Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[autoparallel] fix param hook issue in transform pass #1755

Merged
merged 1 commit into from
Oct 24, 2022
Merged

[autoparallel] fix param hook issue in transform pass #1755

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 7 additions & 3 deletions colossalai/fx/passes/experimental/adding_shape_consistency_pass_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -70,10 +70,14 @@ def solution_annotatation_pass(gm: torch.fx.GraphModule, solution: List[int], de
comm_spec_to_use = comm_action.comm_spec
if operation_data.type == OperationDataType.PARAM and operation_data.name == name and comm_action.comm_type == CommType.HOOK:

def hook_fn(grad):
_all_reduce(grad, comm_spec_to_use)
def wrapper(param, comm_spec):

param_sharded.register_hook(hook_fn)
def hook_fn(grad):
_all_reduce(grad, comm_spec)

param.register_hook(hook_fn)

wrapper(param_sharded, comm_spec_to_use)

sharded_buffer_dict = {}
for name, buffer in target_module.named_buffers():
Expand Down
86 changes: 78 additions & 8 deletions tests/test_auto_parallel/test_tensor_shard/test_resnet_block_runtime.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,22 +171,92 @@ def check_apply_bottleneck(rank, world_size, port):
torch.cuda.set_rng_state(cuda_rng_state)
origin_output.sum().backward()
if rank == 0:
print((gm.bn3.weight.grad - test_model.bn3.weight.grad.narrow(0, 0, 4)).abs().sum())
print((gm.conv3.weight.grad - test_model.conv3.weight.grad.narrow(0, 0, 8)).abs().sum())
print(
f"bn3 diff sum in rank {rank}: {(gm.bn3.weight.grad - test_model.bn3.weight.grad.narrow(0, 0, 4)).abs().sum()}"
)
print(
f"conv3 diff sum in rank {rank}: {(gm.conv3.weight.grad - test_model.conv3.weight.grad.narrow(0, 0, 8)).abs().sum()}"
)
print(
f"bn2 diff sum in rank {rank}: {(gm.bn2.weight.grad - test_model.bn2.weight.grad.narrow(0, 0, 2)).abs().sum()}"
)
print(
f"conv2 diff sum in rank {rank}: {(gm.conv2.weight.grad - test_model.conv2.weight.grad.narrow(0, 0, 2)).abs().sum()}"
)
print(
f"bn1 diff sum in rank {rank}: {(gm.bn1.weight.grad - test_model.bn1.weight.grad.narrow(0, 0, 1)).abs().sum()}"
)
print(f"conv1 diff sum in rank {rank}: {(gm.conv1.weight.grad - test_model.conv1.weight.grad).sum()}")

assert_close_loose(gm.conv3.weight.grad.sum(), test_model.conv3.weight.grad.narrow(0, 0, 8).sum())
assert_close_loose(gm.conv2.weight.grad.sum(), test_model.conv2.weight.grad.narrow(0, 0, 2).sum())
assert_close_loose(gm.conv1.weight.grad.sum(), test_model.conv1.weight.grad.sum())

if rank == 1:
print((gm.bn3.weight.grad - test_model.bn3.weight.grad.narrow(0, 4, 4)).abs().sum())
print((gm.conv3.weight.grad - test_model.conv3.weight.grad.narrow(0, 0, 8)).abs().sum())
print(
f"bn3 diff sum in rank {rank}: {(gm.bn3.weight.grad - test_model.bn3.weight.grad.narrow(0, 4, 4)).abs().sum()}"
)
print(
f"conv3 diff sum in rank {rank}: {(gm.conv3.weight.grad - test_model.conv3.weight.grad.narrow(0, 0, 8)).abs().sum()}"
)
print(
f"bn2 diff sum in rank {rank}: {(gm.bn2.weight.grad - test_model.bn2.weight.grad.narrow(0, 2, 2)).abs().sum()}"
)
print(
f"conv2 diff sum in rank {rank}: {(gm.conv2.weight.grad - test_model.conv2.weight.grad.narrow(0, 2, 2)).abs().sum()}"
)
print(
f"bn1 diff sum in rank {rank}: {(gm.bn1.weight.grad - test_model.bn1.weight.grad.narrow(0, 1, 1)).abs().sum()}"
)
print(f"conv1 diff sum in rank {rank}: {(gm.conv1.weight.grad - test_model.conv1.weight.grad).sum()}")

assert_close_loose(gm.conv3.weight.grad.sum(), test_model.conv3.weight.grad.narrow(0, 0, 8).sum())
assert_close_loose(gm.conv2.weight.grad.sum(), test_model.conv2.weight.grad.narrow(0, 2, 2).sum())
assert_close_loose(gm.conv1.weight.grad.sum(), test_model.conv1.weight.grad.sum())

if rank == 2:
print((gm.bn3.weight.grad - test_model.bn3.weight.grad.narrow(0, 8, 4)).abs().sum())
print((gm.conv3.weight.grad - test_model.conv3.weight.grad.narrow(0, 8, 8)).abs().sum())
print(
f"bn3 diff sum in rank {rank}: {(gm.bn3.weight.grad - test_model.bn3.weight.grad.narrow(0, 8, 4)).abs().sum()}"
)
print(
f"conv3 diff sum in rank {rank}: {(gm.conv3.weight.grad - test_model.conv3.weight.grad.narrow(0, 8, 8)).abs().sum()}"
)
print(
f"bn2 diff sum in rank {rank}: {(gm.bn2.weight.grad - test_model.bn2.weight.grad.narrow(0, 0, 2)).abs().sum()}"
)
print(
f"conv2 diff sum in rank {rank}: {(gm.conv2.weight.grad - test_model.conv2.weight.grad.narrow(0, 0, 2)).abs().sum()}"
)
print(
f"bn1 diff sum in rank {rank}: {(gm.bn1.weight.grad - test_model.bn1.weight.grad.narrow(0, 2, 1)).abs().sum()}"
)
print(f"conv1 diff sum in rank {rank}: {(gm.conv1.weight.grad - test_model.conv1.weight.grad).sum()}")

assert_close_loose(gm.conv3.weight.grad.sum(), test_model.conv3.weight.grad.narrow(0, 8, 8).sum())
assert_close_loose(gm.conv2.weight.grad.sum(), test_model.conv2.weight.grad.narrow(0, 0, 2).sum())
assert_close_loose(gm.conv1.weight.grad.sum(), test_model.conv1.weight.grad.sum())

if rank == 3:
print((gm.bn3.weight.grad - test_model.bn3.weight.grad.narrow(0, 12, 4)).abs().sum())
print((gm.conv3.weight.grad - test_model.conv3.weight.grad.narrow(0, 8, 8)).abs().sum())
print(
f"bn3 diff sum in rank {rank}: {(gm.bn3.weight.grad - test_model.bn3.weight.grad.narrow(0, 12, 4)).abs().sum()}"
)
print(
f"conv3 diff sum in rank {rank}: {(gm.conv3.weight.grad - test_model.conv3.weight.grad.narrow(0, 8, 8)).abs().sum()}"
)
print(
f"bn2 diff sum in rank {rank}: {(gm.bn2.weight.grad - test_model.bn2.weight.grad.narrow(0, 2, 2)).abs().sum()}"
)
print(
f"conv2 diff sum in rank {rank}: {(gm.conv2.weight.grad - test_model.conv2.weight.grad.narrow(0, 2, 2)).abs().sum()}"
)
print(
f"bn1 diff sum in rank {rank}: {(gm.bn1.weight.grad - test_model.bn1.weight.grad.narrow(0, 3, 1)).abs().sum()}"
)
print(f"conv1 diff sum in rank {rank}: {(gm.conv1.weight.grad - test_model.conv1.weight.grad).sum()}")

assert_close_loose(gm.conv3.weight.grad.sum(), test_model.conv3.weight.grad.narrow(0, 8, 8).sum())
assert_close_loose(gm.conv2.weight.grad.sum(), test_model.conv2.weight.grad.narrow(0, 2, 2).sum())
assert_close_loose(gm.conv1.weight.grad.sum(), test_model.conv1.weight.grad.sum())


@run_on_environment_flag(name='AUTO_PARALLEL')
Expand Down