Skip to content

fix(fsdp): serialize weight-sync buckets only on the gather src to stop CUDA-IPC pinning #65

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

Open
zhihengy wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix(fsdp): serialize weight-sync buckets only on the gather src to stop CUDA-IPC pinning #65

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
59 changes: 27 additions & 32 deletions miles/backends/fsdp_utils/diffusion_update_weight_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -106,6 +106,7 @@ def __init__(self, args: Namespace, models: dict[str, torch.nn.Module]) -> None:
self.args = args
self.models = models
self.weight_version = 0
self._sync_bucket_count = 0

@abc.abstractmethod
def connect_rollout_engines(
Expand Down Expand Up @@ -206,6 +207,10 @@ def update_bucket_weights(
weight_version=None,
weight_update_mode: str | None = None,
) -> None:
# Serializing pins the storage until the payload is deserialized somewhere and released;
# only the src rank's payload is ever consumed, so no other rank may serialize.
if dist.get_rank() != self._ipc_gather_src:
return
monkey_patch_torch_reductions()
logger.info("Using flattened tensor bucket (diffusion updater, module=%s)", target_module)
named_tensors_by_dtypes = {}
Expand All @@ -215,7 +220,6 @@ def update_bucket_weights(
named_tensors_by_dtypes[dtype] = []
named_tensors_by_dtypes[dtype].append((name, tensor))

serialized_tensors = []
for _dtype, named_tensors in named_tensors_by_dtypes.items():
flattened_tensor_bucket = FlattenedTensorBucket(named_tensors=named_tensors)
metadata = flattened_tensor_bucket.get_metadata()
Expand All @@ -230,37 +234,28 @@ def update_bucket_weights(
"metadata": metadata,
}
}
serialized_tensors.append(MultiprocessingSerializer.serialize(flattened_tensor_data, output_str=True))

if self._ipc_gather_src == dist.get_rank():
gathered_serialized_batches = [None for _ in range(dist.get_world_size(self._ipc_gather_group))]
else:
gathered_serialized_batches = None

dist.gather_object(
obj=serialized_tensors,
object_gather_list=gathered_serialized_batches,
dst=self._ipc_gather_src,
group=self._ipc_gather_group,
)

if dist.get_rank() == self._ipc_gather_src:
# TODO: here we assume all ranks have the same number of dtypes.
num_dtypes = len(gathered_serialized_batches[0])
assert num_dtypes > 0
for i in range(num_dtypes):
kwargs = {
"serialized_named_tensors": [tensors[i] for tensors in gathered_serialized_batches],
"load_format": "flattened_bucket",
"target_modules": [target_module],
"weight_version": str(weight_version),
}
if weight_update_mode is not None:
kwargs["weight_update_mode"] = weight_update_mode
kwargs["lora_alpha"] = self.args.lora_alpha
kwargs["lora_rank"] = self.args.lora_rank
ref = self._ipc_engine.update_weights_from_tensor.remote(**kwargs)
ray.get(ref)
kwargs = {
"serialized_named_tensors": [
MultiprocessingSerializer.serialize(flattened_tensor_data, output_str=True)
],
"load_format": "flattened_bucket",
"target_modules": [target_module],
"weight_version": str(weight_version),
}
if weight_update_mode is not None:
kwargs["weight_update_mode"] = weight_update_mode
kwargs["lora_alpha"] = self.args.lora_alpha
kwargs["lora_rank"] = self.args.lora_rank
ref = self._ipc_engine.update_weights_from_tensor.remote(**kwargs)
ray.get(ref)

self._sync_bucket_count += 1
if self._sync_bucket_count % 16 == 1:
logger.info(
"[wsync-mem] bucket=%d allocated=%.2fGB",
self._sync_bucket_count,
torch.cuda.memory_allocated() / 2**30,
)


# TODO: update weights only for sgl-d LoRA params
Expand Down
Loading