causal_softmax.py 4.48 KB
Newer Older
xgqdut2016's avatar
xgqdut2016 committed
1
import torch
PanZezhongQY's avatar
PanZezhongQY committed
2
import ctypes
xgqdut2016's avatar
xgqdut2016 committed
3
4
from ctypes import POINTER, Structure, c_int32, c_size_t, c_uint64, c_void_p, c_float
from libinfiniop import (
PanZezhongQY's avatar
PanZezhongQY committed
5
6
    infiniopHandle_t,
    infiniopTensorDescriptor_t,
xgqdut2016's avatar
xgqdut2016 committed
7
8
9
    open_lib,
    to_tensor,
    get_test_devices,
PanZezhongQY's avatar
PanZezhongQY committed
10
    check_error,
xgqdut2016's avatar
xgqdut2016 committed
11
    rearrange_if_needed,
PanZezhongQY's avatar
PanZezhongQY committed
12
    create_workspace,
xgqdut2016's avatar
xgqdut2016 committed
13
14
15
16
17
    test_operator,
    get_args,
    debug,
    get_tolerance,
    profile_operation,
PanZezhongQY's avatar
PanZezhongQY committed
18
19
)

xgqdut2016's avatar
xgqdut2016 committed
20
21
22
23
24
# ==============================================================================
#  Configuration (Internal Use Only)
# ==============================================================================
# These are not meant to be imported from other modules
_TEST_CASES = [
xgqdut2016's avatar
xgqdut2016 committed
25
26
27
28
29
30
31
32
    # x_shape, x_stride
    ((32, 512), None),
    ((32, 512), (1024, 1)),
    ((32, 5, 5), None),
    ((32, 20, 512), None),
    ((32, 20, 512), (20480, 512, 1)),  # Ascend 暂不支持非连续
]

xgqdut2016's avatar
xgqdut2016 committed
33
# Data types used for testing
xgqdut2016's avatar
xgqdut2016 committed
34
_TENSOR_DTYPES = [torch.float16]
xgqdut2016's avatar
xgqdut2016 committed
35
36
37

# Tolerance map for different data types
_TOLERANCE_MAP = {
xgqdut2016's avatar
xgqdut2016 committed
38
    torch.float16: {"atol": 0, "rtol": 1e-2},
xgqdut2016's avatar
xgqdut2016 committed
39
40
41
42
43
44
}

DEBUG = False
PROFILE = False
NUM_PRERUN = 10
NUM_ITERATIONS = 1000
PanZezhongQY's avatar
PanZezhongQY committed
45

xgqdut2016's avatar
xgqdut2016 committed
46

PanZezhongQY's avatar
PanZezhongQY committed
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
class CausalSoftmaxDescriptor(Structure):
    _fields_ = [("device", c_int32)]


infiniopCausalSoftmaxDescriptor_t = POINTER(CausalSoftmaxDescriptor)


def causal_softmax(x):
    type = x.dtype
    mask = torch.tril(torch.ones_like(x), diagonal=-1).flip(dims=[-2, -1])
    y = x.clone()
    masked = torch.where(mask == 1, -torch.inf, y.to(torch.float32))
    return torch.nn.functional.softmax(masked, dim=-1).to(type)


xgqdut2016's avatar
xgqdut2016 committed
62
def test(lib, handle, torch_device, x_shape, x_stride=None, dtype=torch.float16):
PanZezhongQY's avatar
PanZezhongQY committed
63
    print(
xgqdut2016's avatar
xgqdut2016 committed
64
        f"Testing CausalSoftmax on {torch_device} with x_shape:{x_shape} x_stride:{x_stride} dtype:{dtype}"
PanZezhongQY's avatar
PanZezhongQY committed
65
    )
xgqdut2016's avatar
xgqdut2016 committed
66

xgqdut2016's avatar
xgqdut2016 committed
67
68
    x = torch.rand(x_shape, dtype=dtype).to(torch_device)

PanZezhongQY's avatar
PanZezhongQY committed
69
    ans = causal_softmax(x)
xgqdut2016's avatar
xgqdut2016 committed
70
71

    x = rearrange_if_needed(x, x_stride)
xgqdut2016's avatar
xgqdut2016 committed
72

PanZezhongQY's avatar
PanZezhongQY committed
73
    x_tensor = to_tensor(x, lib)
xgqdut2016's avatar
xgqdut2016 committed
74

PanZezhongQY's avatar
PanZezhongQY committed
75
76
77
    descriptor = infiniopCausalSoftmaxDescriptor_t()
    check_error(
        lib.infiniopCreateCausalSoftmaxDescriptor(
xgqdut2016's avatar
xgqdut2016 committed
78
            handle, ctypes.byref(descriptor), x_tensor.descriptor
PanZezhongQY's avatar
PanZezhongQY committed
79
80
        )
    )
xgqdut2016's avatar
xgqdut2016 committed
81
82
83
84

    # Invalidate the shape and strides in the descriptor to prevent them from being directly used by the kernel
    x_tensor.descriptor.contents.invalidate()

PanZezhongQY's avatar
PanZezhongQY committed
85
86
87
88
89
90
91
    workspace_size = c_uint64(0)
    check_error(
        lib.infiniopGetCausalSoftmaxWorkspaceSize(
            descriptor, ctypes.byref(workspace_size)
        )
    )
    workspace = create_workspace(workspace_size.value, x.device)
xgqdut2016's avatar
xgqdut2016 committed
92

xgqdut2016's avatar
xgqdut2016 committed
93
94
95
96
97
98
99
100
101
    def lib_causal_softmax():
        check_error(
            lib.infiniopCausalSoftmax(
                descriptor,
                workspace.data_ptr() if workspace is not None else None,
                workspace_size.value,
                x_tensor.data,
                None,
            )
PanZezhongQY's avatar
PanZezhongQY committed
102
        )
xgqdut2016's avatar
xgqdut2016 committed
103

xgqdut2016's avatar
xgqdut2016 committed
104
    lib_causal_softmax()
xgqdut2016's avatar
xgqdut2016 committed
105

xgqdut2016's avatar
xgqdut2016 committed
106
107
108
109
110
111
112
113
114
115
116
    atol, rtol = get_tolerance(_TOLERANCE_MAP, dtype)
    if DEBUG:
        debug(x, ans, atol=atol, rtol=rtol)
    assert torch.allclose(x, ans, atol=atol, rtol=rtol)

    # Profiling workflow
    if PROFILE:
        # fmt: off
        profile_operation("PyTorch", lambda: causal_softmax(x), torch_device, NUM_PRERUN, NUM_ITERATIONS)
        profile_operation("    lib", lambda: lib_causal_softmax(), torch_device, NUM_PRERUN, NUM_ITERATIONS)
        # fmt: on
PanZezhongQY's avatar
PanZezhongQY committed
117

xgqdut2016's avatar
xgqdut2016 committed
118
    check_error(lib.infiniopDestroyCausalSoftmaxDescriptor(descriptor))
119

PanZezhongQY's avatar
PanZezhongQY committed
120
121
122
123

if __name__ == "__main__":
    args = get_args()
    lib = open_lib()
xgqdut2016's avatar
xgqdut2016 committed
124

PanZezhongQY's avatar
PanZezhongQY committed
125
126
127
128
129
130
    lib.infiniopCreateCausalSoftmaxDescriptor.restype = c_int32
    lib.infiniopCreateCausalSoftmaxDescriptor.argtypes = [
        infiniopHandle_t,
        POINTER(infiniopCausalSoftmaxDescriptor_t),
        infiniopTensorDescriptor_t,
    ]
xgqdut2016's avatar
xgqdut2016 committed
131

PanZezhongQY's avatar
PanZezhongQY committed
132
133
134
135
136
    lib.infiniopGetCausalSoftmaxWorkspaceSize.restype = c_int32
    lib.infiniopGetCausalSoftmaxWorkspaceSize.argtypes = [
        infiniopCausalSoftmaxDescriptor_t,
        POINTER(c_uint64),
    ]
xgqdut2016's avatar
xgqdut2016 committed
137

PanZezhongQY's avatar
PanZezhongQY committed
138
139
140
141
142
143
144
145
    lib.infiniopCausalSoftmax.restype = c_int32
    lib.infiniopCausalSoftmax.argtypes = [
        infiniopCausalSoftmaxDescriptor_t,
        c_void_p,
        c_uint64,
        c_void_p,
        c_void_p,
    ]
xgqdut2016's avatar
xgqdut2016 committed
146

PanZezhongQY's avatar
PanZezhongQY committed
147
148
149
150
    lib.infiniopDestroyCausalSoftmaxDescriptor.restype = c_int32
    lib.infiniopDestroyCausalSoftmaxDescriptor.argtypes = [
        infiniopCausalSoftmaxDescriptor_t,
    ]
xgqdut2016's avatar
xgqdut2016 committed
151

xgqdut2016's avatar
xgqdut2016 committed
152
153
154
155
156
    # Configure testing options
    DEBUG = args.debug
    PROFILE = args.profile
    NUM_PRERUN = args.num_prerun
    NUM_ITERATIONS = args.num_iterations
xgqdut2016's avatar
xgqdut2016 committed
157

xgqdut2016's avatar
xgqdut2016 committed
158
159
    for device in get_test_devices(args):
        test_operator(lib, device, test, _TEST_CASES, _TENSOR_DTYPES)
PanZezhongQY's avatar
PanZezhongQY committed
160
161

    print("\033[92mTest passed!\033[0m")