[plugin][distributed] use active platform's backend in get_default_distributed_backend

[AgainstEntropy]

Motivation

Follow-up to #21388.

get_default_distributed_backend(device) currently maps device to torch distributed backend via a hard-coded _DEVICE_TO_DISTRIBUTED_BACKEND dict (cuda → nccl, cpu → gloo, …). Out-of-tree platform plugins currently have to patch the dict for new backends.

The platform interface already declares get_torch_distributed_backend_str() as the source of truth for a platform's torch.distributed backend, and the same pattern has already been used in multimodal_gen runtime when initializing process groups (multimodal_gen/runtime/distributed/parallel_state.py).

This PR aligns the SRT distributed init with that pattern while keeping the dict as a safe fallback.

Modifications

python/sglang/srt/distributed/parallel_state.py — get_default_distributed_backend(device) now, only when device == current_platform.device_type, calls current_platform.get_torch_distributed_backend_str() first:

• success → return the platform-supplied backend
• NotImplementedError → silently fall through to the dict (today's in-tree behavior is preserved: the base method in device_mixin.py is raise NotImplementedError and no in-tree platform overrides it yet)
• other exception → log a warning, fall through

The device == current_platform.device_type guard preserves correctness for callers that ask for a non-active device's backend (e.g. a cpu auxiliary gloo group on a CUDA process); those keep going through the dict.

current_platform is imported lazily inside the function to avoid a circular import with sglang.srt.platforms.

Tests

test/registered/unit/distributed/test_parallel_state.py covers all paths through get_default_distributed_backend using real SRTPlatform subclasses.

Registered for stage-a-test-cpu:

python test/registered/unit/distributed/test_parallel_state.py

.....
----------------------------------------------------------------------
Ran 5 tests in 0.001s

OK

Also have run some tp/dp manual tests to confirm the correctness.

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!

[AgainstEntropy]

/tag-and-rerun-ci , try again

[AgainstEntropy]

/rerun-failed-ci, try again, and again, and again