Star847
1325
代码
代码
Issues1502
Pull Requests2268
流水线
项目讨论
Wiki
分析
Star847
1325
已合并
test(distributed): add test for FileStore,Store,TCPStore #31911
Nice try创建于 3月17日
test(distributed): add test for FileStore,Store,TCPStore #31911
已合并
Nice try创建于 3月17日
从Nice_try/pytorchv2.7.1-apt合入到Ascend/pytorchv2.7.1
共 5 个文件变更+201-0
Mdocs/zh/native_apis/pytorch_2-6-0/torch-distributed.md+16-0
@@ -22,6 +22,7 @@
22|torch.distributed.HashStore|是|-|22|torch.distributed.HashStore|是|-|
23|torch.distributed.FileStore|是|-|23|torch.distributed.FileStore|是|-|
24|torch.distributed.PrefixStore|是|-|24|torch.distributed.PrefixStore|是|-|
25+|torch.distributed.Store.\_\_init\_\_|是|-|
25|torch.distributed.Store.set|是|-|26|torch.distributed.Store.set|是|-|
26|torch.distributed.Store.get|是|-|27|torch.distributed.Store.get|是|-|
27|torch.distributed.Store.add|是|-|28|torch.distributed.Store.add|是|-|
@@ -30,6 +31,21 @@
30|torch.distributed.Store.num_keys|是|-|31|torch.distributed.Store.num_keys|是|-|
31|torch.distributed.Store.delete_key|是|-|32|torch.distributed.Store.delete_key|是|-|
32|torch.distributed.Store.set_timeout|是|-|33|torch.distributed.Store.set_timeout|是|-|
34+|torch.distributed.Store.append|是|-|
35+|torch.distributed.Store.check|是|-|
36+|torch.distributed.Store.has_extended_api|是|-|
37+|torch.distributed.Store.multi_set|是|-|
38+|torch.distributed.Store.multi_get|是|-|
39+|torch.distributed.Store.timeout|是|-|
40+|torch.distributed.TCPStore.\_\_init\_\_|是|-|
41+|torch.distributed.TCPStore.host|是|-|
42+|torch.distributed.TCPStore.libuvBackend|是|-|
43+|torch.distributed.TCPStore.port|是|-|
44+|torch.distributed.HashStore.\_\_init\_\_|是|-|
45+|torch.distributed.FileStore.\_\_init\_\_|是|-|
46+|torch.distributed.FileStore.path|是|-|
47+|torch.distributed.PrefixStore.\_\_init\_\_|是|-|
48+|torch.distributed.PrefixStore.underlying_store|是|-|
33|torch.distributed.new_group|是|当pg_options函数传入类型为torch_npu._C._distributed_c10d.ProcessGroupHCCL.Options()时，配置该变量属性hccl_config可控制HCCL通信域缓存区大小。具体示例可参考《PyTorch 训练模型迁移调优指南》的“hccl_buffer_size”章节，配置变量属性hccl_config的group_name字段可以设置HCCL通信域的通信组自定义名称，取值为长度不超过32的字符串。|49|torch.distributed.new_group|是|当pg_options函数传入类型为torch_npu._C._distributed_c10d.ProcessGroupHCCL.Options()时，配置该变量属性hccl_config可控制HCCL通信域缓存区大小。具体示例可参考《PyTorch 训练模型迁移调优指南》的“hccl_buffer_size”章节，配置变量属性hccl_config的group_name字段可以设置HCCL通信域的通信组自定义名称，取值为长度不超过32的字符串。|
34|torch.distributed.get_group_rank|是|-|50|torch.distributed.get_group_rank|是|-|
35|torch.distributed.get_global_rank|是|-|51|torch.distributed.get_global_rank|是|-|
Mdocs/zh/native_apis/pytorch_2-7-1/torch-distributed.md+16-0
@@ -22,6 +22,7 @@
22|torch.distributed.HashStore|是|-|22|torch.distributed.HashStore|是|-|
23|torch.distributed.FileStore|是|-|23|torch.distributed.FileStore|是|-|
24|torch.distributed.PrefixStore|是|-|24|torch.distributed.PrefixStore|是|-|
25+|torch.distributed.Store.\_\_init\_\_|是|-|
25|torch.distributed.Store.set|是|-|26|torch.distributed.Store.set|是|-|
26|torch.distributed.Store.get|是|-|27|torch.distributed.Store.get|是|-|
27|torch.distributed.Store.add|是|-|28|torch.distributed.Store.add|是|-|
@@ -30,6 +31,21 @@
30|torch.distributed.Store.num_keys|是|-|31|torch.distributed.Store.num_keys|是|-|
31|torch.distributed.Store.delete_key|是|-|32|torch.distributed.Store.delete_key|是|-|
32|torch.distributed.Store.set_timeout|是|-|33|torch.distributed.Store.set_timeout|是|-|
34+|torch.distributed.Store.append|是|-|
35+|torch.distributed.Store.check|是|-|
36+|torch.distributed.Store.has_extended_api|是|-|
37+|torch.distributed.Store.multi_set|是|-|
38+|torch.distributed.Store.multi_get|是|-|
39+|torch.distributed.Store.timeout|是|-|
40+|torch.distributed.TCPStore.\_\_init\_\_|是|-|
41+|torch.distributed.TCPStore.host|是|-|
42+|torch.distributed.TCPStore.libuvBackend|是|-|
43+|torch.distributed.TCPStore.port|是|-|
44+|torch.distributed.HashStore.\_\_init\_\_|是|-|
45+|torch.distributed.FileStore.\_\_init\_\_|是|-|
46+|torch.distributed.FileStore.path|是|-|
47+|torch.distributed.PrefixStore.\_\_init\_\_|是|-|
48+|torch.distributed.PrefixStore.underlying_store|是|-|
33|torch.distributed.new_group|是|当pg_options函数传入类型为torch_npu._C._distributed_c10d.ProcessGroupHCCL.Options()时，配置该变量属性hccl_config可控制HCCL通信域缓存区大小。具体示例可参考《PyTorch 训练模型迁移调优指南》的“hccl_buffer_size”章节，配置变量属性hccl_config的group_name字段可以设置HCCL通信域的通信组自定义名称，取值为长度不超过32的字符串。|49|torch.distributed.new_group|是|当pg_options函数传入类型为torch_npu._C._distributed_c10d.ProcessGroupHCCL.Options()时，配置该变量属性hccl_config可控制HCCL通信域缓存区大小。具体示例可参考《PyTorch 训练模型迁移调优指南》的“hccl_buffer_size”章节，配置变量属性hccl_config的group_name字段可以设置HCCL通信域的通信组自定义名称，取值为长度不超过32的字符串。|
34|torch.distributed.get_group_rank|是|-|50|torch.distributed.get_group_rank|是|-|
35|torch.distributed.get_global_rank|是|-|51|torch.distributed.get_global_rank|是|-|
Mdocs/zh/native_apis/pytorch_2-8-0/torch-distributed.md+16-0
@@ -22,6 +22,7 @@
22|torch.distributed.HashStore|是|-|22|torch.distributed.HashStore|是|-|
23|torch.distributed.FileStore|是|-|23|torch.distributed.FileStore|是|-|
24|torch.distributed.PrefixStore|是|-|24|torch.distributed.PrefixStore|是|-|
25+|torch.distributed.Store.\_\_init\_\_|是|-|
25|torch.distributed.Store.set|是|-|26|torch.distributed.Store.set|是|-|
26|torch.distributed.Store.get|是|-|27|torch.distributed.Store.get|是|-|
27|torch.distributed.Store.add|是|-|28|torch.distributed.Store.add|是|-|
@@ -30,6 +31,21 @@
30|torch.distributed.Store.num_keys|是|-|31|torch.distributed.Store.num_keys|是|-|
31|torch.distributed.Store.delete_key|是|-|32|torch.distributed.Store.delete_key|是|-|
32|torch.distributed.Store.set_timeout|是|-|33|torch.distributed.Store.set_timeout|是|-|
34+|torch.distributed.Store.append|是|-|
35+|torch.distributed.Store.check|是|-|
36+|torch.distributed.Store.has_extended_api|是|-|
37+|torch.distributed.Store.multi_set|是|-|
38+|torch.distributed.Store.multi_get|是|-|
39+|torch.distributed.Store.timeout|是|-|
40+|torch.distributed.TCPStore.\_\_init\_\_|是|-|
41+|torch.distributed.TCPStore.host|是|-|
42+|torch.distributed.TCPStore.libuvBackend|是|-|
43+|torch.distributed.TCPStore.port|是|-|
44+|torch.distributed.HashStore.\_\_init\_\_|是|-|
45+|torch.distributed.FileStore.\_\_init\_\_|是|-|
46+|torch.distributed.FileStore.path|是|-|
47+|torch.distributed.PrefixStore.\_\_init\_\_|是|-|
48+|torch.distributed.PrefixStore.underlying_store|是|-|
33|torch.distributed.new_group|是|当pg_options函数传入类型为torch_npu._C._distributed_c10d.ProcessGroupHCCL.Options()时，配置该变量属性hccl_config可控制HCCL通信域缓存区大小。具体示例可参考《PyTorch 训练模型迁移调优指南》的“hccl_buffer_size”章节，配置变量属性hccl_config的group_name字段可以设置HCCL通信域的通信组自定义名称，取值为长度不超过32的字符串。|49|torch.distributed.new_group|是|当pg_options函数传入类型为torch_npu._C._distributed_c10d.ProcessGroupHCCL.Options()时，配置该变量属性hccl_config可控制HCCL通信域缓存区大小。具体示例可参考《PyTorch 训练模型迁移调优指南》的“hccl_buffer_size”章节，配置变量属性hccl_config的group_name字段可以设置HCCL通信域的通信组自定义名称，取值为长度不超过32的字符串。|
34|torch.distributed.get_group_rank|是|-|50|torch.distributed.get_group_rank|是|-|
35|torch.distributed.get_global_rank|是|-|51|torch.distributed.get_global_rank|是|-|
Mdocs/zh/native_apis/pytorch_2-9-0/torch-distributed.md+16-0
@@ -22,6 +22,7 @@
22|torch.distributed.HashStore|是|-|22|torch.distributed.HashStore|是|-|
23|torch.distributed.FileStore|是|-|23|torch.distributed.FileStore|是|-|
24|torch.distributed.PrefixStore|是|-|24|torch.distributed.PrefixStore|是|-|
25+|torch.distributed.Store.\_\_init\_\_|是|-|
25|torch.distributed.Store.set|是|-|26|torch.distributed.Store.set|是|-|
26|torch.distributed.Store.get|是|-|27|torch.distributed.Store.get|是|-|
27|torch.distributed.Store.add|是|-|28|torch.distributed.Store.add|是|-|
@@ -30,6 +31,21 @@
30|torch.distributed.Store.num_keys|是|-|31|torch.distributed.Store.num_keys|是|-|
31|torch.distributed.Store.delete_key|是|-|32|torch.distributed.Store.delete_key|是|-|
32|torch.distributed.Store.set_timeout|是|-|33|torch.distributed.Store.set_timeout|是|-|
34+|torch.distributed.Store.append|是|-|
35+|torch.distributed.Store.check|是|-|
36+|torch.distributed.Store.has_extended_api|是|-|
37+|torch.distributed.Store.multi_set|是|-|
38+|torch.distributed.Store.multi_get|是|-|
39+|torch.distributed.Store.timeout|是|-|
40+|torch.distributed.TCPStore.\_\_init\_\_|是|-|
41+|torch.distributed.TCPStore.host|是|-|
42+|torch.distributed.TCPStore.libuvBackend|是|-|
43+|torch.distributed.TCPStore.port|是|-|
44+|torch.distributed.HashStore.\_\_init\_\_|是|-|
45+|torch.distributed.FileStore.\_\_init\_\_|是|-|
46+|torch.distributed.FileStore.path|是|-|
47+|torch.distributed.PrefixStore.\_\_init\_\_|是|-|
48+|torch.distributed.PrefixStore.underlying_store|是|-|
33|torch.distributed.new_group|是|当pg_options函数传入类型为torch_npu._C._distributed_c10d.ProcessGroupHCCL.Options()时，配置该变量属性hccl_config可控制HCCL通信域缓存区大小。具体示例可参考《PyTorch 训练模型迁移调优指南》的“hccl_buffer_size”章节，配置变量属性hccl_config的group_name字段可以设置HCCL通信域的通信组自定义名称，取值为长度不超过32的字符串。|49|torch.distributed.new_group|是|当pg_options函数传入类型为torch_npu._C._distributed_c10d.ProcessGroupHCCL.Options()时，配置该变量属性hccl_config可控制HCCL通信域缓存区大小。具体示例可参考《PyTorch 训练模型迁移调优指南》的“hccl_buffer_size”章节，配置变量属性hccl_config的group_name字段可以设置HCCL通信域的通信组自定义名称，取值为长度不超过32的字符串。|
34|torch.distributed.get_group_rank|是|-|50|torch.distributed.get_group_rank|是|-|
35|torch.distributed.get_global_rank|是|-|51|torch.distributed.get_global_rank|是|-|
Atest/distributed/test_store_api.py+137-0
@@ -0,0 +1,137 @@
1+"""
2+Add validation cases for torch.distributed APIs on NPU:
3+1. test/distributed/test_store.py from PyTorch community lacks sufficient API validations, so this file is added.
4+2. This file validates
5+torch.distributed.FileStore.path
6+torch.distributed.Store.__init__
7+torch.distributed.Store.add
8+torch.distributed.Store.timeout
9+torch.distributed.TCPStore.host
10+torch.distributed.TCPStore.port
11+(extendable).
12+"""
13+ 
14+import os
15+import time
16+import socket
17+import tempfile
18+from datetime import timedelta
19+ 
20+import torch
21+import torch.distributed as dist
22+from torch.testing._internal.common_utils import TestCase, run_tests
23+ 
24+import torch_npu
25+ 
26+ 
27+def find_free_port() -> int:
28+ with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
29+ s.bind(('', 0))
30+ return s.getsockname()[1]
31+ 
32+ 
33+class TestStoreAPIs(TestCase):
34+ """Test cases for specific behavioral validations of Store APIs."""
35+ 
36+ def test_filestore_path_behavior(self):
37+ """Test if FileStore actually uses the specified path for data exchange."""
38+ with tempfile.TemporaryDirectory() as temp_dir:
39+ filename = os.path.join(temp_dir, "npu_filestore.txt")
40+
41+ store_master = dist.FileStore(filename, 2)
42+ self.assertEqual(store_master.path, filename)
43+ store_master.set("shared_key", "npu_data")
44+ 
45+ store_worker = dist.FileStore(filename, 2)
46+ val = store_worker.get("shared_key")
47+
48+ self.assertEqual(val, b"npu_data")
49+ 
50+ def test_store_init(self):
51+ """Test the initialization (super().__init__) of the base Store class."""
52+ class MyCustomStore(dist.Store):
53+ def __init__(self):
54+ super().__init__()
55+ self.is_initialized = True
56+ 
57+ custom_store = MyCustomStore()
58+ self.assertTrue(custom_store.is_initialized)
59+ self.assertIsInstance(custom_store, dist.Store)
60+ 
61+ def test_store_add_behavior(self):
62+ """Test the add operation mathematically on a HashStore."""
63+ store = dist.HashStore()
64+ key = "test_add_key"
65+
66+ res1 = store.add(key, 5)
67+ self.assertEqual(res1, 5)
68+
69+ res2 = store.add(key, 10)
70+ self.assertEqual(res2, 15)
71+
72+ self.assertEqual(store.get(key), b"15")
73+
74+ res3 = store.add(key, -3)
75+ self.assertEqual(res3, 12)
76+ 
77+ def test_store_timeout_behavior(self):
78+ """Test if the timeout property actively interrupts blocking operations."""
79+ store = dist.HashStore()
80+
81+ timeout_seconds = 1
82+ test_timeout = timedelta(seconds=timeout_seconds)
83+ store.set_timeout(test_timeout)
84+ self.assertEqual(store.timeout, test_timeout)
85+ 
86+ start_time = time.time()
87+ with self.assertRaises(RuntimeError) as context:
88+ store.wait(["non_existent_key"], test_timeout)
89+ elapsed_time = time.time() - start_time
90+ 
91+ self.assertTrue(
92+ "Timeout" in str(context.exception) or "Wait timeout" in str(context.exception),
93+ f"Exception message does not indicate timeout: {context.exception}"
94+ )
95+
96+ self.assertTrue(
97+ 0.8 <= elapsed_time <= 2.5,
98+ f"Actual wait time {elapsed_time:.2f}s did not respect the {timeout_seconds}s timeout."
99+ )
100+ 
101+ def test_tcpstore_host_and_port_behavior(self):
102+ """Test TCPStore host and port by establishing an actual Client-Server connection."""
103+ host = "127.0.0.1"
104+ port = find_free_port()
105+
106+ server_store = dist.TCPStore(
107+ host_name=host,
108+ port=port,
109+ world_size=2,
110+ is_master=True,
111+ timeout=timedelta(seconds=5),
112+ wait_for_workers=False
113+ )
114+
115+ self.assertEqual(server_store.host, host)
116+ self.assertEqual(server_store.port, port)
117+
118+ server_store.set("tcp_key", "tcp_value")
119+ 
120+ client_store = dist.TCPStore(
121+ host_name=host,
122+ port=port,
123+ world_size=2,
124+ is_master=False,
125+ timeout=timedelta(seconds=5)
126+ )
127+
128+ client_store.wait(["tcp_key"], timedelta(seconds=5))
129+ val = client_store.get("tcp_key")
130+ self.assertEqual(val, b"tcp_value")
131+ 
132+ del client_store
133+ del server_store
134+ 
135+ 
136+if __name__ == "__main__":
137+ run_tests()