def test_fsdp_integration_with_linear_eval(self):
with in_temporary_directory() as pretrain_dir:
# Start pre-training
config = self._create_pretraining_config(
with_fsdp=True,
with_activation_checkpointing=True,
with_mixed_precision=False,
auto_wrap_threshold=0,
)
run_integration_test(config)
# Consolidate the weights
CheckpointFormatConverter.sharded_to_consolidated_checkpoint(
"checkpoint.torch", "checkpoint_conso.torch")
# Load the checkpoint and perform a linear evaluation on it
losses = self.run_linear_eval(
checkpoint_path=os.path.join(pretrain_dir,
"checkpoint_conso.torch"),
with_fsdp=True,
with_mixed_precision=False,
auto_wrap_threshold=0,
)
self.assertEqual(8, len(losses))
print(losses)
def test_fsdp_integration_with_linear_eval(self):
with in_temporary_directory() as pretrain_dir:
# Start pre-training
config = self._create_pretraining_config(
with_fsdp=True,
with_activation_checkpointing=True,
with_mixed_precision=False,
auto_wrap_threshold=0,
)
run_integration_test(config)
# Consolidate the weights (3 different ways)
CheckpointFormatConverter.sharded_to_consolidated_checkpoint(
"checkpoint.torch", "checkpoint_conso.torch")
CheckpointFormatConverter.sharded_to_sliced_checkpoint(
"checkpoint.torch", "checkpoint_sliced.torch")
CheckpointFormatConverter.consolidated_to_sliced_checkpoint(
"checkpoint_conso.torch", "checkpoint_sliced_2.torch")
# Load the sharded checkpoint and perform a inear evaluation on it
ref_losses = self.run_linear_eval(
checkpoint_path=os.path.join(pretrain_dir, "checkpoint.torch"),
with_fsdp=True,
with_mixed_precision=False,
auto_wrap_threshold=0,
)
self.assertEqual(8, len(ref_losses))
# Then check that the results are the same for the other kind of
# checkpoints after consolidation has taken place
for checkpoint_name in [
"checkpoint_conso.torch",
"checkpoint_sliced.torch",
"checkpoint_sliced_2.torch",
]:
losses = self.run_linear_eval(
checkpoint_path=os.path.join(pretrain_dir,
checkpoint_name),
with_fsdp=True,
with_mixed_precision=False,
auto_wrap_threshold=0,
)
self.assertEqual(8, len(losses))
self.assertAlmostEqual(
losses[0],
ref_losses[0],
places=4,
msg=f"Failed for {checkpoint_name}",
)
def convert_checkpoint(input_path: str, output_path: str, output_type: str):
assert g_pathmgr.exists(
input_path), f"Checkpoint input path: {input_path} not found."
# Make the output directory if it doesn't exist.
makedir(os.path.split(output_path)[0])
setup_logging(__name__)
if output_type == CheckpointType.consolidated.name:
CheckpointFormatConverter.sharded_to_consolidated_checkpoint(
input_path, output_path)
elif output_type == CheckpointType.sliced.name:
CheckpointFormatConverter.to_sliced_checkpoint(input_path, output_path)
shutdown_logging()
def test_benchmarking_from_a_consolidated_checkpoint_2(self):
with in_temporary_directory() as checkpoint_folder:
# Run a pre-training in DDP mode and convert to a consolidated checkpoint
config = self._create_pretraining_config(with_fsdp=True)
run_integration_test(config)
sharded_checkpoint_path = os.path.join(checkpoint_folder,
"checkpoint.torch")
checkpoint_path = os.path.join(checkpoint_folder,
"checkpoint.torch")
CheckpointFormatConverter.sharded_to_consolidated_checkpoint(
sharded_checkpoint_path, checkpoint_path)
# Now, run both DDP and FSDP linear evaluation and compare the traces
ddp_losses, ddp_accuracies = self.run_benchmarking(checkpoint_path,
with_fsdp=False)
fsdp_losses, fsdp_accuracies = self.run_benchmarking(
checkpoint_path, with_fsdp=True)
self.assertEqual(ddp_losses, fsdp_losses)
self.assertEqual(ddp_accuracies, fsdp_accuracies)
def test_benchmarking_with_checkpoint_resharding(self):
with in_temporary_directory() as checkpoint_folder:
# Run a pre-training in FSDP mode and save a sharded checkpoint
config = self._create_pretraining_config(with_fsdp=True)
run_integration_test(config)
checkpoint_path = os.path.join(checkpoint_folder,
"checkpoint.torch")
# List the files inside the current working directory
# to later test what files have been created
files_before_conversion = set(os.listdir(checkpoint_folder))
# Transform the sharded checkpoint to a consolidated checkpoint
eval_checkpoint_path_1 = os.path.join(checkpoint_folder,
"checkpoint_eval_1.torch")
CheckpointFormatConverter.sharded_to_consolidated_checkpoint(
input_checkpoint_path=checkpoint_path,
output_checkpoint_path=eval_checkpoint_path_1,
)
# Transform the sharded checkpoint to a sliced checkpoint
eval_checkpoint_path_2 = os.path.join(checkpoint_folder,
"checkpoint_eval_2.torch")
CheckpointFormatConverter.sharded_to_sliced_checkpoint(
input_checkpoint_path=checkpoint_path,
output_checkpoint_path=eval_checkpoint_path_2,
)
# Verify the content of the directory after checkpoint conversion
files_after_conversion = set(os.listdir(checkpoint_folder))
new_files = files_after_conversion - files_before_conversion
expected_new_files = {
"checkpoint_eval_1.torch",
"checkpoint_eval_2.torch",
"checkpoint_eval_2_layers",
}
self.assertEqual(
new_files,
expected_new_files,
"checkpoint 2 slices should be packaged in a directory",
)
# Run a benchmark in FSDP mode and record the losses and accuracies
eval_losses, eval_accuracies = self.run_benchmarking(
checkpoint_path, with_fsdp=True)
self.assertGreater(len(eval_losses), 0)
self.assertEqual(4, len(eval_accuracies))
# Check that these losses and accuracies are the same with the
# consolidated and sliced checkpoints
for eval_checkpoint in [
eval_checkpoint_path_1, eval_checkpoint_path_2
]:
fsdp_losses, fsdp_accuracies = self.run_benchmarking(
eval_checkpoint, with_fsdp=True)
self.assertEqual(fsdp_losses, eval_losses)
self.assertEqual(fsdp_accuracies, eval_accuracies)
# Check that the consolidated and sliced checkpoints, contrary to
# the sharded checkpoint, can be used with a different number of GPUs
for eval_checkpoint in [
eval_checkpoint_path_1, eval_checkpoint_path_2
]:
fsdp_losses, fsdp_accuracies = self.run_benchmarking(
eval_checkpoint, with_fsdp=True, num_gpu=1)
self.assertGreater(len(fsdp_losses), 0)
self.assertEqual(len(fsdp_accuracies), 4)
def _worker(gpu_id: int, sync_file: str, world_size: int):
torch.manual_seed(0)
os.environ["RANK"] = str(gpu_id)
init_distributed_on_file(world_size=world_size,
gpu_id=gpu_id,
sync_file=sync_file)
torch.backends.cudnn.deterministic = True
config = TestCheckpointConversion._create_fsdp_model_config(
with_fsdp=True)
model = build_model(config.MODEL, config.OPTIMIZER).cuda(gpu_id)
model = fsdp_wrapper(model, **config.MODEL.FSDP_CONFIG)
optimizer = optim.SGD(model.parameters(), lr=1e-4)
# Fake inputs
num_iterations = 5
batch_size = 3
torch.manual_seed(gpu_id)
fake_inputs = torch.randn(size=(num_iterations, batch_size, 3, 96, 96))
fake_targets = torch.randn(size=(num_iterations, batch_size))
# Fake training loop
criterion = nn.MSELoss()
for iteration in range(num_iterations):
fake_input = fake_inputs[iteration].cuda(gpu_id)
fake_target = fake_targets[iteration].cuda(gpu_id)
output1, output2 = model(fake_input)[0]
loss = criterion(output1.sum(axis=-1), fake_target) + criterion(
output2.sum(axis=-1), fake_target)
if gpu_id == 0:
print(loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Save a bunch of checkpoint, one by shard
checkpoint_writer = CheckpointWriter(
checkpoint_folder=".",
is_final_train_phase=True,
mode="iteration",
mode_num=0,
backend="disk",
)
content = {
"classy_state_dict": {
"base_model": {
"model": {
"trunk": model.trunk.local_state_dict()
},
"meta": {
"trunk": model.trunk.local_metadata_dict()
},
}
}
}
checkpoint_writer.save_sharded_checkpoint(content,
shard_rank=gpu_id,
world_size=world_size)
dist.barrier()
print(os.listdir("."))
# Convert the checkpoint to consolidated and sliced checkpoints
if gpu_id == 0:
CheckpointFormatConverter.sharded_to_consolidated_checkpoint(
"checkpoint.torch", "checkpoint_conso.torch")
CheckpointFormatConverter.sharded_to_sliced_checkpoint(
"checkpoint.torch", "checkpoint_sliced.torch")
dist.barrier()
print(os.listdir("."))
# Now create models initialized from the previous checkpoint and compare them
fake_test_input = torch.randn(size=(1, 3, 96, 96)).cuda(gpu_id)
shard_cp = CheckpointLoader.load_and_broadcast_init_weights(
"checkpoint.torch", device=torch.device("cpu"))
shard_model = build_model(config.MODEL, config.OPTIMIZER).cuda(gpu_id)
shard_model = fsdp_wrapper(shard_model, **config.MODEL.FSDP_CONFIG)
shard_model.init_model_from_weights_params_file(config, shard_cp)
conso_cp = CheckpointLoader.load_and_broadcast_init_weights(
"checkpoint_conso.torch", device=torch.device("cpu"))
conso_model = build_model(config.MODEL, config.OPTIMIZER).cuda(gpu_id)
conso_model = fsdp_wrapper(conso_model, **config.MODEL.FSDP_CONFIG)
conso_model.init_model_from_weights_params_file(config, conso_cp)
slice_cp = CheckpointLoader.load_and_broadcast_init_weights(
"checkpoint_sliced.torch", device=torch.device("cpu"))
slice_model = build_model(config.MODEL, config.OPTIMIZER).cuda(gpu_id)
slice_model = fsdp_wrapper(slice_model, **config.MODEL.FSDP_CONFIG)
slice_model.init_model_from_weights_params_file(config, slice_cp)
# Verifying that the models are equivalent
if gpu_id == 0:
slice_state_dict = slice_model.local_state_dict()
conso_state_dict = conso_model.local_state_dict()
assert set(slice_state_dict.keys()) == set(conso_state_dict.keys())
for k in slice_state_dict.keys():
slice_val = slice_state_dict[k]
conso_val = conso_state_dict[k]
assert torch.allclose(
slice_val, conso_val
), f"Difference for key {k}: {slice_val} VS {conso_val}"
dist.barrier()
with torch.no_grad():
ref_out = model.trunk(fake_test_input)[0]
shard_out = shard_model.trunk(fake_test_input)[0]
conso_out = conso_model.trunk(fake_test_input)[0]
slice_out = slice_model.trunk(fake_test_input)[0]
assert torch.allclose(
ref_out, shard_out), f"{ref_out.sum()} vs {shard_out.sum()}"
assert torch.allclose(
ref_out, conso_out), f"{ref_out.sum()} vs {conso_out.sum()}"
assert torch.allclose(
ref_out, slice_out), f"{ref_out.sum()} vs {slice_out.sum()}"