[bingxche]

Motivation

The MI35x stage-b test fails during collection because test_triton_mla_prefill_gfx950.py still imports the deleted extend_attention_split_dim module. The dedicated split-dim kernel and its dispatch predicate were removed in 4e2fd9f1e9; the general extend_attention_fwd path now serves these Kimi-K3 shapes. The leftover import prevents all live numerical tests in this file from running.

Failing job: https://github.com/sgl-project/sglang/actions/runs/34070685413/job/101587367463

Modifications

• Remove the stale import and the dispatch-gate test for the deleted split-dim implementation.
• Keep the three numerical tests for the current extend_attention_fwd path: BF16 ragged prefill, zero-prefix FP8 prefill, and absorbed FP8 cached-prefix prefill.

This is a test-only change. It does not change kernels, runtime dispatch, accuracy thresholds, or model behavior.

Accuracy Tests

No model behavior changes.

Speed Tests and Profiling

Not applicable. Runtime code is unchanged.

Checklist

• Format your code according to the contribution guide.
• Add or update tests according to the contribution guide, or explain why not applicable.
• Update documentation according to the contribution guide, or explain why not applicable.
• Provide accuracy and speed benchmark results, or explain why not applicable.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #34090399165
Latest PR Test (Extra): ❌ Run #34090398694
Latest PR Test (AMD ROCm 7.2): ❌ Run #34090399112