github
/
ColossalAI
mirror of https://github.com/hpcaitech/ColossalAI
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

[hotfix] fx shard 1d pass bug fixing (#1220)

pull/1219/head^2
Jiarui Fang 2022-07-07 13:37:31 +08:00 committed by GitHub
parent 11973d892d
commit db1bef9032
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 8 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
colossalai/fx/passes/shard_1d_pass.py
View File

@ -1,13 +1,9 @@
import torch import torch
from torch.fx.node import map_arg from torch.fx.node import map_arg
from torch.fx.node import Node from colossalai.tensor import ColoTensorSpec, distspec, ProcessGroup, ComputeSpec, ComputePattern
from torch.fx.passes.split_module import split_module
import colossalai
from colossalai.tensor import ColoTensor, TensorSpec, distspec, ProcessGroup, ComputeSpec, ComputePattern
def weight_split(weight: torch.nn.parameter.Parameter, dim: int) -> torch.nn.parameter.Parameter: def weight_split(weight: torch.Tensor, dim: int) -> torch.nn.parameter.Parameter:
"""weight_split """weight_split
split a nn.Parameter split a nn.Parameter
@ -18,22 +14,20 @@ def weight_split(weight: torch.nn.parameter.Parameter, dim: int) -> torch.nn.par
Returns: Returns:
_type_: _description_ _type_: _description_
""" """
#TODO: This func temporarily works with no materialization
# Append a Tensor spec to target_module.weight.shard # Append a Tensor spec to target_module.weight.shard
# Convert to ColoTensor: colo_tensor = ColoTensor.from_torch_tensor(tensor, spec) # Convert to ColoTensor: colo_tensor = ColoTensor.from_torch_tensor(tensor, spec)
# assert isinstance(weight, torch.nn.parameter.Parameter), \ assert isinstance(weight, torch.Tensor), \
# f'The type of the input tensor should be torch.nn.parameter' \ f'The type of the input tensor should be torch.nn.parameter' \
# f'Your Input tensor is {type(weight)}' f'Your Input tensor is {type(weight)}'
# FIXME() I initialized a PG for this tensor. Only has TP comm group. # FIXME() I initialized a PG for this tensor. Only has TP comm group.
# we only consider the TP-only caes. # we only consider the TP-only caes.
world_size = torch.distributed.get_world_size() world_size = torch.distributed.get_world_size()
pg = ProcessGroup(tp_degree=world_size) pg = ProcessGroup(tp_degree=world_size)
spec = TensorSpec(distspec.shard(pg, [dim], [pg.tp_world_size()]), ComputeSpec(ComputePattern.TP1D)) spec = ColoTensorSpec(pg, distspec.shard([dim], [pg.tp_world_size()]), ComputeSpec(ComputePattern.TP1D))
# As you has constructed a Spec, why not directly convert the tensor to ColoTensor. # As you has constructed a Spec, why not directly convert the tensor to ColoTensor.
# setattr(weight, "fx_attr", spec) setattr(weight, "fx_attr", spec)
weight.data = ColoTensor(data=weight.data, spec=spec)
return weight return weight
@ -58,6 +52,7 @@ def row_shard_linear_pass(gm: torch.fx.GraphModule):
target_module = node.graph.owning_module.get_submodule(node.target) target_module = node.graph.owning_module.get_submodule(node.target)
if isinstance(target_module, torch.nn.Linear): if isinstance(target_module, torch.nn.Linear):
target_module.weight = weight_split(target_module.weight, dim=-1) target_module.weight = weight_split(target_module.weight, dim=-1)
gm.recompile() gm.recompile()
return gm return gm