[Tensor Parallelism] split fix bug (#33015)

Author: JZ-LIANG

python/paddle/distributed/collective.py

@@ -977,6 +977,11 @@ def _parallel_linear(x,
                      group=None):
     """
     Parallel Linear
+
+    axis the dimension of the parameter of linear layer. 
+    axis = 0: the row dimension
+    axid = 1: the col dimension
+    
     """
     if group is not None and not group.is_member():
         return
@@ -1008,6 +1013,12 @@ def _parallel_linear(x,
     main_block = paddle.static.default_main_program().global_block()
     startup_block.vars[linear.weight.name].is_distributed = True
     main_block.vars[linear.weight.name].is_distributed = True
+    # set is_distributed for splited bias
+    # if a linear layer is splited by row, each rank would hold a complete bias and they should be the same in each rank.
+    # if a linear layer is splited by col, the bias would also be split into each rank as its weight
+    if axis == 1 and linear._bias_attr != False:
+        startup_block.vars[linear.bias.name].is_distributed = True
+        main_block.vars[linear.bias.name].is_distributed = True
 
     if not gather_out: return linear_out
 

python/paddle/distributed/fleet/base/distributed_strategy.py

@@ -814,7 +814,7 @@ def sharding_configs(self):
                 "sharding_segment_strategy": "segment_broadcast_MB",
                 "segment_broadcast_MB": 32,
                 "sharding_degree": 8,
-                "sharding_degree": 2,
+                "dp_degree": 2,
                 "gradient_merge_acc_step": 4,
                 }
         """

python/paddle/fluid/contrib/mixed_precision/fp16_lists.py

@@ -145,6 +145,7 @@ def _update_list(self):
     'sign',
     'cast',
     'fused_bn_add_activation',
+    'c_identity',
 }
 
 # The set of ops that don't support fp16 calculation

python/paddle/fluid/tests/unittests/column_parallel_linear_api.py

@@ -69,7 +69,7 @@ def get_model(self, main_prog, startup_program, rank):
                 axis=1,
                 num_partitions=2,
                 weight_attr=param_attr,
-                bias_attr=False, )
+                bias_attr=True, )
 
             return [linear_out]
 

python/paddle/fluid/tests/unittests/row_parallel_linear_api.py

@@ -65,12 +65,12 @@ def get_model(self, main_prog, startup_program, rank):
 
             linear_out = paddle.distributed.split(
                 data,
-                size=(1000, 8),
+                size=(1000, 16),
                 operation='linear',
                 axis=0,
                 num_partitions=2,
                 weight_attr=param_attr,
-                bias_attr=False, )
+                bias_attr=True, )
 
             return [linear_out]
 

python/paddle/fluid/tests/unittests/test_collective_api_base.py

@@ -154,7 +154,10 @@ def _run_cluster(self, model_file, envs):
         #update environment
         env0.update(envs)
         env1.update(envs)
-        tr_cmd = "%s %s"
+        if os.getenv('WITH_COVERAGE', 'OFF') == 'ON':
+            tr_cmd = "%s -m coverage run --branch -p %s"
+        else:
+            tr_cmd = "%s %s"
         tr0_cmd = tr_cmd % (self._python_interp, model_file)
         tr1_cmd = tr_cmd % (self._python_interp, model_file)
         tr0_pipe = open("/tmp/tr0_err_%d.log" % os.getpid(), "w")