mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-08 07:39:33 +01:00
ff8e33556e
Enables the deduplication of saved entries by load balancing duplicates across ranks.
Tested with existing and modified tests. Additionally tested with the following code snippet, which saves a 20GB DDP model in **~3 seconds on 8 ranks**. Before this PR, the same operation has been measured at ~19 seconds.
```
def run(local_rank, world_size, param_size, num_params, work_dir):
os.environ["RANK"] = str(local_rank)
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = "12355"
device = torch.device(f"cuda:{local_rank}")
torch.cuda.set_device(device)
dist.init_process_group(backend="nccl", rank=local_rank, world_size=world_size)
model = Model(param_size=param_size, num_params=num_params)
model = DistributedDataParallel(model, gradient_as_bucket_view=True)
_patch_model_state_dict(model)
sz = sum(t.nelement() * t.element_size() for t in model.parameters())
rank_0_print(f"Model size: {sz / 1_000_000_000.0} GB")
rank_0_print("Saving the model with DCP...")
checkpointer = _FileSystemCheckpointer(
f"{args.work_dir}/dcp",
sync_files=False,
single_file_per_rank=False,
thread_count=1
)
begin_ts = time.monotonic()
checkpointer.save(state_dict={"model": model})
end_ts = time.monotonic()
rank_0_print(f"Took {end_ts - begin_ts} seconds with DCP")
```
Differential Revision: [D52435926](https://our.internmc.facebook.com/intern/diff/D52435926/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/116469
Approved by: https://github.com/fegin, https://github.com/wz337
342 lines
14 KiB
Python
342 lines
14 KiB
Python
# Owner(s): ["oncall: distributed"]
|
|
|
|
import sys
|
|
|
|
import torch
|
|
|
|
from torch.distributed._shard.sharded_tensor import (
|
|
Shard,
|
|
ShardedTensor,
|
|
ShardedTensorMetadata,
|
|
ShardMetadata,
|
|
)
|
|
from torch.distributed._shard.sharded_tensor.metadata import (
|
|
TensorProperties as TensorProperties_Shard,
|
|
)
|
|
from torch.distributed.checkpoint._dedup_save_plans import dedup_save_plans
|
|
|
|
from torch.distributed.checkpoint.default_planner import (
|
|
_create_default_local_metadata,
|
|
create_default_global_save_plan,
|
|
create_default_local_load_plan,
|
|
create_default_local_save_plan,
|
|
)
|
|
from torch.distributed.checkpoint.metadata import (
|
|
BytesStorageMetadata,
|
|
ChunkStorageMetadata,
|
|
MetadataIndex,
|
|
TensorProperties,
|
|
TensorStorageMetadata,
|
|
)
|
|
from torch.distributed.checkpoint.planner import LoadItemType, WriteItemType
|
|
|
|
from torch.distributed.checkpoint.planner_helpers import (
|
|
create_read_items_for_chunk_list,
|
|
)
|
|
|
|
from torch.testing._internal.common_utils import (
|
|
run_tests,
|
|
TEST_WITH_DEV_DBG_ASAN,
|
|
TestCase,
|
|
)
|
|
|
|
from torch.testing._internal.distributed.distributed_utils import (
|
|
with_dist,
|
|
with_fake_comms,
|
|
)
|
|
|
|
|
|
if TEST_WITH_DEV_DBG_ASAN:
|
|
print(
|
|
"Skip dev-asan as torch + multiprocessing spawn have known issues",
|
|
file=sys.stderr,
|
|
)
|
|
sys.exit(0)
|
|
|
|
|
|
def create_sharded_tensor(rank, world_size, shards_per_rank, shard_size=8):
|
|
shards_metadata = []
|
|
local_shards = []
|
|
for idx in range(0, world_size * shards_per_rank):
|
|
shard_rank = idx // shards_per_rank
|
|
shard_md = ShardMetadata(
|
|
shard_offsets=[idx * shard_size],
|
|
shard_sizes=[shard_size],
|
|
placement=f"rank:{shard_rank}/cpu",
|
|
)
|
|
shards_metadata.append(shard_md)
|
|
if shard_rank == rank:
|
|
shard = Shard.from_tensor_and_offsets(
|
|
torch.rand(*shard_md.shard_sizes),
|
|
shard_offsets=shard_md.shard_offsets,
|
|
rank=rank,
|
|
)
|
|
local_shards.append(shard)
|
|
|
|
sharded_tensor_md = ShardedTensorMetadata(
|
|
shards_metadata=shards_metadata,
|
|
size=torch.Size([shard_size * len(shards_metadata)]),
|
|
tensor_properties=TensorProperties_Shard.create_from_tensor(torch.zeros(1)),
|
|
)
|
|
|
|
return ShardedTensor._init_from_local_shards_and_global_metadata(
|
|
local_shards=local_shards, sharded_tensor_metadata=sharded_tensor_md
|
|
)
|
|
|
|
|
|
class TestSavePlan(TestCase):
|
|
@with_fake_comms(rank=1, world_size=4)
|
|
def test_local_plan(self):
|
|
tensor = torch.rand(10)
|
|
val = [1, 2, 3]
|
|
st = create_sharded_tensor(rank=1, world_size=4, shards_per_rank=1)
|
|
state_dict = {"tensor": tensor, "value": val, "st": st}
|
|
plan = create_default_local_save_plan(state_dict, False)
|
|
self.assertEqual(2, len(plan.items))
|
|
wi = plan.items[0]
|
|
self.assertEqual(wi.index, MetadataIndex("tensor", [0]))
|
|
self.assertEqual(wi.type, WriteItemType.TENSOR)
|
|
self.assertEqual(wi.tensor_data.size, tensor.size())
|
|
self.assertEqual(
|
|
wi.tensor_data.properties,
|
|
TensorProperties.create_from_tensor(torch.zeros(1)),
|
|
)
|
|
self.assertEqual(wi.tensor_data.chunk.offsets, torch.Size([0]))
|
|
self.assertEqual(wi.tensor_data.chunk.sizes, torch.Size([10]))
|
|
|
|
st_wi = plan.items[1]
|
|
self.assertEqual(st_wi.index, MetadataIndex("st", [8]))
|
|
self.assertEqual(st_wi.type, WriteItemType.SHARD)
|
|
self.assertEqual(st_wi.tensor_data.size, st.size())
|
|
self.assertEqual(
|
|
st_wi.tensor_data.properties,
|
|
TensorProperties.create_from_tensor(torch.zeros(1)),
|
|
)
|
|
self.assertEqual(st_wi.tensor_data.chunk.offsets, torch.Size([8]))
|
|
self.assertEqual(st_wi.tensor_data.chunk.sizes, torch.Size([8]))
|
|
|
|
# Coordinator rank, should include replicated items as well
|
|
plan = create_default_local_save_plan(state_dict, True)
|
|
self.assertEqual(3, len(plan.items))
|
|
|
|
tensor_wi = next(wi for wi in plan.items if wi.type == WriteItemType.TENSOR)
|
|
self.assertEqual(tensor_wi.index, MetadataIndex("tensor", [0]))
|
|
self.assertEqual(tensor_wi.tensor_data.size, tensor.size())
|
|
self.assertEqual(
|
|
tensor_wi.tensor_data.properties,
|
|
TensorProperties.create_from_tensor(tensor),
|
|
)
|
|
self.assertEqual(tensor_wi.tensor_data.chunk.offsets, torch.Size([0]))
|
|
self.assertEqual(tensor_wi.tensor_data.chunk.sizes, torch.Size([10]))
|
|
|
|
bytes_wi = next(wi for wi in plan.items if wi.type == WriteItemType.BYTE_IO)
|
|
self.assertEqual(bytes_wi.index, MetadataIndex("value"))
|
|
self.assertIsNone(bytes_wi.tensor_data)
|
|
|
|
def test_global_plan(self):
|
|
def create_data(rank):
|
|
with with_dist(rank=rank, world_size=4):
|
|
tensor = torch.rand(10)
|
|
val = [1, 2, 3]
|
|
st = create_sharded_tensor(rank=rank, world_size=4, shards_per_rank=1)
|
|
state_dict = {"tensor": tensor, "value": val, "st": st}
|
|
return create_default_local_save_plan(state_dict, rank == 0)
|
|
|
|
all_plans = [create_data(0), create_data(1), create_data(2), create_data(3)]
|
|
all_plans = dedup_save_plans(all_plans)
|
|
final_plans, metadata = create_default_global_save_plan(all_plans=all_plans)
|
|
|
|
# The default global plan updates all indexes to include hints
|
|
for new_plan, old_plan in zip(final_plans, all_plans):
|
|
for new_item, old_item in zip(new_plan.items, old_plan.items):
|
|
self.assertEqual(new_item.index, old_item.index)
|
|
self.assertEqual(new_item.type, old_item.type)
|
|
self.assertEqual(new_item.tensor_data, old_item.tensor_data)
|
|
self.assertIn(new_item.index.fqn, metadata.state_dict_metadata)
|
|
|
|
item_md = metadata.state_dict_metadata[new_item.index.fqn]
|
|
if new_item.type == WriteItemType.BYTE_IO:
|
|
self.assertTrue(isinstance(item_md, BytesStorageMetadata))
|
|
else:
|
|
self.assertTrue(isinstance(item_md, TensorStorageMetadata))
|
|
self.assertEqual(item_md.size, old_item.tensor_data.size)
|
|
self.assertEqual(
|
|
item_md.properties, old_item.tensor_data.properties
|
|
)
|
|
|
|
self.assertIsNotNone(new_item.index.index)
|
|
# Make sure the hint is correct
|
|
self.assertEqual(
|
|
item_md.chunks[new_item.index.index], old_item.tensor_data.chunk
|
|
)
|
|
|
|
def test_local_load_plan(self):
|
|
def create_state_dict(rank):
|
|
with with_dist(rank=rank, world_size=4):
|
|
tensor = torch.rand(10)
|
|
val = [1, 2, 3]
|
|
st = create_sharded_tensor(rank=rank, world_size=4, shards_per_rank=1)
|
|
return {"tensor": tensor, "value": val, "st": st}
|
|
|
|
state_dict = create_state_dict(1)
|
|
metadata = _create_default_local_metadata(state_dict)
|
|
|
|
load_plan = create_default_local_load_plan(state_dict, metadata)
|
|
# This will create 3 entries
|
|
self.assertEqual(3, len(load_plan.items))
|
|
st_item = next(ri for ri in load_plan.items if ri.dest_index.fqn == "st")
|
|
tensor_item = next(
|
|
ri for ri in load_plan.items if ri.dest_index.fqn == "tensor"
|
|
)
|
|
bytes_item = next(ri for ri in load_plan.items if ri.dest_index.fqn == "value")
|
|
|
|
self.assertEqual(st_item.type, LoadItemType.TENSOR)
|
|
# This is an exact copy
|
|
self.assertEqual(st_item.dest_index, MetadataIndex("st", [8]))
|
|
self.assertEqual(st_item.dest_offsets, torch.Size([0]))
|
|
self.assertEqual(st_item.storage_index, MetadataIndex("st", [8]))
|
|
self.assertEqual(st_item.storage_offsets, torch.Size([0]))
|
|
self.assertEqual(st_item.lengths, torch.Size([8]))
|
|
|
|
self.assertEqual(tensor_item.type, LoadItemType.TENSOR)
|
|
self.assertEqual(tensor_item.dest_index, MetadataIndex("tensor", [0]))
|
|
self.assertEqual(tensor_item.dest_offsets, torch.Size([0]))
|
|
self.assertEqual(tensor_item.storage_index, MetadataIndex("tensor", [0]))
|
|
self.assertEqual(tensor_item.storage_offsets, torch.Size([0]))
|
|
self.assertEqual(tensor_item.lengths, torch.Size([10]))
|
|
|
|
self.assertEqual(bytes_item.type, LoadItemType.BYTE_IO)
|
|
self.assertEqual(bytes_item.dest_index, MetadataIndex("value"))
|
|
|
|
def test_load_with_resharding(self):
|
|
def create_state_dict(rank, world_size):
|
|
with with_dist(rank=rank, world_size=world_size):
|
|
return {
|
|
"st": create_sharded_tensor(
|
|
rank=rank,
|
|
world_size=world_size,
|
|
shards_per_rank=1,
|
|
shard_size=128 // world_size,
|
|
)
|
|
}
|
|
|
|
# Rank 1 has a 16 bytes shard from [16, 32[
|
|
world8_state_dict = create_state_dict(rank=1, world_size=8)
|
|
world8_metadata = _create_default_local_metadata(world8_state_dict)
|
|
|
|
# Rank 1 has a 32 bytes shard from [32, 64[
|
|
world4_state_dict = create_state_dict(rank=1, world_size=4)
|
|
world4_metadata = _create_default_local_metadata(world4_state_dict)
|
|
|
|
# First scenario, going from world=8 to world=4, need to load 2 shards
|
|
# Each 4-world shard has 32 elements, so it needs to load 2 shards
|
|
load_plan = create_default_local_load_plan(world4_state_dict, world8_metadata)
|
|
self.assertEqual(2, len(load_plan.items))
|
|
low_ri = next(
|
|
ri for ri in load_plan.items if ri.dest_offsets == torch.Size([0])
|
|
)
|
|
high_ri = next(
|
|
ri for ri in load_plan.items if ri.dest_offsets == torch.Size([16])
|
|
)
|
|
|
|
self.assertEqual(low_ri.storage_index, MetadataIndex("st", [32]))
|
|
self.assertEqual(low_ri.storage_offsets, torch.Size([0]))
|
|
self.assertEqual(low_ri.dest_index, MetadataIndex("st", [32]))
|
|
self.assertEqual(low_ri.dest_offsets, torch.Size([0]))
|
|
self.assertEqual(low_ri.lengths, torch.Size([16]))
|
|
|
|
self.assertEqual(high_ri.storage_index, MetadataIndex("st", [48]))
|
|
self.assertEqual(high_ri.storage_offsets, torch.Size([0]))
|
|
self.assertEqual(high_ri.dest_index, MetadataIndex("st", [32]))
|
|
self.assertEqual(high_ri.dest_offsets, torch.Size([16]))
|
|
self.assertEqual(high_ri.lengths, torch.Size([16]))
|
|
|
|
# Second scenario, going from world=4 to world=8, need to load half of 1 shard
|
|
# rank1 on 8-world needs to load the upper half of the rank0 4-world shard
|
|
load_plan = create_default_local_load_plan(world8_state_dict, world4_metadata)
|
|
self.assertEqual(1, len(load_plan.items))
|
|
ri = load_plan.items[0]
|
|
self.assertEqual(ri.storage_index, MetadataIndex("st", [0]))
|
|
self.assertEqual(ri.storage_offsets, torch.Size([16]))
|
|
self.assertEqual(ri.dest_index, MetadataIndex("st", [16]))
|
|
self.assertEqual(ri.dest_offsets, torch.Size([0]))
|
|
self.assertEqual(ri.lengths, torch.Size([16]))
|
|
|
|
def test_load_with_world_size_diff_by_one(self):
|
|
def create_state_dict(rank, world_size):
|
|
with with_dist(rank=rank, world_size=world_size):
|
|
return {
|
|
"st": create_sharded_tensor(
|
|
rank=rank,
|
|
world_size=world_size,
|
|
shards_per_rank=1,
|
|
shard_size=120 // world_size,
|
|
)
|
|
}
|
|
|
|
# rank 1 has a 30 bytes shard from [30, 60[
|
|
world4_state_dict = create_state_dict(rank=1, world_size=4)
|
|
world4_metadata = _create_default_local_metadata(world4_state_dict)
|
|
|
|
# rank 1 has a 40 bytes shard from [40, 80[
|
|
world3_state_dict = create_state_dict(rank=1, world_size=3)
|
|
|
|
load_plan = create_default_local_load_plan(world3_state_dict, world4_metadata)
|
|
self.assertEqual(2, len(load_plan.items))
|
|
# this is [30, 60] to load [40, 60]
|
|
low_ri = next(
|
|
ri for ri in load_plan.items if ri.dest_offsets == torch.Size([0])
|
|
)
|
|
# this is [60, 90] to load [60, 80]
|
|
high_ri = next(
|
|
ri for ri in load_plan.items if ri.dest_offsets == torch.Size([20])
|
|
)
|
|
|
|
self.assertEqual(low_ri.storage_index, MetadataIndex("st", [30]))
|
|
self.assertEqual(low_ri.storage_offsets, torch.Size([10]))
|
|
self.assertEqual(low_ri.dest_index, MetadataIndex("st", [40]))
|
|
self.assertEqual(low_ri.dest_offsets, torch.Size([0]))
|
|
self.assertEqual(low_ri.lengths, torch.Size([20]))
|
|
|
|
self.assertEqual(high_ri.storage_index, MetadataIndex("st", [60]))
|
|
self.assertEqual(high_ri.storage_offsets, torch.Size([0]))
|
|
self.assertEqual(high_ri.dest_index, MetadataIndex("st", [40]))
|
|
self.assertEqual(high_ri.dest_offsets, torch.Size([20]))
|
|
self.assertEqual(high_ri.lengths, torch.Size([20]))
|
|
|
|
|
|
class TestPlannerHelpers(TestCase):
|
|
def test_create_read_item_from_chunks(self):
|
|
tensor_md = TensorStorageMetadata(
|
|
properties=TensorProperties.create_from_tensor(torch.empty([16])),
|
|
size=torch.Size([16]),
|
|
chunks=[
|
|
ChunkStorageMetadata(offsets=torch.Size([0]), sizes=torch.Size([8])),
|
|
ChunkStorageMetadata(offsets=torch.Size([8]), sizes=torch.Size([8])),
|
|
],
|
|
)
|
|
|
|
chunk = ChunkStorageMetadata(offsets=torch.Size([4]), sizes=torch.Size([7]))
|
|
read_items = create_read_items_for_chunk_list("foo", tensor_md, [chunk])
|
|
|
|
self.assertEqual(2, len(read_items))
|
|
self.assertEqual(MetadataIndex("foo", [4]), read_items[0].dest_index)
|
|
self.assertEqual(torch.Size([0]), read_items[0].dest_offsets)
|
|
|
|
self.assertEqual(MetadataIndex("foo", [0]), read_items[0].storage_index)
|
|
self.assertEqual(torch.Size([4]), read_items[0].storage_offsets)
|
|
|
|
self.assertEqual(torch.Size([4]), read_items[0].lengths)
|
|
|
|
self.assertEqual(MetadataIndex("foo", [4]), read_items[1].dest_index)
|
|
self.assertEqual(torch.Size([4]), read_items[1].dest_offsets)
|
|
|
|
self.assertEqual(MetadataIndex("foo", [8]), read_items[1].storage_index)
|
|
self.assertEqual(torch.Size([0]), read_items[1].storage_offsets)
|
|
|
|
self.assertEqual(torch.Size([3]), read_items[1].lengths)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
run_tests()
|