import unittest
import os
import numpy as np
import torch
import torch.distributed as dist
import torch.multiprocessing as mp
import torch_npu
from torch_npu.testing.testcase import TestCase, run_tests
from torch_npu.testing.common_utils import create_common_tensor
from torch_npu.testing.common_distributed import skipIfUnsupportMultiNPU
class HcomAllReduceTest(TestCase):
@classmethod
def _init_dist_hccl(cls, rank, world_size):
os.environ['MASTER_ADDR'] = '127.0.0.1'
os.environ['MASTER_PORT'] = '29500'
os.environ['HCCL_WHITELIST_DISABLE'] = '1'
torch_npu.npu.set_device(rank)
dist.init_process_group(backend='hccl', world_size=world_size, rank=rank)
return dist
@classmethod
def _test_all_reduce(cls, rank, input1, world_size, init_pg, c2p, reduce_op=dist.ReduceOp.SUM, done_event=None):
dist_group = init_pg(rank, world_size)
dst = 0
input1 = input1.npu()
dist_group.all_reduce(input1, reduce_op)
c2p.put((rank, dst, input1.cpu()))
if done_event is not None:
done_event.wait(timeout=5)
@classmethod
def _test_all_reduce_with_internal_format_and_offset(cls, rank, input1, world_size, init_pg, reduce_op=dist.ReduceOp.SUM):
torch_npu.npu.config.allow_internal_format = True
dist_group = init_pg(rank, world_size)
dst = 0
first_dim = input1.shape[0]
other_dims = input1.shape[1:]
input2 = torch_npu.npu_format_cast(input1.repeat(2, *[1 for i in other_dims]).npu(), 29)
test_case = TestCase()
error_expect = "For a tensor of internal format, it's storage_offset must be 0"
with test_case.assertRaisesRegex(RuntimeError, error_expect):
dist_group.all_reduce(input2[first_dim:], reduce_op)
def _test_multiprocess_with_error(self, f, init_pg, input1, world_size, reduce_op=dist.ReduceOp.SUM):
ctx = mp.get_context('spawn')
ps = []
for i in range(world_size):
p = ctx.Process(
target=f,
args=(i, input1.cpu(), world_size, init_pg, reduce_op))
p.start()
ps.append(p)
for p in ps:
p.join()
self.assertEqual(p.exitcode, 0, "subprocess exit with abnormal code.")
def _test_multiprocess(self, f, init_pg, expected, input1, world_size, reduce_op=dist.ReduceOp.SUM):
ctx = mp.get_context('spawn')
c2p = ctx.Queue(world_size)
done_event = ctx.Event()
ps = []
for i in range(world_size):
p = ctx.Process(
target=f,
args=(i, input1.cpu(), world_size, init_pg, c2p, reduce_op, done_event))
p.start()
ps.append(p)
for _ in range(world_size):
rank, dst, output = c2p.get()
if rank == dst:
self.assertEqual(output, expected,
"rank {} world_size {} dtype {} shape {} Expect receive tensor {} but got {}.".format(
rank, world_size, expected.dtype, expected.shape, expected, output))
done_event.set()
for p in ps:
p.join()
def _construct_excepted_result(self, inputs, world_size, dtype=np.float32, reduce_op=dist.ReduceOp.SUM):
expected = 0
for _ in range(world_size):
expected += inputs
if reduce_op == dist.ReduceOp.AVG:
if dtype in [np.int32, np.int8, np.int64, np.uint8]:
expected //= world_size
else:
expected /= world_size
return expected
@skipIfUnsupportMultiNPU(2)
def test_dist_all_reduce(self):
ranks = [2]
dtype_list = [np.float32]
format_list = [0, 2, 3]
shape_format = [
[i, j, [2, 3, 16]] for i in dtype_list for j in format_list
]
for world_size in ranks:
for shape in shape_format:
exp_input, input1 = create_common_tensor(shape, -10, 10)
expected = self._construct_excepted_result(exp_input, world_size)
self._test_multiprocess(HcomAllReduceTest._test_all_reduce,
HcomAllReduceTest._init_dist_hccl, expected, input1, world_size)
@skipIfUnsupportMultiNPU(2)
def test_dist_all_reduce_int64(self):
ranks = [2]
shape_format = [np.int64, 2, [1]]
for world_size in ranks:
exp_input, input1 = create_common_tensor(shape_format, -10, 10)
expected = self._construct_excepted_result(exp_input, world_size)
self._test_multiprocess(HcomAllReduceTest._test_all_reduce,
HcomAllReduceTest._init_dist_hccl, expected, input1, world_size)
@skipIfUnsupportMultiNPU(2)
def test_dist_all_reduce_avg(self):
ranks = [2]
dtype_list = [np.int32]
shape_format = [
[i, 2, [3, 16]] for i in dtype_list
]
for world_size in ranks:
for shape in shape_format:
exp_input, input1 = create_common_tensor(shape, -10, 10)
expected = self._construct_excepted_result(exp_input, world_size, shape[0], dist.ReduceOp.AVG)
self._test_multiprocess(HcomAllReduceTest._test_all_reduce,
HcomAllReduceTest._init_dist_hccl, expected, input1, world_size,
dist.ReduceOp.AVG)
@skipIfUnsupportMultiNPU(2)
def test_dist_all_reduce_uint8_avg(self):
ranks = [2]
shape_format = [[np.uint8, 2, [3, 16]]]
for world_size in ranks:
for shape in shape_format:
exp_input, input1 = create_common_tensor(shape, 0, 10)
expected = self._construct_excepted_result(exp_input, world_size, shape[0], dist.ReduceOp.AVG)
self._test_multiprocess(HcomAllReduceTest._test_all_reduce,
HcomAllReduceTest._init_dist_hccl, expected, input1, world_size,
dist.ReduceOp.AVG)
@skipIfUnsupportMultiNPU(2)
def test_dist_all_reduce_with_internal_format_and_offset(self):
print("in case")
ranks = [2]
shape_format = [[np.float32, 2, [16, 16]]]
for world_size in ranks:
for shape in shape_format:
exp_input, input1 = create_common_tensor(shape, 0, 10)
self._test_multiprocess_with_error(HcomAllReduceTest._test_all_reduce_with_internal_format_and_offset,
HcomAllReduceTest._init_dist_hccl, input1, world_size,
dist.ReduceOp.AVG)
if __name__ == '__main__':
run_tests()