Fix fused DeepSeek top-k capability fallback

[nv-tusharma]

Motivation

SGLang currently selects FlashInfer's fused_topk_deepseek based on shape constraints without checking whether the active GPU architecture is supported by that kernel. On A10G (SM86), FlashInfer raises BackendSupportedError during DeepSeek grouped top-k routing instead of allowing SGLang to use its existing fallback.

This was observed in the Dynamo nightly SGLang data-parallel test: https://github.com/ai-dynamo/dynamo/actions/runs/32114902544/job/95672470424

Modifications

• Read FlashInfer's is_compute_capability_supported metadata before selecting fused_topk_deepseek.
• Preserve compatibility with FlashInfer versions that do not expose capability metadata.

Accuracy Tests

The numerical implementation is unchanged. Unsupported devices now use the existing SGLang fallback instead of entering a kernel that raises before producing output.

Speed Tests and Profiling

Not applicable. Supported devices retain the fused path; the added check is one device-capability lookup during top-k dispatch.

Validation

• python3 -m compileall -q python/sglang/srt/layers/moe/topk.py test/registered/kernels/ops/moe/test_fused_topk_deepseek.py
• GPU pytest not run locally because the available environment does not include PyTorch/CUDA.

Checklist

• Format code according to the project guidance.
• Keep the change limited to the capability-aware dispatch guard; no new unit test is needed.
• No documentation update is needed for this internal dispatch fix.
• Accuracy and performance impact are described above.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #32763510057
Latest PR Test (Extra): ❌ Run #32763509769
Latest PR Test (AMD ROCm 7.2): ❌ Run #32763509983

[nv-tusharma]

/tag-and-rerun-ci