v2.9.0 enable internal format by default for hccl tests on A3-pytorch-AtomGit

Star854

Star854

已合并

v2.9.0 enable internal format by default for hccl tests on A3 #27477

freezee创建于 2025年12月4日从freezee/pytorch： v2.9.0 合入到 Ascend/pytorch： v2.9.0

v2.9.0 enable internal format by default for hccl tests on A3 #27477

已合并

freezee创建于 2025年12月4日

从freezee/pytorch v2.9.0合入到Ascend/pytorch v2.9.0

共 9 个文件变更+14-0

Mtest/distributed/test_all_to_all.py+2-0

	@@ -64,6 +64,7 @@ class HcclAlltoAllTest(TestCase):
64		64
65	@classmethod	65	@classmethod
66	def _test_alltoall_2p_size_with_input_internal_format_and_offset(cls, rank, world_size, init_pg):	66	def _test_alltoall_2p_size_with_input_internal_format_and_offset(cls, rank, world_size, init_pg):
		67	+ torch_npu.npu.config.allow_internal_format = True
67	pg = init_pg(rank, world_size)	68	pg = init_pg(rank, world_size)
68	input_list = [(torch.zeros(rank + 1 + world_size, rank + 1) + rank).float().npu() for i in range(2)]	69	input_list = [(torch.zeros(rank + 1 + world_size, rank + 1) + rank).float().npu() for i in range(2)]
69	input_list = [torch_npu.npu_format_cast(i, 29)[world_size:] for i in input_list]	70	input_list = [torch_npu.npu_format_cast(i, 29)[world_size:] for i in input_list]
	@@ -75,6 +76,7 @@ class HcclAlltoAllTest(TestCase):
75		76
76	@classmethod	77	@classmethod
77	def _test_alltoall_2p_size_with_output_internal_format_and_offset(cls, rank, world_size, init_pg):	78	def _test_alltoall_2p_size_with_output_internal_format_and_offset(cls, rank, world_size, init_pg):
		79	+ torch_npu.npu.config.allow_internal_format = True
78	pg = init_pg(rank, world_size)	80	pg = init_pg(rank, world_size)
79	input_list = [(torch.zeros(rank + 1, rank + 1) + rank).float().npu() for i in range(2)]	81	input_list = [(torch.zeros(rank + 1, rank + 1) + rank).float().npu() for i in range(2)]
80	output_list = [torch.empty(i + 1 + world_size, i + 1).float().npu() for i in range(2)]	82	output_list = [torch.empty(i + 1 + world_size, i + 1).float().npu() for i in range(2)]

Mtest/distributed/test_allgather.py+2-0

	@@ -119,6 +119,7 @@ class HcclAllGatherTest(HcclAllGatherTestBase):
119		119
120	@classmethod	120	@classmethod
121	def _test_all_gather_with_input_internal_format_and_offset(cls, rank, input1, world_size, init_pg):	121	def _test_all_gather_with_input_internal_format_and_offset(cls, rank, input1, world_size, init_pg):
		122	+ torch_npu.npu.config.allow_internal_format = True
122	pg = init_pg(rank, world_size)	123	pg = init_pg(rank, world_size)
123	first_dim = input1.shape[0]	124	first_dim = input1.shape[0]
124	other_dims = input1.shape[1:]	125	other_dims = input1.shape[1:]
	@@ -132,6 +133,7 @@ class HcclAllGatherTest(HcclAllGatherTestBase):
132		133
133	@classmethod	134	@classmethod
134	def _test_all_gather_with_output_internal_format_and_offset(cls, rank, input1, world_size, init_pg):	135	def _test_all_gather_with_output_internal_format_and_offset(cls, rank, input1, world_size, init_pg):
		136	+ torch_npu.npu.config.allow_internal_format = True
135	pg = init_pg(rank, world_size)	137	pg = init_pg(rank, world_size)
136	first_dim = input1.shape[0]	138	first_dim = input1.shape[0]
137	other_dims = input1.shape[1:]	139	other_dims = input1.shape[1:]

Mtest/distributed/test_allgather_into_tensor.py+2-0

	@@ -31,6 +31,7 @@ class HcclAllGatherIntoTensorTest(HcclAllGatherTestBase):
31		31
32	@classmethod	32	@classmethod
33	def _test_all_gather_into_tensor_with_input_internal_format_and_offset(cls, rank, input1, world_size, init_pg):	33	def _test_all_gather_into_tensor_with_input_internal_format_and_offset(cls, rank, input1, world_size, init_pg):
		34	+ torch_npu.npu.config.allow_internal_format = True
34	pg = init_pg(rank, world_size)	35	pg = init_pg(rank, world_size)
35	input1 = input1.npu()	36	input1 = input1.npu()
36	first_dim = input1.shape[0]	37	first_dim = input1.shape[0]
	@@ -46,6 +47,7 @@ class HcclAllGatherIntoTensorTest(HcclAllGatherTestBase):
46		47
47	@classmethod	48	@classmethod
48	def _test_all_gather_into_tensor_with_output_internal_format_and_offset(cls, rank, input1, world_size, init_pg):	49	def _test_all_gather_into_tensor_with_output_internal_format_and_offset(cls, rank, input1, world_size, init_pg):
		50	+ torch_npu.npu.config.allow_internal_format = True
49	pg = init_pg(rank, world_size)	51	pg = init_pg(rank, world_size)
50	input1 = input1.npu()	52	input1 = input1.npu()
51	shape = list(input1.size())	53	shape = list(input1.size())

Mtest/distributed/test_allreduce.py+1-0

	@@ -37,6 +37,7 @@ class HcomAllReduceTest(TestCase):
37	@classmethod	37	@classmethod
38	# pylint:disable=huawei-too-many-arguments	38	# pylint:disable=huawei-too-many-arguments
39	def _test_all_reduce_with_internal_format_and_offset(cls, rank, input1, world_size, init_pg, reduce_op=dist.ReduceOp.SUM):	39	def _test_all_reduce_with_internal_format_and_offset(cls, rank, input1, world_size, init_pg, reduce_op=dist.ReduceOp.SUM):
		40	+ torch_npu.npu.config.allow_internal_format = True
40	dist_group = init_pg(rank, world_size)	41	dist_group = init_pg(rank, world_size)
41	dst = 0	42	dst = 0
42	first_dim = input1.shape[0]	43	first_dim = input1.shape[0]

Mtest/distributed/test_batch_isend_irecv.py+1-0

	@@ -46,6 +46,7 @@ class HcomBatchIsendIrecvTest(TestCase):
46		46
47	@classmethod	47	@classmethod
48	def _test_batch_isend_irecv_with_internal_format_and_offset(cls, rank, world_size, init_pg):	48	def _test_batch_isend_irecv_with_internal_format_and_offset(cls, rank, world_size, init_pg):
		49	+ torch_npu.npu.config.allow_internal_format = True
49	_ = init_pg(rank, world_size)	50	_ = init_pg(rank, world_size)
50	recv_tensors = [None for _ in range(world_size)]	51	recv_tensors = [None for _ in range(world_size)]
51	p2p_op_list = []	52	p2p_op_list = []

Mtest/distributed/test_broadcast.py+1-0

	@@ -25,6 +25,7 @@ class HcclBroadcastTest(TestCase):
25		25
26	@classmethod	26	@classmethod
27	def _test_broadcast_with_internal_format_and_offset(cls, rank, input1, world_size, init_pg):	27	def _test_broadcast_with_internal_format_and_offset(cls, rank, input1, world_size, init_pg):
		28	+ torch_npu.npu.config.allow_internal_format = True
28	pg = init_pg(rank, world_size)	29	pg = init_pg(rank, world_size)
29	first_dim = input1.shape[0]	30	first_dim = input1.shape[0]
30	other_dims = input1.shape[1:]	31	other_dims = input1.shape[1:]

Mtest/distributed/test_reduce_scatter.py+2-0

	@@ -85,6 +85,7 @@ class HcclReduceScatterTest(HcclReduceScatterTestBase):
85	@classmethod	85	@classmethod
86	# pylint:disable=huawei-too-many-arguments	86	# pylint:disable=huawei-too-many-arguments
87	def _test_reduce_scatter_with_input_internal_format_and_offset(cls, rank, input_list, world_size, init_pg):	87	def _test_reduce_scatter_with_input_internal_format_and_offset(cls, rank, input_list, world_size, init_pg):
		88	+ torch_npu.npu.config.allow_internal_format = True
88	pg = init_pg(rank, world_size)	89	pg = init_pg(rank, world_size)
89	input_list_npu = []	90	input_list_npu = []
90	for inp in input_list:	91	for inp in input_list:
	@@ -101,6 +102,7 @@ class HcclReduceScatterTest(HcclReduceScatterTestBase):
101	@classmethod	102	@classmethod
102	# pylint:disable=huawei-too-many-arguments	103	# pylint:disable=huawei-too-many-arguments
103	def _test_reduce_scatter_with_output_internal_format_and_offset(cls, rank, input_list, world_size, init_pg):	104	def _test_reduce_scatter_with_output_internal_format_and_offset(cls, rank, input_list, world_size, init_pg):
		105	+ torch_npu.npu.config.allow_internal_format = True
104	pg = init_pg(rank, world_size)	106	pg = init_pg(rank, world_size)
105	input_list_npu = [input.npu() for input in input_list]	107	input_list_npu = [input.npu() for input in input_list]
106	output = torch.empty_like(input_list_npu[rank])	108	output = torch.empty_like(input_list_npu[rank])

Mtest/distributed/test_reduce_scatter_tensor.py+2-0

	@@ -31,6 +31,7 @@ class HcclReduceScatterTensorTest(HcclReduceScatterTestBase):
31	@classmethod	31	@classmethod
32	# pylint:disable=huawei-too-many-arguments	32	# pylint:disable=huawei-too-many-arguments
33	def _test_reduce_scatter_tensor_with_input_internal_format_and_offset(cls, rank, input_list, world_size, init_pg):	33	def _test_reduce_scatter_tensor_with_input_internal_format_and_offset(cls, rank, input_list, world_size, init_pg):
		34	+ torch_npu.npu.config.allow_internal_format = True
34	pg = init_pg(rank, world_size)	35	pg = init_pg(rank, world_size)
35	input_list_npu = [input.npu() for input in input_list]	36	input_list_npu = [input.npu() for input in input_list]
36	input_tensor = torch.cat(input_list_npu)	37	input_tensor = torch.cat(input_list_npu)
	@@ -46,6 +47,7 @@ class HcclReduceScatterTensorTest(HcclReduceScatterTestBase):
46	@classmethod	47	@classmethod
47	# pylint:disable=huawei-too-many-arguments	48	# pylint:disable=huawei-too-many-arguments
48	def _test_reduce_scatter_tensor_with_output_internal_format_and_offset(cls, rank, input_list, world_size, init_pg):	49	def _test_reduce_scatter_tensor_with_output_internal_format_and_offset(cls, rank, input_list, world_size, init_pg):
		50	+ torch_npu.npu.config.allow_internal_format = True
49	pg = init_pg(rank, world_size)	51	pg = init_pg(rank, world_size)
50	input_list_npu = [input.npu() for input in input_list]	52	input_list_npu = [input.npu() for input in input_list]
51	input_tensor = torch.cat(input_list_npu)	53	input_tensor = torch.cat(input_list_npu)

Mtest/distributed/test_send_recv.py+1-0

	@@ -48,6 +48,7 @@ class HcclSendRecvDistTest(TestCase):
48		48
49	@classmethod	49	@classmethod
50	def _test_send_recv_dist_with_internal_format_and_offset(cls, rank, shared_tensors, world_size, init_pg):	50	def _test_send_recv_dist_with_internal_format_and_offset(cls, rank, shared_tensors, world_size, init_pg):
		51	+ torch_npu.npu.config.allow_internal_format = True
51	pg = init_pg(rank, world_size)	52	pg = init_pg(rank, world_size)
52	first_dim = shared_tensors.shape[0]	53	first_dim = shared_tensors.shape[0]
53	other_dims = shared_tensors.shape[1:]	54	other_dims = shared_tensors.shape[1:]