import os
import unittest
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, SupportedDevices
from torch_npu.testing.common_distributed import skipIfUnsupportMultiNPU
class TestAllGatherBaseMm(TestCase):
@classmethod
def _init_dist_hccl(cls, rank, world_size):
os.environ['MASTER_ADDR'] = '127.0.0.1'
os.environ['MASTER_PORT'] = '50000'
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_npu_all_gather_base_mm(cls, rank, input_list):
x1_list, x2_list, x1_scale_list, x2_scale_list, world_size, comm_mode, output_dtype, init_pg, c2p = input_list
x1 = x1_list[rank]
x2 = x2_list[rank]
pg = init_pg(rank, world_size)
group = pg.distributed_c10d._get_default_group()
if torch.__version__ > '2.0':
hcom_name = group._get_backend(torch.device('npu')).get_hccl_comm_name(rank)
else:
hcom_name = group.get_hccl_comm_name(rank)
x1 = x1.npu()
x2 = x2.npu()
x1_scale = x1_scale_list[rank].npu() if x1_scale_list else None
x2_scale = x2_scale_list[rank].npu() if x2_scale_list else None
out, gather_out = torch_npu.npu_all_gather_base_mm(x1,
x2,
hcom_name,
world_size,
bias=None,
x1_scale=x1_scale,
x2_scale=x2_scale,
gather_index=0,
gather_output=True,
output_dtype=output_dtype,
comm_turn=0,
comm_mode=comm_mode)
c2p.put((rank, out.cpu().numpy(), gather_out.cpu().numpy()))
pg.barrier()
def _test_multiprocess(self, f, init_pg, input_list):
expt_out_list, expt_gather, x1, x2, x1_scale_list, x2_scale_list, world_size, comm_mode, output_dtype = input_list
ctx = mp.get_context('spawn')
c2p = ctx.Queue(world_size)
ps = []
for i in range(world_size):
p = ctx.Process(
target=f,
args=(i, [x1, x2, x1_scale_list, x2_scale_list, world_size, comm_mode, output_dtype, init_pg, c2p]))
p.start()
ps.append(p)
for _ in range(world_size):
rank, output, gather_output = c2p.get()
output, gather_output = torch.from_numpy(output), torch.from_numpy(gather_output)
self.assertEqual(output, expt_out_list[rank],
("rank {} Expect receive tensor {} but got {}.").format(rank, expt_out_list[rank], output))
self.assertEqual(gather_output, expt_gather,
("rank {} Expect receive tensor {} but got {}.").format(rank, expt_gather, gather_output))
for p in ps:
p.join()
def _construct_excepted_result(self, x1_list, x2_list, world_size, x1_scale_list=None, x2_scale_list=None, output_dtype=None):
gather_out = torch.cat(x1_list)
if x1_scale_list:
x1_scale = torch.cat(x1_scale_list)
out_list = []
if output_dtype:
out_dtype = output_dtype
else:
out_dtype = gather_out.dtype
for i in range(world_size):
gather_out_npu, x2_list_npu = gather_out.npu(), x2_list[i].npu()
if x1_scale_list:
mm_res = torch_npu.npu_quant_matmul(x1=gather_out_npu, x2=x2_list_npu, scale=x2_scale_list[i].squeeze(0).npu(), pertoken_scale=x1_scale.squeeze(-1).npu(), output_dtype=out_dtype)
elif x2_scale_list:
mm_res = torch_npu.npu_quant_matmul(x1=gather_out_npu, x2=x2_list_npu, scale=x2_scale_list[i].squeeze(0).npu(), output_dtype=out_dtype)
else:
mm_res = torch.matmul(gather_out_npu, x2_list_npu)
out_list.append(mm_res.to(out_dtype).cpu())
return out_list, gather_out_npu.cpu()
@skipIfUnsupportMultiNPU(8)
@SupportedDevices(['Ascend910B'])
def test_npu_all_gather_base_mm(self):
world_size = 8
dtype = np.float16
data_format = -1
x1_shape = [dtype, data_format, [16, 512]]
x2_shape = [dtype, data_format, [512, 256]]
x1_list = []
x2_list = []
for _ in range(world_size):
x1, _ = create_common_tensor(x1_shape, -1, 1)
x2, _ = create_common_tensor(x2_shape, -1, 1)
x1_list.append(x1)
x2_list.append(x2)
expt_out_list, expt_gather = self._construct_excepted_result(x1_list, x2_list, world_size)
for comm_mode in ['aiv', 'ai_cpu']:
self._test_multiprocess(TestAllGatherBaseMm._test_npu_all_gather_base_mm,
TestAllGatherBaseMm._init_dist_hccl, [expt_out_list, expt_gather, x1_list, x2_list, None, None, world_size, comm_mode, None])
@skipIfUnsupportMultiNPU(8)
@SupportedDevices(['Ascend910B'])
def test_npu_all_gather_quant_mm(self):
world_size = 8
m, k, n = 16, 512, 256
output_dtype = torch.float16
x1_list = []
x2_list = []
x1_scale_list = []
x2_scale_list = []
for _ in range(world_size):
x1 = torch.randint(-10, 10, size=(m, k), dtype=torch.int8)
x2 = torch.randint(-10, 10, size=(k, n), dtype=torch.int8)
x1_scale = torch.randn((m, 1), dtype=torch.float32)
x2_scale = torch.randn((1, n), dtype=torch.float32)
x1_list.append(x1)
x2_list.append(x2)
x1_scale_list.append(x1_scale)
x2_scale_list.append(x2_scale)
expt_out_list, expt_gather = self._construct_excepted_result(x1_list, x2_list, world_size, x1_scale_list, x2_scale_list, output_dtype)
self._test_multiprocess(TestAllGatherBaseMm._test_npu_all_gather_base_mm,
TestAllGatherBaseMm._init_dist_hccl, [expt_out_list, expt_gather, x1_list, x2_list, x1_scale_list, x2_scale_list, world_size, 'aiv', output_dtype])
if __name__ == '__main__':
run_tests()