solove_device_guard_bugs (#32915)

python/paddle/distributed/collective.py

@@ -927,7 +927,8 @@ def _linear(x, weight, bias=None, name=None):
     else:
         helper = LayerHelper('linear', **locals())
         dtype = x.dtype
-        assert x.ndim < 4, "X latitude is not supported greater than 3 now."
+        assert len(
+            x.shape) < 4, "X latitude is not supported greater than 3 now."
 
         check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'],
                                  'linear')

python/paddle/distributed/fleet/ascend_utils.py

@@ -74,10 +74,17 @@ def _get_ascend_rankfile(rank_table_file_path):
     device_count = 0
     server_list = json_data['server_list']
     for server in server_list:
-        node_ips.append(server['server_id'])
         device_list = server['device']
         device_count = len(device_list)
-
+        if os.getenv("FLAGS_MODELARTS", None):
+            nodes = os.getenv("DLS_TASK_NUMBER", None)
+            assert nodes is not None, "DLS_TASK_NUMBER didn't set!"
+            for node in range(int(nodes)):
+                node_ip = os.getenv(f"VC_CUSTOM{node}_HOSTS", None)
+                assert node_ip is not None, f"VC_CUSTOM{node}_HOSTS didn't set!"
+                node_ips.append(node_ip)
+            return node_ips, device_count
+        node_ips.append(server['server_id'])
     return node_ips, device_count
 
 

python/paddle/fluid/optimizer.py

@@ -4258,7 +4258,7 @@ class PipelineOptimizer(object):
         device = op.attr(self._op_device_key) \
             if op.has_attr(self._op_device_key) else None
         if device:
-            assert device[0:3] == 'gpu' or dev_type == 'npu', "Now, only gpu and npu devices are " \
+            assert device[0:3] == 'gpu' or device[0:3] == 'npu', "Now, only gpu and npu devices are " \
                 "supported in pipeline parallemism."
         return device