[HiCache] Reject load-back specs that claim nodes pinned by an in-flight load-back (#35931)

python/sglang/srt/mem_cache/unified_cache/unified_tree_core.py

@@ -1927,6 +1927,20 @@ class UnifiedTreeCore(UnifiedTreeCoreInterface):
             )
             if t:
                 comp_xfers[comp.component_type] = t
+        # Reject transfers that would claim a node pinned by another load-back
+        # anchor; the empty spec makes the caller back off and recompute.
+        if any(
+            self.node_by_id(nid).load_back_pending_id not in (None, node_id)
+            for xfers in ([kv_xfer], *comp_xfers.values())
+            for xfer in xfers
+            for nid in xfer.nodes_to_load or ()
+        ):
+            empty_kv = PoolTransfer(
+                name=PoolName.KV,
+                host_indices=torch.empty((0,), dtype=torch.int64, device="cpu"),
+                nodes_to_load=[],
+            )
+            return empty_kv, {}
         return kv_xfer, comp_xfers
 
     def prefetch_anchor_info(

python/sglang/srt/mem_cache/unified_radix_cache.py

@@ -1465,8 +1465,10 @@ class UnifiedRadixCache(BasePrefixCache):
 
         # Skip if there is nothing to load, or if the Full-KV transfer is too
         # small / exceeds memory quota. Aux transfers should still run even
-        # when the Full-KV load is skipped by thresholding.
-        if (kv_tokens < self.load_back_threshold and not comp_xfers) or (
+        # when the Full-KV load is skipped by thresholding. max(1, ...): an
+        # 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
         ):
             self.dec_lock_ref(node_id, ancestor_lock_params)