Motivation

I am developing CXL memory-based KV cache sharing (Maru). HiCache's 3-tier hierarchy (L1 GPU → L2 Host → L3 Storage) is a great fit for this, so I integrated it as an L3 backend and started validation testing.

During KV sharing tests between two SGLang instances, I found that write-through completions are never flushed to L3 storage while the scheduler is idle. check_hicache_events() is called inside _get_new_batch_prefill_raw() after an early return that fires when waiting_queue is empty. During idle periods, the method returns before reaching check_hicache_events(), so the write-through pipeline (GPU→Host→L3) never completes and the other instance sees zero cache hits.
This bug affects all HiCache L3 backends (not just Maru) when using write-through policy.

Modifications

Move check_hicache_events() call before the early-return condition in _get_new_batch_prefill_raw(), so write-through completions and prefetch events are processed on every scheduler tick regardless of whether new prefill work is available.

Checklist

• Format your code according to the Format code with pre-commit.