[None][fix] Pass dtype to AllReduce ctor to enable MNNVL all-reduce fo…

[nv-guomingz]

…r Qwen3.5

On NVL multi-node systems, AllReduce must be given dtype at construction so it can build the MNNVL all-reduce path (its Lamport workspace is sized by the dtype's element size). If dtype is omitted, mnnvl_allreduce is None, so the op falls back to the generic NCCL all-reduce across nodes — functionally correct but lower performance than the NVLink-fabric MNNVL path.

[nv-guomingz]

/bot run --disable-fail-fast

[coderabbitai]

📝 Walkthrough

Walkthrough

In modeling_qwen3_next.py, three AllReduce constructor calls are updated to pass dtype=config.torch_dtype explicitly. The affected constructors are in Qwen3NextSparseMoeBlock, Qwen3NextLinearDecoderLayer, and Qwen3NextFullAttentionDecoderLayer.

Changes

AllReduce dtype propagation in Qwen3Next

Layer / File(s)	Summary
Pass dtype to AllReduce in all three decoder components tensorrt_llm/_torch/models/modeling_qwen3_next.py	AllReduce construction in Qwen3NextSparseMoeBlock (line 128), Qwen3NextLinearDecoderLayer (line 350), and Qwen3NextFullAttentionDecoderLayer (line 515) now each supply dtype=config.torch_dtype; previously dtype was omitted from all three calls.

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~2 minutes

🚥 Pre-merge checks | ✅ 3 | ❌ 2

❌ Failed checks (1 warning, 1 inconclusive)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The pull request description provides sufficient context about the technical rationale for the changes but lacks required sections from the template.	Add missing template sections: provide a clear short description, explicitly list test coverage validation, and complete the PR checklist items required for this repository.
Title check	❓ Inconclusive	The title is truncated and appears incomplete ('fo…' suggests the text was cut off). While it references the main change (passing dtype to AllReduce), the truncation makes it unclear and prevents full assessment of clarity.	Complete the pull request title to fully convey the change. A complete title might be: '[fix] Pass dtype to AllReduce ctor to enable MNNVL all-reduce for Qwen3' or similar, ensuring all key information is visible.

✅ Passed checks (3 passed)

Check name	Status	Explanation
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.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[tensorrt-cicd]

PR_Github #55247 [ run ] triggered by Bot. Commit: cc449b3 Link to invocation