[AMD] Fix DeepSeekV4 server cutlass error

[At1a8]

Motivation

PR #25220 introduced a top-level import of the CuTeDSL MQA logits module, which depends on NVIDIA-only cutlass packages.

On AMD/ROCm environments without cutlass, this import fails during DSV4 model registration and prevents the server from starting, even though the CuTeDSL backend is not used.

Modifications

Guard the CuTeDSL imports behind not is_hip(). On non-ROCm platforms, preserve the original eager import behavior so NVIDIA/CuTeDSL logic remains unchanged.

Accuracy Tests

root@smci355-ccs-aus-n12-13:/sgl-workspace/sglang# python3 /sgl-workspace/sglang/benchmark/gsm8k/bench_sglang.py --port 8000 --num-questions 1319

100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1319/1319 [01:08<00:00, 19.29it/s]
Accuracy: 0.948
Invalid: 0.000
Latency: 68.365 s
Output throughput: 1728.667 token/s

Speed Tests and Profiling

This PR has almost no impact on performance.

TP8 DP8, 8k/1k, concurrency = 256, num prompts = 4 × concurrency	Total token throughput (tok/s)	Mean TTFT (ms)	Mean TPOT (ms)
Main branch	failed	failed	failed
This PR	34904.71	6871.52	57.17

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 #28877915192
Latest PR Test (Extra): ❌ Run #28877913869

[gemini-code-assist]

Code Review

This pull request introduces conditional imports for pick_dsl_expand and CuteDSLPagedMQALogitsRunner to avoid eager loading on ROCm (HIP) platforms. The review feedback notes that on ROCm platforms, pick_dsl_expand is left completely undefined in the module namespace, which will trigger static analysis warnings. It is recommended to define a fallback value of None when _is_hip is True to resolve this issue.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

When _is_hip is True, pick_dsl_expand is not imported and remains completely undefined in the module namespace. Although the code path that uses it (use_cute_dsl) is not reachable on ROCm, static analysis tools (such as Pyright, mypy, or flake8) will still flag pick_dsl_expand as an undefined name.\n\nDefining a fallback value of None when _is_hip is True resolves this static analysis issue.

if not _is_hip:\n    # Preserve the original eager import behavior on non-ROCm platforms.\n    from sglang.jit_kernel.dsa import pick_dsl_expand\nelse:\n    pick_dsl_expand = None

[kkHuang-amd]

Please check this review comment and fix that

[At1a8]

Please check this review comment and fix that

Thanks for your comments! I've addressed the issue based on your suggestions.

[b8zhong]

Thanks. Sorry about the breaking

[At1a8]

DeepSeek V4 passed in local terminal.