[mem_cache] Carry swa_evicted_seqlen into SWARadixCache.cache_unfinished_req (#36909)

2026-08-28 14:29:27 -07:00
parent 4d78d59e51
commit 60f6d77a98
2 changed files with 58 additions and 0 deletions
@@ -539,11 +539,14 @@ class SWARadixCache(BasePrefixCache):

        # Radix Cache takes one ref in memory pool
        # Note: the insert function already frees the overlapped kv_indices
+        # The prefix below swa_evicted_seqlen has no SWA peers left; the insert
+        # tombstones it instead of claiming live SWA KV.
        result = self.insert(
            InsertParams(
                key=radix_key,
                value=values,
                prev_prefix_len=old_prefix_len,
+                swa_evicted_seqlen=req.kv.swa_evicted_seqlen,
            )
        )
        new_prefix_len = result.prefix_len