修复API入参校验和用例适配-pytorch-AtomGit

Mtest/nn/test_module_hooks.py+3-7

	@@ -1457,10 +1457,6 @@ class TestModuleHookNN(NNTestCase):
1457	out, _ = mod(True, inp)	1457	out, _ = mod(True, inp)
1458	out.sum().backward()	1458	out.sum().backward()
1459		1459
1460	- @unittest.skip(
1461	- "Skip: pre-existing error message mismatch "
1462	- "(BackwardHookFunctionBackward vs BackwardHookFunction) on ARM CI"
1463	- )
1464	def test_hook_inplace(self):	1460	def test_hook_inplace(self):
1465	class MyModule(nn.Module):	1461	class MyModule(nn.Module):
1466	def forward(self, inp, do_inplace):	1462	def forward(self, inp, do_inplace):
	@@ -1488,7 +1484,7 @@ class TestModuleHookNN(NNTestCase):
1488	self.assertEqual(hook_called[0], 1)	1484	self.assertEqual(hook_called[0], 1)
1489		1485
1490	# Input inplace error should throw an error	1486	# Input inplace error should throw an error
1491	- with self.assertRaisesRegex(RuntimeError, "Output 0 of BackwardHookFunctionBackward is "	1487	+ with self.assertRaisesRegex(RuntimeError, "Output 0 of BackwardHookFunction(Backward)? is "
1492	"a view and is being modified inplace."):	1488	"a view and is being modified inplace."):
1493	mod(inp.clone(), True)	1489	mod(inp.clone(), True)
1494		1490
	@@ -1497,14 +1493,14 @@ class TestModuleHookNN(NNTestCase):
1497	local_inp = inp.clone()	1493	local_inp = inp.clone()
1498	out = mod(local_inp, False)	1494	out = mod(local_inp, False)
1499	local_inp[0] *= 1	1495	local_inp[0] *= 1
1500	- with self.assertRaisesRegex(RuntimeError, "Output 0 of BackwardHookFunctionBackward is "	1496	+ with self.assertRaisesRegex(RuntimeError, "Output 0 of BackwardHookFunction(Backward)? is "
1501	"a view and its base or another view"):	1497	"a view and its base or another view"):
1502	# Any operation involving the view will fail here	1498	# Any operation involving the view will fail here
1503	mod.inp + 2	1499	mod.inp + 2
1504		1500
1505	# Output inplace error should throw an error	1501	# Output inplace error should throw an error
1506	out = mod(inp, False)	1502	out = mod(inp, False)
1507	- with self.assertRaisesRegex(RuntimeError, "BackwardHookFunctionBackward is a view "	1503	+ with self.assertRaisesRegex(RuntimeError, "BackwardHookFunction(Backward)? is a view "
1508	"and is being modified inplace."):	1504	"and is being modified inplace."):
1509	out += 1	1505	out += 1
1510		1506

Mtest/npu/test_torch_npu.py+25-0

	@@ -140,6 +140,31 @@ class TorchNPUDeviceTestCase(TestCase):
140	torch.npu.reset_stream_limit(stream1)	140	torch.npu.reset_stream_limit(stream1)
141	self.assertEqual(ans_dict_3, torch.npu.get_stream_limit(stream1))	141	self.assertEqual(ans_dict_3, torch.npu.get_stream_limit(stream1))
142		142
		143	+ def test_set_device_limit_device_type_check(self):
		144	+ invalid_devices = ["0", 0.0, None, True, False]
		145	+ origin_called = getattr(torch_npu.npu.set_device_limit, "called", None)
		146	+
		147	+ try:
		148	+ for device in invalid_devices:
		149	+ with self.subTest(device=device):
		150	+ if hasattr(torch_npu.npu.set_device_limit, "called"):
		151	+ torch_npu.npu.set_device_limit.called = False
		152	+
		153	+ with self.assertRaisesRegex(TypeError, "device must be an int"):
		154	+ torch_npu.npu.set_device_limit(device)
		155	+ finally:
		156	+ if origin_called is not None:
		157	+ torch_npu.npu.set_device_limit.called = origin_called
		158	+
		159	+ def test_get_device_limit_device_type_check(self):
		160	+ invalid_devices = ["0", 0.0, None, True, False]
		161	+
		162	+ for device in invalid_devices:
		163	+ with self.subTest(device=device):
		164	+ with self.assertRaisesRegex(TypeError, "device must be an int"):
		165	+ torch_npu.npu.get_device_limit(device)
		166	+
		167	+
143	class TorchNPUMemoryApiTestCase(TestCase):	168	class TorchNPUMemoryApiTestCase(TestCase):
144	def test_npu_memory_stats(self):	169	def test_npu_memory_stats(self):
145	res = torch_npu.npu.memory_stats()	170	res = torch_npu.npu.memory_stats()

Mtorch_npu/npu/npu_config.py+30-8

	@@ -161,6 +161,12 @@ class _allowHF32Matmul:
161	@classmethod	161	@classmethod
162	def __setattr__(cls, name, value):	162	def __setattr__(cls, name, value):
163	if name == "allow_hf32":	163	if name == "allow_hf32":
		164	+ if not isinstance(value, bool):
		165	+ raise TypeError(
		166	+ "allow_hf32 must be a bool, but got {}{}".format(
		167	+ type(value).__name__, pta_error(ErrCode.TYPE)
		168	+ )
		169	+ )
164	option = {"ALLOW_MATMUL_HF32": "enable" if value else "disable"}	170	option = {"ALLOW_MATMUL_HF32": "enable" if value else "disable"}
165	torch_npu._C._npu_setOption(option)	171	torch_npu._C._npu_setOption(option)
166	elif name == "cube_math_type":	172	elif name == "cube_math_type":
	@@ -185,6 +191,12 @@ class _allowHF32Conv:
185	@classmethod	191	@classmethod
186	def __setattr__(cls, name, value):	192	def __setattr__(cls, name, value):
187	if name == "allow_hf32":	193	if name == "allow_hf32":
		194	+ if not isinstance(value, bool):
		195	+ raise TypeError(
		196	+ "allow_hf32 must be a bool, but got {}{}".format(
		197	+ type(value).__name__, pta_error(ErrCode.TYPE)
		198	+ )
		199	+ )
188	option = {"ALLOW_CONV_HF32": "enable" if value else "disable"}	200	option = {"ALLOW_CONV_HF32": "enable" if value else "disable"}
189	torch_npu._C._npu_setOption(option)	201	torch_npu._C._npu_setOption(option)
190		202
	@@ -214,24 +226,34 @@ class _call_once_class:
214	@_call_once_class	226	@_call_once_class
215	def set_device_limit(device, cube_num=-1, vector_num=-1):	227	def set_device_limit(device, cube_num=-1, vector_num=-1):
216	from torch_npu.npu import device_count	228	from torch_npu.npu import device_count
217	- device_id = _get_device_index(device, optional=True)	229	+ if isinstance(device, bool) or not isinstance(device, int):
218	- if device_id < 0 or device_id >= device_count():	230	+ raise TypeError(
		231	+ "device must be an int, but got {}{}".format(
		232	+ type(device).__name__, pta_error(ErrCode.TYPE)
		233	+ )
		234	+ )
		235	+ if device < 0 or device >= device_count():
219	raise AssertionError("Invalid device id" + pta_error(ErrCode.VALUE))	236	raise AssertionError("Invalid device id" + pta_error(ErrCode.VALUE))
220	torch_npu.npu._lazy_init()	237	torch_npu.npu._lazy_init()
221	if cube_num != -1:	238	if cube_num != -1:
222	- torch_npu._C._npu_set_device_res_limit(device_id, 0, cube_num)	239	+ torch_npu._C._npu_set_device_res_limit(device, 0, cube_num)
223	if vector_num != -1:	240	if vector_num != -1:
224	- torch_npu._C._npu_set_device_res_limit(device_id, 1, vector_num)	241	+ torch_npu._C._npu_set_device_res_limit(device, 1, vector_num)
225		242
226		243
227	def get_device_limit(device):	244	def get_device_limit(device):
228	from torch_npu.npu import device_count	245	from torch_npu.npu import device_count
229	- device_id = _get_device_index(device, optional=True)	246	+ if isinstance(device, bool) or not isinstance(device, int):
230	- if device_id < 0 or device_id >= device_count():	247	+ raise TypeError(
		248	+ "device must be an int, but got {}{}".format(
		249	+ type(device).__name__, pta_error(ErrCode.TYPE)
		250	+ )
		251	+ )
		252	+ if device < 0 or device >= device_count():
231	raise AssertionError("Invalid device id" + pta_error(ErrCode.VALUE))	253	raise AssertionError("Invalid device id" + pta_error(ErrCode.VALUE))
232	torch_npu.npu._lazy_init()	254	torch_npu.npu._lazy_init()
233	- return {"cube_core_num": torch_npu._C._npu_get_device_res_limit(device_id, 0), \	255	+ return {"cube_core_num": torch_npu._C._npu_get_device_res_limit(device, 0), \
234	- "vector_core_num": torch_npu._C._npu_get_device_res_limit(device_id, 1)}	256	+ "vector_core_num": torch_npu._C._npu_get_device_res_limit(device, 1)}
235		257
236		258
237	def set_stream_limit(stream, cube_num=-1, vector_num=-1):	259	def set_stream_limit(stream, cube_num=-1, vector_num=-1):