[autoparallel] add batch norm metainfo (#1815)

* [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

colossalai/auto_parallel/meta_profiler/meta_registry/__init__.py

@@ -1,2 +1,3 @@
 from .conv import *
 from .linear import *
+from .norm import *

colossalai/auto_parallel/meta_profiler/meta_registry/norm.py

@@ -0,0 +1,100 @@
+from typing import Callable, Dict, List, Tuple, Union
+
+import torch
+
+from colossalai.auto_parallel.tensor_shard.sharding_strategy import (
+    MemoryCost,
+    OperationData,
+    OperationDataType,
+    ShardingStrategy,
+    StrategiesVector,
+    TrainCycleItem,
+)
+from colossalai.fx.profiler.memory_utils import activation_size
+from colossalai.fx.profiler.opcount import flop_mapping
+from colossalai.tensor.sharding_spec import ShardingSpec
+
+from ..registry import meta_register
+
+__all__ = ['batchnormnd_meta_info']
+
+
+@meta_register.register(torch.nn.BatchNorm1d)
+@meta_register.register(torch.nn.BatchNorm2d)
+@meta_register.register(torch.nn.BatchNorm3d)
+def batchnormnd_meta_info(*args) -> Tuple[TrainCycleItem, TrainCycleItem, List[torch.Tensor]]:
+    """BatchNorm1d, BatchNorm2d, BatchNorm3d, meta info generator
+    The aten graph of BatchNorm2d is like
+
+    graph():
+    %input_2 : [#users=2] = placeholder[target=placeholder](default=)
+    %cudnn_batch_norm_default : [#users=4] = call_function[target=torch.ops.aten.cudnn_batch_norm.default](args = (%input_2, None, None, None, None, None, None, None), kwargs = {})
+    %zeros_like_default : [#users=1] = call_function[target=torch.ops.aten.zeros_like.default](args = (%cudnn_batch_norm_default,), kwargs = {dtype: None, layout: None, device: None, pin_memory: None})
+    %detach_default : [#users=1] = call_function[target=torch.ops.aten.detach.default](args = (%input_2,), kwargs = {})
+    %detach_default_1 : [#users=1] = call_function[target=torch.ops.aten.detach.default](args = (%cudnn_batch_norm_default,), kwargs = {})
+    %detach_default_2 : [#users=1] = call_function[target=torch.ops.aten.detach.default](args = (%cudnn_batch_norm_default,), kwargs = {})
+    %detach_default_3 : [#users=1] = call_function[target=torch.ops.aten.detach.default](args = (%cudnn_batch_norm_default,), kwargs = {})
+    %cudnn_batch_norm_backward_default : [#users=3] = call_function[target=torch.ops.aten.cudnn_batch_norm_backward.default](args = (%detach_default, %zeros_like_default, None, None, None, %detach_default_1, %detach_default_2, None, %detach_default_3), kwargs = {})
+    %detach_default_4 : [#users=1] = call_function[target=torch.ops.aten.detach.default](args = (%cudnn_batch_norm_backward_default,), kwargs = {})
+    %detach_default_5 : [#users=0] = call_function[target=torch.ops.aten.detach.default](args = (%detach_default_4,), kwargs = {})
+    %detach_default_6 : [#users=1] = call_function[target=torch.ops.aten.detach.default](args = (%cudnn_batch_norm_backward_default,), kwargs = {})
+    %detach_default_7 : [#users=0] = call_function[target=torch.ops.aten.detach.default](args = (%detach_default_6,), kwargs = {})
+    %detach_default_8 : [#users=1] = call_function[target=torch.ops.aten.detach.default](args = (%cudnn_batch_norm_backward_default,), kwargs = {})
+    %detach_default_9 : [#users=0] = call_function[target=torch.ops.aten.detach.default](args = (%detach_default_8,), kwargs = {})
+    Returns:
+        Tuple[TrainCycleItem, TrainCycleItem, List[torch.Tensor]]: compute cost, memory cost and forward inputs
+    """
+
+    input_tensor = next(filter(lambda x: x.type == OperationDataType.ARG, args)).data
+    output_tensor = next(filter(lambda x: x.type == OperationDataType.OUTPUT, args)).data
+    weight_tensor = next(filter(lambda x: x.name == "weight", args)).data
+    bias_tensor = next(filter(lambda x: x.name == "bias", args)).data
+    mean_tensor = next(filter(lambda x: x.name == "running_mean", args)).data
+    var_tensor = next(filter(lambda x: x.name == "running_var", args)).data
+    num_batch = next(filter(lambda x: x.name == "num_batches_tracked", args)).data
+
+    # construct fwd args
+    # the fwd inputs are input, weight, bias, running_mean, running_var and some other args
+    # indicating the status of the module
+    # the fwd outputs are output, saved mean, saved inv std and num batches tracked
+    fwd_in_args = [input_tensor, weight_tensor, bias_tensor, mean_tensor, var_tensor, True, 0.1, 1e-5]
+    fwd_out_args = [output_tensor, mean_tensor, var_tensor, num_batch]
+
+    # construct bwd args
+    # the bwd inputs are upstream grad, input, weight, running_mean, running_var, saved mean,
+    # saved inv std and some other args indicating the status of the module
+    # the bwd outputs are input grad, weight grad and bias grad
+    bwd_in_args = [
+        output_tensor, output_tensor, weight_tensor, mean_tensor, var_tensor, mean_tensor, var_tensor, 1e-5, num_batch
+    ]
+    bwd_out_args = [input_tensor, weight_tensor, bias_tensor]
+
+    # calculate cost
+    fwd_compute_cost = flop_mapping[torch.ops.aten.cudnn_batch_norm.default](fwd_in_args, fwd_out_args)
+    bwd_compute_cost = flop_mapping[torch.ops.aten.cudnn_batch_norm_backward.default](bwd_in_args, bwd_out_args)
+    compute_cost = TrainCycleItem(fwd=fwd_compute_cost, bwd=bwd_compute_cost, total=fwd_compute_cost + bwd_compute_cost)
+
+    # calculate memory cost
+    # the fwd activation cost is output plus saved mean and saved inv std
+    fwd_memory_cost = MemoryCost(activation=activation_size([output_tensor, mean_tensor, var_tensor]),
+                                 parameter=activation_size([weight_tensor, bias_tensor]),
+                                 temp=0,
+                                 buffer=activation_size([mean_tensor, var_tensor]))
+
+    # the bwd memory cost is quite tricky here, BatchNorm will remove saved mean
+    # and saved inv std during backward phase
+    bwd_memory_cost = MemoryCost(activation=activation_size([input_tensor]),
+                                 parameter=activation_size([weight_tensor, bias_tensor]),
+                                 temp=activation_size([mean_tensor, var_tensor]),
+                                 buffer=activation_size([mean_tensor, var_tensor]))
+
+    # total cost is the sum of forward and backward cost
+    total_cost = MemoryCost(activation=fwd_memory_cost.activation + bwd_memory_cost.activation,
+                            parameter=fwd_memory_cost.parameter + bwd_memory_cost.parameter)
+
+    memory_cost = TrainCycleItem(fwd=fwd_memory_cost, bwd=bwd_memory_cost, total=total_cost)
+
+    # store fwd_in
+    fwd_in = [input_tensor]
+
+    return compute_cost, memory_cost, fwd_in

tests/test_auto_parallel/test_tensor_shard/test_metainfo/test_batchnorm_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 _batchnorm_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.BatchNorm2d(128)).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 = 4
+    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_batchnorm_meta_concrete_info_match():
+    world_size = 4
+    run_func_module = partial(_batchnorm_module_mem_test, world_size=world_size, port=free_port())
+    mp.spawn(run_func_module, nprocs=world_size)
+
+
+if __name__ == '__main__':
+    test_batchnorm_meta_concrete_info_match()