[BugFix] NCCL deadlock in HiCache writing_check by making all_reduce unconditional (#26923)

Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu>
Co-authored-by: Zhangheng <hzh0425@apache.org>

python/sglang/srt/mem_cache/hi_mamba_radix_cache.py

@@ -396,10 +396,11 @@ class HiMambaRadixCache(MambaRadixCache):
                 assert len(self.ongoing_write_through) == 0
             return
 
-        if len(self.ongoing_write_through) == 0:
+        # Every rank must enter the all_reduce below; ongoing_write_through can
-            return
+        # diverge across ranks because loading_check processes DMA completions
-
+        # independently (no cross-rank sync).
         finish_count = 0
+        if len(self.ongoing_write_through) > 0:
             for _, finish_event, ack_list in self.cache_controller.ack_write_queue:
                 if not finish_event.query():
                     break
@@ -426,17 +427,30 @@ class HiMambaRadixCache(MambaRadixCache):
             finish_count -= 1
 
     def loading_check(self):
+        # Every rank must enter the all_reduce below; ongoing_load_back can
+        # diverge across ranks.
         finish_count = 0
         for _, finish_event, ack_list in self.cache_controller.ack_load_queue:
             if not finish_event.query():
-                # the KV cache loading is still ongoing
                 break
             finish_count += 1
+
+        queue_size = torch.tensor(finish_count, dtype=torch.int, device="cpu")
+        if self.tp_world_size > 1:
+            torch.distributed.all_reduce(
+                queue_size,
+                op=torch.distributed.ReduceOp.MIN,
+                group=self.tp_group,
+            )
+        finish_count = int(queue_size.item())
+
+        while finish_count > 0:
+            _, finish_event, ack_list = self.cache_controller.ack_load_queue.pop(0)
+            finish_event.synchronize()
             for ack_id in ack_list:
                 end_node = self.ongoing_load_back.pop(ack_id)
                 self.dec_lock_ref(end_node)
-
+            finish_count -= 1
-        del self.cache_controller.ack_load_queue[:finish_count]
 
     def ready_to_load_host_cache(self) -> int:
         return self.cache_controller.start_loading()

python/sglang/srt/mem_cache/hiradix_cache.py

@@ -916,10 +916,10 @@ class HiRadixCache(RadixCache):
                 assert len(self.ongoing_write_through) == 0
             return
 
-        # NOTE: all ranks has the same ongoing_write_through, can skip sync if empty
+        # Every rank must enter the all_reduce below; ongoing_write_through can
-        if len(self.ongoing_write_through) == 0:
+        # diverge across ranks (e.g. write_backup returning 0 on a subset under
-            return
+        # host memory pressure), so a conditional skip desyncs the NCCL op
-
+        # sequence and deadlocks under TP > 1. (Matches UnifiedRadixCache.)
         finish_count = 0
         if self.pp_rank == 0:
             for _, finish_event, ack_list in self.cache_controller.ack_write_queue: