-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcheck_py_dispatch_decomp.py
154 lines (131 loc) · 6.87 KB
/
check_py_dispatch_decomp.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
import torch
def fn(x):
out = torch.nn.functional.interpolate(x, (5, 5), mode="bicubic")
return out
def check(req_grad):
print("-- Call fn on cpu tensor")
x = torch.arange(4 * 3 * 10 * 10, dtype=torch.float32).reshape(4, 3, 10, 10)
if req_grad:
x = x.requires_grad_()
expected = fn(x)
print(expected.shape, expected.dtype, expected.device)
print("-- Call fn on meta tensor")
x = torch.arange(4 * 3 * 10 * 10, dtype=torch.float32, device="meta").reshape(4, 3, 10, 10)
if req_grad:
x = x.requires_grad_()
expected = fn(x)
print(expected.shape, expected.dtype, expected.device)
print("-- Call compiled fn on cpu tensor")
x = torch.arange(4 * 3 * 10 * 10, dtype=torch.float32).reshape(4, 3, 10, 10)
if req_grad:
x = x.requires_grad_()
cfn = torch.compile(fn)
output = cfn(x)
print(output.shape, output.dtype, output.device)
print("\n---- Grad mode, input without grad")
check(req_grad=False)
print("\n---- Grad mode, input with grad")
check(req_grad=True)
with torch.no_grad():
print("\n---- NoGrad mode, input without grad")
check(req_grad=False)
print("\n---- NoGrad mode, input with grad")
check(req_grad=True)
with torch.inference_mode():
print("\n---- Inf mode, input without grad")
check(req_grad=False)
print("\n---- Inf mode, input with grad")
check(req_grad=True)
# Output:
"""
---- Grad mode, input without grad
-- Call fn on cpu tensor
Call C++ upsample_nearest1d_kernel_impl
torch.Size([4, 3, 5]) torch.float32 cpu
-- Call fn on meta tensor
Call meta upsample_nearest1d
torch.Size([4, 3, 5]) torch.float32 meta
-- Call compiled fn on cpu tensor
call decomp upsample_nearest1d_vec <class 'torch._subclasses.fake_tensor.FakeTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.fake_tensor.FakeTensor'> <class 'list'>
call decomp upsample_nearest1d_vec <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d_vec <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
torch.Size([4, 3, 5]) torch.float32 cpu
---- Grad mode, input with grad
-- Call fn on cpu tensor
Call C++ upsample_nearest1d_kernel_impl
torch.Size([4, 3, 5]) torch.float32 cpu
-- Call fn on meta tensor
Call meta upsample_nearest1d
torch.Size([4, 3, 5]) torch.float32 meta
-- Call compiled fn on cpu tensor
call decomp upsample_nearest1d_vec <class 'torch._subclasses.fake_tensor.FakeTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.fake_tensor.FakeTensor'> <class 'list'>
call decomp upsample_nearest1d_vec <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d_vec <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
torch.Size([4, 3, 5]) torch.float32 cpu
---- NoGrad mode, input without grad
-- Call fn on cpu tensor
Call C++ upsample_nearest1d_kernel_impl
torch.Size([4, 3, 5]) torch.float32 cpu
-- Call fn on meta tensor
Call meta upsample_nearest1d
torch.Size([4, 3, 5]) torch.float32 meta
-- Call compiled fn on cpu tensor
call decomp upsample_nearest1d_vec <class 'torch._subclasses.fake_tensor.FakeTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.fake_tensor.FakeTensor'> <class 'list'>
call decomp upsample_nearest1d_vec <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d_vec <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
torch.Size([4, 3, 5]) torch.float32 cpu
---- NoGrad mode, input with grad
-- Call fn on cpu tensor
Call C++ upsample_nearest1d_kernel_impl
torch.Size([4, 3, 5]) torch.float32 cpu
-- Call fn on meta tensor
Call meta upsample_nearest1d
torch.Size([4, 3, 5]) torch.float32 meta
-- Call compiled fn on cpu tensor
call decomp upsample_nearest1d_vec <class 'torch._subclasses.fake_tensor.FakeTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.fake_tensor.FakeTensor'> <class 'list'>
call decomp upsample_nearest1d_vec <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d_vec <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
torch.Size([4, 3, 5]) torch.float32 cpu
---- Inf mode, input without grad
-- Call fn on cpu tensor
Call C++ upsample_nearest1d_kernel_impl
torch.Size([4, 3, 5]) torch.float32 cpu
-- Call fn on meta tensor
Call meta upsample_nearest1d
torch.Size([4, 3, 5]) torch.float32 meta
-- Call compiled fn on cpu tensor
call decomp upsample_nearest1d_vec <class 'torch._subclasses.fake_tensor.FakeTensor'> <class 'list'>
Call meta upsample_nearest1d
call decomp upsample_nearest1d_vec <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d_vec <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
torch.Size([4, 3, 5]) torch.float32 cpu
---- Inf mode, input with grad
-- Call fn on cpu tensor
Call C++ upsample_nearest1d_kernel_impl
torch.Size([4, 3, 5]) torch.float32 cpu
-- Call fn on meta tensor
Call meta upsample_nearest1d
torch.Size([4, 3, 5]) torch.float32 meta
-- Call compiled fn on cpu tensor
call decomp upsample_nearest1d_vec <class 'torch._subclasses.fake_tensor.FakeTensor'> <class 'list'>
Call meta upsample_nearest1d
call decomp upsample_nearest1d_vec <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d_vec <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
call decomp upsample_nearest1d: <class 'torch._subclasses.functional_tensor.FunctionalTensor'> <class 'list'>
torch.Size([4, 3, 5]) torch.float32 cpu
"""