[HiCache] Reset decode ReplaySSM ring cursor on load-back Mamba slots

[rodamani]

Motivation

With decode ReplaySSM (--enable-linear-replayssm) and HiCache, MambaComponent.prepare_load_back allocates the request's device Mamba slot directly from the allocator. That bypasses HybridReqToTokenPool.alloc, which is where a fresh slot's replayssm_write_pos is reset; the request already holds a slot by the time alloc runs, so the reset branch is skipped. The H2D load-back writes only temporal + conv state (the host tier carries no ring state).

Cursors are reset on release only by the finish-and-insert path. Retract and abort release with is_insert=False and free the slot without touching write_pos. So a load-back can land in a slot whose previous tenant was retracted or aborted mid-decode, and the decode kernel then replays that tenant's ring on top of the loaded checkpoint (cross-request recurrent-state bleed). Cache eviction is not a source: tree-held slots enter the tree with cursor 0.

Speculative ReplaySSM (--enable-linear-replayssm-spec) keeps its cursors per request row and is not affected.

Modifications

• Reset replayssm_write_pos on the freshly allocated load-back slot in prepare_load_back, mirroring the fresh-slot reset in HybridReqToTokenPool.alloc. No-op when decode ReplaySSM is off.
• test/registered/unit/mem_cache/test_mamba_load_back_replayssm_reset.py.

Relationship to open PRs: #36345 resets write_pos only on its new buffer-only load-back path; #37837 resets it on the extra_buffer donate path. Neither covers prepare_load_back.

Accuracy Tests

CPU: 2 passed. With the reset removed: test_load_back_slot_starts_with_empty_ring fails (cursor stays nonzero). The test seeds the stale cursor directly; that retract/abort leave one behind is from reading the release paths. Not run end-to-end on GPU.

Speed Tests and Profiling

No hot-path change beyond the fix itself; not separately benchmarked.

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 #36767714730
Latest PR Test (Extra): ❌ Run #36767714122
Latest PR Test (AMD ROCm 10): ❌ Run #36767714748

[rodamani]

/rerun-test -c test_retraction_mamba_backup.py test_mamba_unittest.py test_linear_replayssm_decode.py

[github-actions]

Results for /rerun-test -c test_retraction_mamba_backup.py test_mamba_unittest.py test_linear_replayssm_decode.py:

🚀 ubuntu-latest (1 test): ✅ View workflow run

cd test/ && python3 registered/unit/mem_cache/test_retraction_mamba_backup.py

🚀 1-gpu-5090 (1 test): ✅ View workflow run

cd test/ && python3 registered/unit/mem_cache/test_mamba_unittest.py

🚀 1-gpu-h100 (1 test): ✅ View workflow run

cd test/ && python3 registered/attention/unittests/gdn/test_linear_replayssm_decode.py

⛔ test/registered/unit/mem_cache/test_mamba_load_back_replayssm_reset.py: File not found: test/registered/unit/mem_cache/test_mamba_load_back_replayssm_reset.py

[rodamani]

/tag-and-rerun-ci