistent; no test-list files were modified.

QA Engineer Review

No test changes.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 6bb80c5b-8d1c-4544-956e-edba49f556d6

📥 Commits

Reviewing files that changed from the base of the PR and between 45d1c11 and 8265424.

📒 Files selected for processing (2)

• tensorrt_llm/_torch/speculative/mtp_dynamic_tree.py
• tests/scripts/perf-sanity/disaggregated/gb200_gpt-oss-120b-fp4_1k1k_con2048_ctx1_tp1_gen1_dep2_eplb0_mtp0_ccb-NIXL.yaml

💤 Files with no reviewable changes (1)

• tests/scripts/perf-sanity/disaggregated/gb200_gpt-oss-120b-fp4_1k1k_con2048_ctx1_tp1_gen1_dep2_eplb0_mtp0_ccb-NIXL.yaml

---

Walkthrough

The dynamic-tree worker entry method is renamed to _forward_impl without changing its signature or body. Two disaggregated NIXL cache transceiver configurations remove the PYTHON runtime setting while retaining their existing backend and buffer values.

Changes

Dynamic tree worker entrypoint

Layer / File(s)	Summary
Rename dynamic-tree worker entrypoint tensorrt_llm/_torch/speculative/mtp_dynamic_tree.py	MTPEagleDynamicTreeWorker.forward is renamed to _forward_impl with the same parameters and implementation.

NIXL cache transceiver configuration

Layer / File(s)	Summary
Remove Python runtime selection tests/scripts/perf-sanity/disaggregated/gb200_gpt-oss-120b-fp4_1k1k_con2048_ctx1_tp1_gen1_dep2_eplb0_mtp0_ccb-NIXL.yaml	The generation and context cache transceiver blocks remove transceiver_runtime: PYTHON while retaining backend: NIXL and max_tokens_in_buffer: 1024.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related PRs

• NVIDIA/TensorRT-LLM#16724: Renames the same dynamic-tree worker method.
• NVIDIA/TensorRT-LLM#16727: Applies the same method rename in the same file.
• NVIDIA/TensorRT-LLM#16728: Matches the same signature-preserving method rename.

Suggested reviewers: sunnyqgg

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title matches the main fix and follows the requested [ticket][type] summary format.
Description check	✅ Passed	The description includes the problem, fix, perf data, test plan, and bug link, though it uses different section headings than the template.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}

[Shixiaowei02]

@nv-xtf Please pay attention to the GPT-OSS perf regression case here. Thanks!