[ROCm][Perf][GLM-5.3-Flash] Carry vllm#58584 rebased onto ac9126e5

[amd-sriram]

Purpose

Carry of vllm-project/vllm#58584 (BF16 split-K kernel for GLM-5.3-Flash sparse MLA decode, by @simondanielsson) rebased onto ac9126e58aa7bbab1856ba6593ba4d5003fea516, so it can be built on that nightly.

The upstream PR no longer applies there: its branch merged main after vllm-project#58569 landed, its early commits predate vllm-project#58569, and its last commit (test: tighten tests) edits tests that only exist after vllm-project#58569.

Changes

• The PR's own nine commits, cherry-picked in order. The first eight apply unchanged (the same eight an earlier ac9126e5 build carried).
• test: tighten tests: one conflict in tests/kernels/attention/test_rocm_triton_attn_dsv4.py, resolved to the commit's intent (drop test_sparse_attn_decode_bf16_writes_caller_output, narrow num_splits) without importing the [ROCm][Perf][GLM-5.3-Flash] Remove redundant copy after ragged sparse MLA vllm-project/vllm#58569 prefill tests that the upstream side of the conflict carried.

No source change beyond vllm-project#58584. Close this once a nightly containing vllm-project#58569 is used, or vllm-project#58584 merges.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 0342db85-00c6-414f-b1da-2acf0bd32f2b

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Autopilot is currently an internal CodeRabbit preview.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}