diff --git a/torch_npu/csrc/core/npu/NPUCachingAllocator.cpp b/torch_npu/csrc/core/npu/NPUCachingAllocator.cpp
index 42800ac6c0e736695445435f88f7aba6d8234a5a..e42e354a8c003dae43fca375a97ec5a11e8a7db4 100644
--- a/torch_npu/csrc/core/npu/NPUCachingAllocator.cpp
+++ b/torch_npu/csrc/core/npu/NPUCachingAllocator.cpp
@@ -1388,7 +1388,7 @@ public:
                 // Make sure taskqueue is empty, then execute release_cached_blocks
                 c10_npu::npuSynchronizeDevice(true);
             }
-            c10_npu::NPUWorkspaceAllocator::emptyCache(device, true, true);
+            c10_npu::NPUWorkspaceAllocator::emptyCache(device, true);
             block_found = (release_cached_blocks(true, context) && alloc_block(params, true, context, lock));
         }
 
@@ -1744,7 +1744,7 @@ public:
         // Make sure event deque from taskqueue, then synchronize Event
         c10_npu::npuSynchronizeDevice(check_error);
         std::lock_guard<std::recursive_mutex> lock(mutex);
-        c10_npu::NPUWorkspaceAllocator::emptyCache(device, true, check_error);
+        c10_npu::NPUWorkspaceAllocator::emptyCache(device, check_error);
         release_cached_blocks(check_error, context);
     }
 
diff --git a/torch_npu/csrc/core/npu/NPUWorkspaceAllocator.cpp b/torch_npu/csrc/core/npu/NPUWorkspaceAllocator.cpp
index bf01df2a8d2009a8d08e9b609a89fb05bc285176..d599744c7ef4764d47da48d50425974588b04f0a 100644
--- a/torch_npu/csrc/core/npu/NPUWorkspaceAllocator.cpp
+++ b/torch_npu/csrc/core/npu/NPUWorkspaceAllocator.cpp
@@ -221,21 +221,9 @@ public:
     }
 
     // return to the system allocator
-    void empty_cache(bool need_empty_queue, bool check_error)
+    void empty_cache(bool check_error)
     {
-        if (need_empty_queue) {
-            ASCEND_LOGI("NPUWorkspaceAllocator empty_cache in main_thread.");
-            c10_npu::emptyAllNPUStream(check_error);
-        } else {
-            ASCEND_LOGI("NPUWorkspaceAllocator empty_cache in acl_thread.");
-        }
-
-        auto acl_ret = c10_npu::acl::AclrtSynchronizeDeviceWithTimeout();
-        if (check_error) {
-            NPU_CHECK_ERROR(acl_ret, "AclrtSynchronizeDeviceWithTimeout");
-        } else {
-            NPU_CHECK_WARN(acl_ret);
-        }
+        ASCEND_LOGD("NPUWorkspaceAllocator begin empty cache with check_error = %d", check_error);
 
         for (const auto& block_pair : blocks) {
             if (block_pair.second->data_ptr != nullptr) {
@@ -274,6 +262,7 @@ public:
         }
 
         blocks.clear();
+        ASCEND_LOGD("NPUWorkspaceAllocator end empty cache with check_error = %d", check_error);
     }
 
     void record_history(bool enabled, CreateContextFn context_recorder, RecordContext when)
@@ -431,13 +420,6 @@ public:
         auto src_ptr = static_cast<void*>(device_allocator[device]->getStreamPtr(stream));
         *new_ptr = static_cast<void*>(device_allocator[device]->malloc(size, stream));
 
-        // Free all cached blocks and try again.
-        if ((*new_ptr) == nullptr) {
-            device_allocator[device]->empty_cache(false, true);
-            c10_npu::NPUCachingAllocator::emptyCache(true);
-            *new_ptr = static_cast<void*>(device_allocator[device]->malloc(size, stream));
-        }
-
         if ((*new_ptr) == nullptr) {
             size_t device_free;
             size_t device_total;
@@ -450,7 +432,8 @@ public:
                 format_size(device_total),
                 " total capacity; ",
                 format_size(device_free),
-                " free)",
+                " free). If you want to reduce memory usage, ",
+                "take a try to set the environment variable TASK_QUEUE_ENABLE=1.\n" +
                 PTA_ERROR(ErrCode::MEMORY));
         }
 
@@ -459,9 +442,9 @@ public:
         }
     }
 
-    void empty_cache(int device, bool need_empty_queue, bool check_error)
+    void empty_cache(int device, bool check_error)
     {
-        device_allocator[device]->empty_cache(need_empty_queue, check_error);
+        device_allocator[device]->empty_cache(check_error);
         allocated_ptrs.clear();
     }
 
@@ -604,9 +587,9 @@ c10::DataPtr malloc_with_stream(size_t size, aclrtStream stream)
     return workspace_allocator.allocate_with_stream(size, stream);
 }
 
-void emptyCache(int device, bool need_empty_queue, bool check_error)
+void emptyCache(int device, bool check_error)
 {
-    workspace_allocator.empty_cache(device, need_empty_queue, check_error);
+    workspace_allocator.empty_cache(device, check_error);
 }
 
 void recordHistory(bool enabled, CreateContextFn context_recorder, RecordContext when)
diff --git a/torch_npu/csrc/core/npu/NPUWorkspaceAllocator.h b/torch_npu/csrc/core/npu/NPUWorkspaceAllocator.h
index 75c30236c3c7ea9ef21464345685d0524a424f83..75ad259d9c9989f5badd40142e3e2ad3ae20ce76 100644
--- a/torch_npu/csrc/core/npu/NPUWorkspaceAllocator.h
+++ b/torch_npu/csrc/core/npu/NPUWorkspaceAllocator.h
@@ -25,7 +25,7 @@ struct DeviceStats {
 c10::Allocator* get();
 void init();
 c10::DataPtr malloc_with_stream(size_t size, aclrtStream stream);
-void emptyCache(int device, bool need_empty_queue, bool check_error = true);
+void emptyCache(int device, bool check_error = true);
 void recordHistory(bool enabled, CreateContextFn context_recorder, RecordContext when);
 SnapshotInfo snapshot();
 DeviceStats getDeviceStats(int device);