[HiCache] Account for newly pinned ancestors in load-back quota (#38481)

python/sglang/srt/mem_cache/unified_radix_cache.py

@@ -1594,7 +1594,7 @@ class UnifiedRadixCache(BasePrefixCache):
         # entirely empty spec (e.g. foreign-pin rejection) must never report
         # success, even at load_back_threshold <= 0.
         if (kv_tokens < max(1, self.load_back_threshold) and not comp_xfers) or (
-            mem_quota is not None and kv_tokens > mem_quota + result.delta
+            mem_quota is not None and kv_tokens + result.delta > mem_quota
         ):
             self.dec_lock_ref(node_id, ancestor_lock_params)
             self.dec_host_lock_ref(node_id, host_anchor_params)

test/registered/unit/mem_cache/test_unified_radix_cache_unittest.py

@@ -6281,7 +6281,8 @@ class UnifiedRadixCacheSuite:
                 best_match_node=req.best_match_node,
                 host_hit_length=req.host_hit_length,
                 req=req,
-                mem_quota=-1_000_000,
+                # Pinning the resident ancestors leaves too little for the tail.
+                mem_quota=len(tokens) - 1,
             )
         )