HiCache Fix (#8288)

Co-authored-by: pansicheng <sicheng.pan.chn@gmail.com>

python/sglang/srt/managers/cache_controller.py

@@ -358,6 +358,7 @@ class HiCacheController:
         if host_indices is None:
             return None
         self.mem_pool_host.protect_write(host_indices)
+        torch.cuda.current_stream().synchronize()
         self.write_queue.put(
             CacheOperation(host_indices, device_indices, node_id, priority)
         )

python/sglang/srt/model_executor/model_runner.py

@@ -378,6 +378,7 @@ class ModelRunner:
                     is_hopper_with_cuda_12_3()
                     and is_no_spec_infer_or_topk_one(server_args)
                     and is_fa3_default_architecture(self.model_config.hf_config)
+                    and (not server_args.enable_hierarchical_cache)
                 ):
                     server_args.attention_backend = "fa3"
                 elif _is_hip:
@@ -390,7 +391,9 @@ class ModelRunner:
                     )
             else:
                 # MLA architecture
-                if is_hopper_with_cuda_12_3():
+                if is_hopper_with_cuda_12_3() and (
+                    not server_args.enable_hierarchical_cache
+                ):
                     server_args.attention_backend = "fa3"
                 elif is_sm100_supported():
                     server_args.attention_backend = "flashinfer"