[shipiyouniao]

Backs native DeepSeek-V4 C128 request state on SGLang 0.5.20 using native alloc_rows/free_rows, including online banks. Native row selection, binding and reuse remain intact. Misses roll back backing; idle shrink releases unpromised standby rows while preserving prefix cache, active state and sentinel pages.

Stacked on #568 and its dependencies; one C128 state commit. Dynamic state requires managed SWA, TP=1 and PP=1, supported non-unified layouts and sufficient pages. Unsupported topology, HIP C128, preemption, disaggregation and unsupported request pools retain native storage/accounting. Admission follows native batch and beam limits and reserves the largest backed native row tail, including pending members. Other capacity queries retain native logical availability; scheduling cost is not measured.

All selected backing managers finish initialization before any resize or reclamation starts. This includes the exact-ID readiness and HiSparse manager-discovery corrections from #570. Regression tests hold each side pool pending and verify that no mapping or resize is published early.

ENABLE_KVCACHED is checked before runner-ownership validation: unset, false, and 0 retain native pools, including direct compressed-pool construction. Native runner tests cover enabled and disabled construction.

Validation at a99c41e: 514 targeted CPU tests, GPU-free C++, all-file pre-commit, MyPy 3.10–3.13 and private Codex review passed. A single T4 passed 18 native SGLang storage/runner cases and 4 exact-ID cases with no skips. Cases cover native row binding/reuse, online banks, sentinel lifetime, allocation misses, rollback and reclamation. They do not load model weights, run compressor/attention kernels or capture graphs.

Full-model DSV4/MTP validation has not been run for this revision. Cross-rank work should first assess #455 and its #487 prerequisite, which still lack current-head multi-GPU TP/PP validation. Performance remains unmeasured: compare with native on identical hardware/model/workload for throughput, TTFT, TPOT, tail latency, memory and allocation/reclamation overhead.