Rm sort for decode gdr

[IzacharyI]

Motivation

Align SGLang with the AITER HIP GDN implementation after removal of the host-side index sorting path.
Related AITER PR: ROCm/aiter#4656

Technical Details

• Remove the HIP_GDN_SORT_IDX_BS environment override.
• Remove sorted-index cache reset loading and per-step reset calls.
• Remove obsolete sort-related tests and CI launch configuration.
• Continue using the unsorted HIP GDN decode path.

Requires AITER commit d74ae325d.

Test Plan

• Run the GDN backend and state-management unit tests.
• Validate CUDA Graph replay at batches 192, 256, and 512.
• Compare output and recurrent state against Triton.

Test Result

• GDN regression suite: 48 passed, 1 skipped.
• B192, B256, and B512 CUDA Graph replay passed.
• Output and recurrent state matched Triton.
• Syntax, lint, and diff checks passed.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[Copilot]

Pull request overview

Removes obsolete HIP GDN sorted-index handling following the corresponding AITER change.

Changes:

• Removes sort-cache loading, reset hooks, and calls.
• Removes the HIP_GDN_SORT_IDX_BS benchmark configuration.
• Deletes tests for retired cache-reset behavior.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
test/registered/attention/test_gdn_aiter_backend.py	Removes obsolete reset tests.
scripts/ci/sglang_benchmark_workflow.sh	Removes the retired environment override.
python/sglang/srt/layers/attention/linear/kernels/gdn_aiter.py	Removes sorted-index cache integration.
python/sglang/srt/layers/attention/linear/gdn_backend.py	Removes per-step cache resets.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.