[AMD] Fix DSV4 JIT build on rocm

[yctseng0211]

Motivation

The Jul-14 scheduled Nightly Test (AMD MI355X 2N 1P1D Disagg) (run 29305625618) failed on every DeepSeek-V4 config (flash/pro × fp8/fp4 × base/dp8ep8/mtp) at scheduler init:

utils.cuh:245: error: use of undeclared identifier 'cudaDevAttrComputeCapabilityMajor' → ninja exited with status 1 → Rank 0 scheduler died during initialization (exit code: -3)

#30438 added getSMVersion() to jit_kernel/include/sgl_kernel/utils.cuh using the CUDA-only enums cudaDevAttrComputeCapabilityMajor/Minor. The JIT path (tvm-ffi + hipcc) does not run hipify — unlike the compiled sgl-kernel, whose ROCm build hipifies cudaDevAttr* → hipDeviceAttribute* — so on gfx950 those identifiers are undeclared, the DeepSeek-V4 compress-plan JIT fails to build, and the scheduler dies. (Kimi-K2.6 doesn't build this JIT, so it was unaffected.)

Modifications

Adopting the reviewed approach from #31141 (thanks @kangwangamd, @DarkSharpness): instead of guarding the CUDA-only calls with #ifndef USE_ROCM, move getSMVersion out of utils.cuh into runtime.cuh, where the cuda*→hip* attribute shims already live — so it compiles on both CUDA and HIP with no guard.

• Add the cudaDevAttrComputeCapabilityMinor → hipDeviceAttributeComputeCapabilityMinor #define (the Major one was already present).
• Add get_cc_minor() mirroring the existing get_cc_major(), and express get_sm_version(id) = get_cc_major(id) * 10 + get_cc_minor(id) to deduplicate the compute-capability queries.
• Remove getSMVersion from utils.cuh. It has no caller in the JIT tree (its only reference was its own definition; SM120 gating is done in Python via is_sm120_supported()), so there is no use-site change. No CUDA-path behavior change.

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.

---

CI States

Latest PR Test (Base): ❌ Run #29352408580
Latest PR Test (Extra): ✅ Run #29384613963

[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!

[yctseng0211]

Verification run triggered: run 29313160939 on this branch, config dsv4pro-fp8-1k1k-1p1d — the exact config that failed at init in the Jul-14 scheduled nightly (run 29305625618, all DSV4 configs). It should now compile the DeepSeek-V4 compress-plan JIT on ROCm/gfx950 and the scheduler should start.

[yctseng0211]

https://github.com/sgl-project/sglang/actions/runs/29316462484
AMD CI passed

[DarkSharpness]

sorry i somehow missed this PR (I review the PR from latest to this...). might be a duplicate of #31141

[yctseng0211]

https://github.com/sgl-project/sglang/actions/runs/29320689120

[yctseng0211]

@DarkSharpness Sure, we didn't coordinate on this upfront. I've since synced with @kangwangamd (author of #31141).

The reason I kept this as a separate PR: the MI355X 2N 1P1D disaggregation nightly only runs on a branch inside sgl-project/sglang (the workflow is gated on github.repository == 'sgl-project/sglang'), so we need an upstream branch to actually drive the hardware tests — a fork branch can't trigger it.

So I've brought the reviewed changes from #31141 over to this branch (move getSMVersion into runtime.cuh as get_sm_version, dedup with get_cc_major / get_cc_minor, drop the USE_ROCM guard) and added @kangwangamd as co-author. Happy to close whichever PR you'd prefer to keep.

[yctseng0211]

https://github.com/sgl-project/sglang/actions/runs/29320689120
result verified
cc: @DarkSharpness @b8zhong @BBuf

[b8zhong]

/rerun-failed-ci

[yctseng0211]

PR Test Base is finished, the only one failure is https://github.com/sgl-project/sglang/actions/runs/29352408580/job/87255423678?pr=31131#logs
which is a known issue

irrelevant to this PR