def __getstate__(self):
pg_state = ShardedTensor.ProcessGroupState(
distributed_c10d.get_rank(self._process_group),
distributed_c10d.get_rank(),
distributed_c10d.get_world_size(self._process_group),
distributed_c10d.get_world_size(),
)
return self._local_shards, self._metadata, pg_state, self._sharding_spec, self._init_rrefs
def allreduce(data: torch.Tensor):
""" return the chunk
"""
c10d.all_reduce(data.view(-1))
rank = c10d.get_rank()
world_size = c10d.get_world_size()
chunk_sz = int(data.numel() / world_size)
return data.view(-1).narrow(0, rank * chunk_sz, chunk_sz)
def reduce_scatter(data: torch.Tensor):
data = data.view(-1)
rank = c10d.get_rank()
world_size = c10d.get_world_size()
chunk_sz = int(data.numel() / world_size)
offset = rank * chunk_sz
chunk = data.narrow(0, offset, chunk_sz)
c10d._reduce_scatter_base(chunk, data)
return chunk
def __setstate__(self, state):
self._sharded_tensor_id = None
if not distributed_c10d.is_initialized():
raise RuntimeError(
'Need to initialize default process group using '
'"init_process_group" before loading ShardedTensor')
self._local_shards, self._metadata, pg_state, self._sharding_spec, self._init_rrefs = state
# Setup process group
global _CURRENT_PROCESS_GROUP
if _CURRENT_PROCESS_GROUP is None:
self._process_group = distributed_c10d._get_default_group()
else:
self._process_group = _CURRENT_PROCESS_GROUP
# Validate process group.
local_rank = distributed_c10d.get_rank(self._process_group)
if pg_state.local_rank != local_rank:
raise RuntimeError(
f'Local rank at save time was {pg_state.local_rank}, but at '
f'load time was {local_rank}')
global_rank = distributed_c10d.get_rank()
if pg_state.global_rank != global_rank:
raise RuntimeError(
f'Global rank at save time was {pg_state.global_rank}, but at '
f'load time was {global_rank}')
local_world_size = distributed_c10d.get_world_size(self._process_group)
if pg_state.local_world_size != local_world_size:
raise RuntimeError(
f'Local world size at save time was {pg_state.local_world_size}, '
f'but at load time was {local_world_size}')
global_world_size = distributed_c10d.get_world_size()
if pg_state.global_world_size != global_world_size:
raise RuntimeError(
f'Global world size at save time was {pg_state.global_world_size}, '
f'but at load time was {global_world_size}')
self._post_init()
def fidelity_check_hierarchy_allgather(local_rank, inter_shard_group,
intra_shard_group):
""""""
torch.manual_seed(10)
n = 10
dtype = torch.half
partition_nelem = 10e6 # 10M elements
world_size = c10d.get_world_size()
rank = c10d.get_rank()
nelem = int(partition_nelem * world_size)
# all ranks suppose to have same
test_tensors = []
for _ in range(n):
test_tensors.append(
torch.rand(nelem, dtype=dtype, device=f'cuda:{local_rank}'))
# get inputs and outputs
outputs_for_vanila, inputs_for_vanila = _get_outputs_inputs(
test_tensors, rank, world_size)
c10d._all_gather_base_coalesced(outputs_for_vanila, inputs_for_vanila)
for i, j in zip(outputs_for_vanila, test_tensors):
print(f"vanila all close {torch.allclose(i, j)}")
# torch.cuda.synchronize()
comm_stream = torch.cuda.Stream()
outputs_for_hierarchy, inputs_for_hierarchy = _get_outputs_inputs(
test_tensors, rank, world_size)
# with torch.cuda.stream(comm_stream):
# with torch.cuda.stream(torch.cuda.default_stream()):
with torch.cuda.stream(torch.cuda.current_stream()):
handle = _hierarchy_allgather(outputs_for_hierarchy,
inputs_for_hierarchy, intra_shard_group,
inter_shard_group, world_size,
local_rank)
handle.wait()
for i, j in zip(outputs_for_hierarchy, test_tensors):
print(f"hierarchy all close {torch.allclose(i, j)}")
def main():
""""""
arg_parser = argparse.ArgumentParser()
arg_parser.add_argument('--data-type', default='bfloat16', type=str)
arg_parser.add_argument('--local_rank', type=int)
args = arg_parser.parse_args()
c10d.init_process_group(backend='nccl')
if args.data_type == 'bfloat16':
data_type = torch.bfloat16
elif args.data_type in ('float16', 'fp16'):
data_type = torch.half
elif args.data_type in ('float32', 'fp32'):
data_type = torch.float32
else:
raise RuntimeError(f"unsupported data type {args.data_type}")
rank = c10d.get_rank()
before_sync_data = data[rank]
for_reduce_scatter = torch.tensor(before_sync_data,
device=f'cuda:{rank}',
dtype=data_type)
for_allreduce = torch.tensor(before_sync_data,
device=f'cuda:{rank}',
dtype=data_type)
rc_part = reduce_scatter(for_reduce_scatter)
ar_part = allreduce(for_allreduce)
errors = []
if rank == 7:
for i, j in zip(
rc_part.to(torch.float32).cpu().numpy().tolist(),
ar_part.to(torch.float32).cpu().numpy().tolist()):
# print(i, j, i - j)
errors.append(i - j)
print(f'error range [{np.min(errors)}, {np.max(errors)}]')
diff_norm = torch.norm(rc_part - ar_part)
print(f'norm diff {diff_norm}')
def print_at_rank0(msg):
if c10d.get_rank() == 0:
print(msg)
