def collective_fn(input_tensor, output_tensor, context):
size = gloo_util.get_tensor_n_elements(input_tensor)
world_size = self._gloo_context.size
pygloo.reduce_scatter(
context, gloo_util.get_tensor_ptr(input_tensor),
gloo_util.get_tensor_ptr(output_tensor), size,
[size // world_size for _ in range(world_size)],
gloo_util.get_gloo_tensor_dtype(output_tensor),
gloo_util.get_gloo_reduce_op(reducescatter_options.reduceOp))
def test_reduce_scatter(rank, world_size, fileStore_path):
'''
rank # Rank of this process within list of participating processes
world_size # Number of participating processes
'''
if rank == 0:
if os.path.exists(fileStore_path):
shutil.rmtree(fileStore_path)
os.makedirs(fileStore_path)
else:
time.sleep(0.5)
context = pygloo.rendezvous.Context(rank, world_size)
attr = pygloo.transport.tcp.attr("localhost")
# Perform rendezvous for TCP pairs
dev = pygloo.transport.tcp.CreateDevice(attr)
fileStore = pygloo.rendezvous.FileStore(fileStore_path)
store = pygloo.rendezvous.PrefixStore(str(world_size), fileStore)
context.connectFullMesh(store, dev)
sendbuf = np.array(
[i + 1 for i in range(sum([j + 1 for j in range(world_size)]))],
dtype=np.float32)
sendptr = sendbuf.ctypes.data
recvbuf = np.zeros((rank + 1, ), dtype=np.float32)
recvptr = recvbuf.ctypes.data
recvElems = [i + 1 for i in range(world_size)]
# sendbuf = torch.Tensor([i+1 for i in range(sum([j+1 for j in range(world_size)]))]).float()
# sendptr = sendbuf.data_ptr()
# recvbuf = torch.zeros(rank+1).float()
# recvptr = recvbuf.data_ptr()
data_size = sendbuf.size if isinstance(
sendbuf, np.ndarray) else sendbuf.numpy().size
datatype = pygloo.glooDataType_t.glooFloat32
op = pygloo.ReduceOp.SUM
pygloo.reduce_scatter(context, sendptr, recvptr, data_size, recvElems,
datatype, op)
print(f"rank {rank} sends {sendbuf}, receives {recvbuf}")