[autoparallel] add torch.nn.ReLU metainfo (#1868)

* [fx] metainfo class for auto parallel

* [fx] add unit test for linear metainfo

* [fx] fix bwd param for linear

* [fx] modify unit test

* [fx] modify unit test

* [fx] modify import

* [fx] modify import

* [fx] modify import

* [fx] move meta profiler to auto parallel

* [fx] add conv metainfo class

* [fx] restore profiler

* [fx] restore meta profiler

* [autoparallel] modify unit test

* [fx] modify unit test

* [autoparallel] add batchnorm metainfo class

* [autoparallel] fix batchnorm unit test function declaration

* [fx] restore profiler

* [fx] add relu metainfo class

* [fx] restore profiler

* [autoparallel] modify metainfo input

tests/test_auto_parallel/test_tensor_shard/test_metainfo/test_activation_metainfo.py

@@ -0,0 +1,61 @@
+from functools import partial
+
+import pytest
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+
+from colossalai.device.device_mesh import DeviceMesh
+from colossalai.fx import ColoGraphModule, ColoTracer
+from colossalai.initialize import launch
+from colossalai.logging import disable_existing_loggers
+from colossalai.testing.pytest_wrapper import run_on_environment_flag
+from colossalai.testing.utils import parameterize, rerun_if_address_is_in_use
+from colossalai.utils import free_port
+from tests.test_auto_parallel.test_tensor_shard.test_metainfo.utils import mem_test_for_node_strategy
+
+
+def _ReLU_module_mem_test(rank, world_size, port):
+    """This function is for conv memory test
+    Test and print real memory cost and estimated, this test will not be executed except with the tag AUTO_PARALLEL
+
+    Args:
+    Args:
+        rank: device rank
+        bias: indicate whether conv module need bias
+        world_size: number of devices
+        port: port for initializing process group
+    """
+    disable_existing_loggers()
+    launch(config={}, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
+    model = nn.Sequential(nn.ReLU()).cuda()
+    input = torch.rand(4, 128, 64, 64).cuda()
+    input.requires_grad = True
+    physical_mesh_id = torch.arange(0, 4)
+    mesh_shape = (2, 2)
+    device_mesh = DeviceMesh(physical_mesh_id, mesh_shape, init_process_group=True)
+
+    # index of conv node in computation graph
+    node_index = 1
+    # total number of conv strategies
+    strategy_number = 1
+    mem_test_for_node_strategy(rank=rank,
+                               model=model,
+                               device_mesh=device_mesh,
+                               node_index=node_index,
+                               strategy_number=strategy_number,
+                               input_args=[input],
+                               meta_arg_names=['input'])
+
+
+@run_on_environment_flag(name='AUTO_PARALLEL')
+@pytest.mark.dist
+@rerun_if_address_is_in_use()
+def test_ReLU_meta_concrete_info_match():
+    world_size = 4
+    run_func_module = partial(_ReLU_module_mem_test, world_size=world_size, port=free_port())
+    mp.spawn(run_func_module, nprocs=world_size)
+
+
+if __name__ == '__main__':
+    test_ReLU_meta_concrete_info_match()

tests/test_auto_parallel/test_tensor_shard/test_metainfo/utils.py

@@ -60,9 +60,10 @@ def mem_test_for_node_strategy(rank: int,
         gm.recompile()
         gm: GraphModule
 
+        num_of_strategies = len(target_node.strategies_vector)
         if rank == 0:
             print("=======================")
-            print(f"#strategy_index: {strategy_index}")
+            print(f"#strategy_index: {strategy_index + 1}/{num_of_strategies}")
             pprint(target_node.strategies_vector[strategy_index])
 
         # warmup
@@ -104,7 +105,7 @@ def mem_test_for_node_strategy(rank: int,
 
             # estimated memory
             metainfo = MetaInfo(target_node.strategies_vector[strategy_index],
-                                target_node.graph.owning_module.get_submodule(target_node.target).__class__)
+                                target_node.graph.owning_module.get_submodule(target_node.target))
             print("estimated memory:")
             print(
                 f"forward activation: {metainfo.memory_cost.fwd.activation / 1024} kb, forward param: {metainfo.memory_cost.fwd.parameter / 1024} kb"