github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-02 09:38:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

[autoparallel] remove redundancy comm node (#1893)

Browse Source
This commit is contained in:
YuliangLiu0306
2022-11-15 10:53:41 +08:00
committed by GitHub
parent 9183e0dec5
commit 36c0f3ea5b
5 changed files with 23 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
colossalai/auto_parallel/passes/runtime_apply_pass.py
View File

@@ -81,6 +81,8 @@ def _shape_consistency_apply(gm: torch.fx.GraphModule):
continue
for user_node_index, user_node in enumerate(node.strategies_vector.successor_nodes):
if node.sharding_spec.sharding_sequence_difference(node.target_sharding_specs[user_node_index]) == 0:
continue
with mod_graph.inserting_before(user_node):
shape_consistency_node = mod_graph.create_node('call_function',
runtime_apply,

8
colossalai/auto_parallel/passes/runtime_preparation_pass.py
View File

@@ -47,6 +47,7 @@ def _solution_annotatation(gm: torch.fx.GraphModule, solution: List[int]):
target_sharding_spec = user_node.best_strategy.get_sharding_spec_by_name(str(node.name))
target_sharding_specs.append(target_sharding_spec)
sharding_spec_convert_dict[index] = target_sharding_specs
setattr(node, 'target_sharding_specs', target_sharding_specs)
# the get_attr node strategy is kind of pending strategy, which means we will change it
# to the same strategy of the user node.
if node.op == 'get_attr':
@@ -95,7 +96,8 @@ def _module_params_sharding(gm: torch.fx.GraphModule, device_mesh):
"""
mod_graph = gm.graph
nodes = tuple(mod_graph.nodes)
# This stream is created for overlaping the communication and computation.
reduction_stream = torch.cuda.Stream()
for node in nodes:
if node.op == 'call_module':
target_module = node.graph.owning_module.get_submodule(node.target)
@@ -122,7 +124,7 @@ def _module_params_sharding(gm: torch.fx.GraphModule, device_mesh):
def wrapper(param, comm_spec):
def hook_fn(grad):
_all_reduce(grad, comm_spec)
_all_reduce(grad, comm_spec, async_op=False)
param.register_hook(hook_fn)
@@ -172,7 +174,7 @@ def _module_params_sharding(gm: torch.fx.GraphModule, device_mesh):
def wrapper(param, comm_spec):
def hook_fn(grad):
_all_reduce(grad, comm_spec)
_all_reduce(grad, comm_spec, async_op=False)
param.register_hook(hook_fn)

8
colossalai/auto_parallel/tensor_shard/node_handler/node_handler.py
View File

@@ -74,11 +74,13 @@ class NodeHandler(ABC):
if op_data.type == OperationDataType.PARAM:
resharding_cost = TrainCycleItem(fwd=0, bwd=0, total=0)
else:
dtype = op_data.data.dtype
size_per_elem_bytes = torch.tensor([], dtype=dtype).element_size()
_, _, resharding_cost = shape_consistency_manager.shape_consistency(
prev_sharding_spec, current_sharding_spec)
resharding_cost = TrainCycleItem(fwd=resharding_cost["forward"],
bwd=resharding_cost["backward"],
total=resharding_cost["total"])
resharding_cost = TrainCycleItem(fwd=resharding_cost["forward"] * size_per_elem_bytes,
bwd=resharding_cost["backward"] * size_per_elem_bytes,
total=resharding_cost["total"] * size_per_elem_bytes)
resharding_costs[node].append(resharding_cost)
strategy.resharding_costs = resharding_costs
return strategy

16
colossalai/auto_parallel/tensor_shard/node_handler/strategy/batch_norm_generator.py
View File

@@ -218,7 +218,7 @@ class BatchNormStrategyGenerator(StrategyGenerator):
sharding_spec=sharding_spec_mapping["output"],
communication_pattern=CollectiveCommPattern.ALLREDUCE_FWD_IDENTITY_BWD,
logical_process_axis=mesh_dim_0,
comm_type=CommType.AFTER)
comm_type=CommType.IMPLICIT)
communication_action_mapping = {"output": output_comm_action}
@@ -254,7 +254,7 @@ class BatchNormStrategyGenerator(StrategyGenerator):
sharding_spec=sharding_spec_mapping["output"],
communication_pattern=CollectiveCommPattern.ALLREDUCE_FWD_IDENTITY_BWD,
logical_process_axis=[mesh_dim_0, mesh_dim_1],
comm_type=CommType.AFTER)
comm_type=CommType.IMPLICIT)
communication_action_mapping = {"output": output_comm_action}
@@ -300,7 +300,7 @@ class BatchNormStrategyGenerator(StrategyGenerator):
sharding_spec=sharding_spec_mapping["output"],
communication_pattern=CollectiveCommPattern.ALLREDUCE_FWD_IDENTITY_BWD,
logical_process_axis=[mesh_dim_0],
comm_type=CommType.AFTER)
comm_type=CommType.IMPLICIT)
communication_action_mapping = {"output": output_comm_action}
@@ -331,14 +331,14 @@ class BatchNormStrategyGenerator(StrategyGenerator):
# TODO: The strategies below should be uncommented after runtime
# passes ready.
# SR = SR x R WITH SYNC_BN
# strategy_list.append(self.split_input_batch(0))
# strategy_list.append(self.split_input_batch(1))
strategy_list.append(self.split_input_batch(0))
strategy_list.append(self.split_input_batch(1))
# SS = SS x S WITH SYNC_BN
# strategy_list.append(self.split_input_both_dim(0, 1))
# strategy_list.append(self.split_input_both_dim(1, 0))
strategy_list.append(self.split_input_both_dim(0, 1))
strategy_list.append(self.split_input_both_dim(1, 0))
# S01R = S01R x R WITH SYNC_BN
# strategy_list.append(self.split_input_batch_1d(0, 1))
strategy_list.append(self.split_input_batch_1d(0, 1))
return strategy_list