Fix/hicache sidecar ok pure mla

[Leoyzen]

Motivation

When a pure-MLA model (e.g. DeepSeek-V4) runs with TP>1, follower TP ranks have no rank-sharded sidecars to back up — all pools are replicated and TP0 handles the writes. The else branch of _page_backup set sidecar_ok = bool(backup_transfers), which evaluated to False when backup_transfers was empty, yielding completed_tokens=0. This prevented L3 restore from ever triggering on follower ranks, causing NaN on DSpark speculative hits (issue #33656).

Modifications

python/sglang/srt/mem_cache/hybrid_cache/hybrid_cache_controller.py — _page_backup() method, else branch (backup_skip=True, i.e. follower TP ranks):

Changed sidecar_ok = bool(backup_transfers) to sidecar_ok = True. When there are no rank-sharded sidecars, the follower rank's backup is vacuously successful — TP0 already persisted all replicated pools. The validation loop still runs for non-empty backup_transfers and can set sidecar_ok=False on real failures.

# Before (buggy):
sidecar_ok = bool(backup_transfers)  # empty → False → completed_tokens=0 → no restore → NaN

# After (fixed):
sidecar_ok = True  # vacuously successful; validation loop still runs and can set False

test/registered/unit/mem_cache/test_hybrid_cache_controller_page_backup.py — 3 new unit tests:

1. test_pure_mla_follower_reports_success_with_no_sidecars — Bug regression: follower rank with empty backup_transfers → completed_tokens > 0 (fails pre-fix, passes post-fix)
2. test_follower_sidecar_write_failure_reports_zero — Negative: non-empty backup_transfers with failed writes → completed_tokens = 0
3. test_follower_sidecar_write_success_reports_full — Positive: non-empty backup_transfers with successful writes → completed_tokens = full

Registered as base-a-test-cpu (est_time=3s). All tests pass locally on CPU.

Accuracy Tests

N/A — this change does not affect model outputs. It fixes the HiCache L3 backup/restore coordination path for pure-MLA models (DSV4 etc.) under TP>1.

Speed Tests and Profiling

N/A — no inference speed impact. The fix eliminates spurious backup-failure reporting on follower ranks when all pools are already replicated.

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.

Related #33656

---

CI States

Latest PR Test (Base): ❌ Run #31193950878
Latest PR Test (Extra): ❌ Run #31193947131

[1e4ves]

completed_tokens tracks local backup work. Followers with no sidecars correctly report 0 but still ACK; restore uses a separate prefetch operation. So this cannot block restore or cause NaNs.

But in the merged #30393, dsv4 mtp and dsv4 dspark should be in the packed path. Only dsv4+eagle3 uses the DRAFT sidecar, and its backup_skip is also fixed by should_backup() . Can you retest it ? Maybe you need also flush mooncake, because of the packed 3 dspark layers. @Leoyzen