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

Handle Manual Wrapping in FSDP. Minor fix of fsdp example. #342

Merged
merged 8 commits into from
May 5, 2022
Merged

Handle Manual Wrapping in FSDP. Minor fix of fsdp example. #342

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
1e97f11
Create peak_memory_uasge_tracker.py
pacman100 Apr 29, 2022
c60075c
fixing the typo in the file name
pacman100 Apr 29, 2022
3df1ac8
reformatting
pacman100 Apr 29, 2022
4fd3c0a
exclude peak_memory_usage_tracker.py from tests
pacman100 Apr 29, 2022
2467bf5
renaming and highlighting proper usage
pacman100 Apr 29, 2022
db3b97d
Update test_examples.py
pacman100 Apr 29, 2022
5fc33cd
Merge branch 'main' into fsdp_blog
pacman100 May 5, 2022
f2c9e8c
Handle manual wrapping in FSDP. Fix fsdp example.
pacman100 May 5, 2022
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
26 changes: 14 additions & 12 deletions examples/by_feature/fsdp_with_peak_mem_tracking.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -251,12 +251,13 @@ def collate_fn(examples):
)
)
# Logging the peak memory usage of the GPU to the tracker
accelerator.log(
{
"train_total_peak_memory": tracemalloc.peaked + b2mb(tracemalloc.begin),
},
step=epoch,
)
if args.with_tracking:
accelerator.log(
{
"train_total_peak_memory": tracemalloc.peaked + b2mb(tracemalloc.begin),
},
step=epoch,
)

# New Code #
# context manager to track the peak memory usage during the evaluation
Expand Down Expand Up @@ -302,12 +303,13 @@ def collate_fn(examples):
"Total Peak Memory consumed during the eval (max): {}".format(tracemalloc.peaked + b2mb(tracemalloc.begin))
)
# Logging the peak memory usage of the GPU to the tracker
accelerator.log(
{
"eval_total_peak_memory": tracemalloc.peaked + b2mb(tracemalloc.begin),
},
step=epoch,
)
if args.with_tracking:
accelerator.log(
{
"eval_total_peak_memory": tracemalloc.peaked + b2mb(tracemalloc.begin),
},
step=epoch,
)

if args.with_tracking:
accelerator.end_training()
Expand Down
25 changes: 14 additions & 11 deletions src/accelerate/accelerator.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -465,17 +465,20 @@ def prepare_model(self, model):
elif self.distributed_type == DistributedType.FSDP:
from torch.distributed.fsdp.fully_sharded_data_parallel import FullyShardedDataParallel as FSDP

fsdp_plugin = self.state.fsdp_plugin
model = FSDP(
model,
sharding_strategy=fsdp_plugin.sharding_strategy,
cpu_offload=fsdp_plugin.cpu_offload,
auto_wrap_policy=fsdp_plugin.auto_wrap_policy,
backward_prefetch=fsdp_plugin.backward_prefetch,
ignored_modules=fsdp_plugin.ignored_modules,
)
if not fsdp_plugin.cpu_offload.offload_params:
model.to(self.device)
# Check if the model is already a FSDP model due to `Manual Wrapping` and if so,
# don't wrap it again
if type(model) != FSDP:
fsdp_plugin = self.state.fsdp_plugin
model = FSDP(
model,
sharding_strategy=fsdp_plugin.sharding_strategy,
cpu_offload=fsdp_plugin.cpu_offload,
auto_wrap_policy=fsdp_plugin.auto_wrap_policy,
backward_prefetch=fsdp_plugin.backward_prefetch,
ignored_modules=fsdp_plugin.ignored_modules,
)
if not fsdp_plugin.cpu_offload.offload_params:
model.to(self.device)
elif self.distributed_type == DistributedType.MULTI_CPU:
kwargs = self.ddp_handler.to_kwargs() if self.ddp_handler is not None else {}
model = torch.nn.parallel.DistributedDataParallel(model, **kwargs)
Expand Down