[Sglang] Update P2P RDMA weight sync with flashinfer compatibility and NIC affinity

[wduan-hai]

Motivation

Update P2P RDMA weight sync with flashinfer compatibility and NIC affinity. This works in tandem with radixark/miles#964 and is miles 964's dependency.

Modifications

NIC affinity: we allow a manual json flag --remote-instance-weight-loader-ib-device to be passed from miles to sglang, so each GPU can find its NIC during the P2P RDMA weight sync.

Flashinfer compatibility: The P2P RDMA sync is 3 steps, first pre process, then weight transfer, and then post process. We add a trigger in pre process such that the 4D tensor of gate_up is transformed to the 3D format before the weight transfer, to be consistent with the format from the miles side. After the weight transfer, the post process will shape it back. Without the reshape in pre processing, the check equality fails.

Accuracy Tests

Tested on 4 node with Qwen3-30B-A3B, 16 node with Qwen3-235B-A22B over RoCE.

Speed Tests and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!