[sgl-kernel][test] Skip dsv3_fused_a_gemm test on consumer Blackwell (sm120/sm121)

[layahaasini]

Motivation

On consumer Blackwell (sm120/sm121, e.g. GB10), test_dsv3_fused_a_gemm fails instead of skipping. The kernel requests ~192KB of dynamic shared memory, but consumer Blackwell caps at ~100KB per block, so the launch returns cudaErrorInvalidValue. Its only device guard is TORCH_CHECK(sm >= 90), which accepts sm12x. This is a hard architectural limit (not a missing dispatch), so the test should skip on these GPUs.

Modifications

• test_dsv3_fused_a_gemm.py: gate the test on is_sm90_or_sm100_supported so it skips on consumer Blackwell.

This PR was originally broader; the other two hunks were dropped after review:

• test_fp8_blockwise_moe.py — sm12x there is unimplemented, not unsupportable; the dispatch is being added in feat(moe): add SM120/SM121 dispatch for fp8_blockwise_scaled_grouped_mm #28125, so skipping would hide cases that PR turns green.
• test_flashmla.py — the flashmla prefill guard is handled by test(sgl-kernel): gate ue8m0 and flashmla prefill tests on supported architectures #29902.

Accuracy Tests

N/A — test-only change; no kernel/model-forward code or output changes.

Speed Tests and Profiling

N/A — same reason.

Checklist

• Format code with pre-commit
• Add unit tests — N/A: test-infrastructure only, no product code
• Update documentation — N/A: no public API / user-facing behavior change
• Provide accuracy and speed benchmark results — N/A
• Follow the SGLang code style guidance

Part of #31365.

---

CI States

Latest PR Test (Base): ❌ Run #29870545071
Latest PR Test (Extra): ❌ Run #29870544904

[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!

[waynehacking8]

One thing on the test_fp8_blockwise_moe.py hunk: sm12x isn't unsupportable there, it's just unimplemented, and #28125 adds the sm120/sm121 dispatch and turns exactly those 10 cases green - independently reproduced today on an RTX 6000D on top of my own RTX PRO 6000 runs. So narrowing is_blackwell_supported to == 10 would skip cases that are about to start passing.

I just backed the same change out of my #29902 for that reason, and dropped my is_blackwell_supported rename so we're not both touching that file. Our flashmla prefill guards still overlap - happy to drop mine if you'd rather carry it, since yours also covers test_dsv3_fused_a_gemm.py which I don't touch.

[layahaasini]

Thanks for pointing this out. I've dropped the fp8 hunk so it doesn't skip cases #28125 turns green.

On flashmla — since #29902 already gates it, I dropped my flashmla hunk too so we're not both touching that file. Happy for #29902 to carry it.

That leaves this PR as just the test_dsv3_fused_a_gemm.py guard, which doesn't overlap.

[layahaasini]

Closing: the AOT test this patched was removed in #30280, and the JIT replacement works on sm121 (verified on a GB10).