random_sample.py 6 KB
Newer Older
xgqdut2016's avatar
xgqdut2016 committed
1
import torch
PanZezhongQY's avatar
PanZezhongQY committed
2
import ctypes
3
from ctypes import POINTER, Structure, c_int32, c_uint64, c_void_p, c_float
xgqdut2016's avatar
xgqdut2016 committed
4
from libinfiniop import (
5
    InfiniDtype,
PanZezhongQY's avatar
PanZezhongQY committed
6
7
    infiniopHandle_t,
    infiniopTensorDescriptor_t,
xgqdut2016's avatar
xgqdut2016 committed
8
9
10
    open_lib,
    to_tensor,
    get_test_devices,
PanZezhongQY's avatar
PanZezhongQY committed
11
12
    check_error,
    create_workspace,
xgqdut2016's avatar
xgqdut2016 committed
13
14
    test_operator,
    get_args,
xgqdut2016's avatar
xgqdut2016 committed
15
    debug_all,
xgqdut2016's avatar
xgqdut2016 committed
16
17
    get_tolerance,
    profile_operation,
xgqdut2016's avatar
xgqdut2016 committed
18
    synchronize_device,
PanZezhongQY's avatar
PanZezhongQY committed
19
20
)

xgqdut2016's avatar
xgqdut2016 committed
21
22
23
24
25
26
# ==============================================================================
#  Configuration (Internal Use Only)
# ==============================================================================
# These are not meant to be imported from other modules
_TEST_CASES = [
    # voc, random_val, topp, topk, temperature
xgqdut2016's avatar
xgqdut2016 committed
27
28
29
30
31
32
33
34
35
36
    (512, 0.8, 0.8, 3, 0.5),
    (4096, 0.05, 0.9, 5, 1.0),
    (16384, 0.15, 0.85, 10, 2.0),
    (512, 0.08, 0, 3, 0.5),
    (4096, 0.5, 0.9, 1, 1.0),
    (16384, 0.15, 0, 1, 2.0),
    (16384, 0.15, 0, 1, 2.0),
    (32000, 0.08, 0.8, 50, 1.0),
    (32000, 0.08, 1.0, 25, 1.0),
    # (119696, 0.01, 1.0, 100, 1.0),
xgqdut2016's avatar
xgqdut2016 committed
37
38
39
]

# Data types used for testing
xgqdut2016's avatar
xgqdut2016 committed
40
41
42
43
44
_TENSOR_DTYPES = [torch.float16]

_TOLERANCE_MAP = {
    torch.float16: {"atol": 0, "rtol": 0},
}
xgqdut2016's avatar
xgqdut2016 committed
45

46

xgqdut2016's avatar
xgqdut2016 committed
47
DEBUG = False
xgqdut2016's avatar
xgqdut2016 committed
48
49
50
PROFILE = False
NUM_PRERUN = 10
NUM_ITERATIONS = 1000
PanZezhongQY's avatar
PanZezhongQY committed
51
52
53
54
55
56
57
58
59


class RandomSampleDescriptor(Structure):
    _fields_ = [("device", c_int32)]


infiniopRandomSampleDescriptor_t = POINTER(RandomSampleDescriptor)


60
def random_sample(data, random_val, topp, topk, voc, temperature):
61
    if topp > 0 and topk > 1:
62
        sorted_vals, sorted_indices = torch.sort(data, descending=True)
63

64
65
66
        scaled_vals = (sorted_vals - sorted_vals[0]) / temperature
        probs = torch.softmax(scaled_vals, dim=0)
        cum_probs = torch.cumsum(probs, dim=0)
67

68
69
        k_index = min(topk, voc) - 1
        threshold = min(cum_probs[k_index], topp) * random_val
70

71
72
        idx = torch.searchsorted(cum_probs, threshold)
        return sorted_indices[idx]
73

74
    return torch.argmax(data)
zhangyue's avatar
zhangyue committed
75
76


77
78
79
80
81
82
83
84
85
def test(
    lib,
    handle,
    torch_device,
    voc,
    random_val,
    topp,
    topk,
    temperature,
86
    dtype=torch.float16,
87
    sync=None,
88
):
89
90
91
    print(
        f"Testing RandomSample on {torch_device} with voc:{voc} random_val:{random_val} topp:{topp} topk:{topk} temperature:{temperature} dtype:{dtype}"
    )
xgqdut2016's avatar
xgqdut2016 committed
92

PanZezhongQY's avatar
PanZezhongQY committed
93
94
    data = torch.arange(voc).float() * 0.0001
    _perm = torch.randperm(voc)
95
    data = data[_perm].to(dtype).to(torch_device)
96
97

    ans = random_sample(
98
        data, random_val, topp, topk, voc, temperature
99
    )  # 这个函数在device速度可能会很慢,可以通过data.to("cpu")方式加快计算过程
xgqdut2016's avatar
xgqdut2016 committed
100

101
    indices = torch.zeros([], dtype=torch.int64).to(torch_device)
xgqdut2016's avatar
xgqdut2016 committed
102
103
104

    x_tensor, indices_tensor = [to_tensor(tensor, lib) for tensor in [data, indices]]

105
    indices_tensor.descriptor.contents.dt = InfiniDtype.U64  # treat int64 as uint64
PanZezhongQY's avatar
PanZezhongQY committed
106

107
108
109
    if sync is not None:
        sync()

PanZezhongQY's avatar
PanZezhongQY committed
110
111
112
    descriptor = infiniopRandomSampleDescriptor_t()
    check_error(
        lib.infiniopCreateRandomSampleDescriptor(
113
114
115
116
            handle,
            ctypes.byref(descriptor),
            indices_tensor.descriptor,
            x_tensor.descriptor,
PanZezhongQY's avatar
PanZezhongQY committed
117
118
119
120
        )
    )

    # Invalidate the shape and strides in the descriptor to prevent them from being directly used by the kernel
xgqdut2016's avatar
xgqdut2016 committed
121
    for tensor in [x_tensor, indices_tensor]:
122
        tensor.destroyDesc(lib)
PanZezhongQY's avatar
PanZezhongQY committed
123
124
125
126
127
128
129

    workspace_size = c_uint64(0)
    check_error(
        lib.infiniopGetRandomSampleWorkspaceSize(
            descriptor, ctypes.byref(workspace_size)
        )
    )
130
    workspace = create_workspace(workspace_size.value, torch_device)
131

xgqdut2016's avatar
xgqdut2016 committed
132
133
134
135
136
137
138
139
140
141
142
143
144
145
    def lib_random_sample():
        check_error(
            lib.infiniopRandomSample(
                descriptor,
                workspace.data_ptr() if workspace is not None else None,
                workspace_size.value,
                indices_tensor.data,
                x_tensor.data,
                random_val,
                topp,
                topk,
                temperature,
                None,
            )
PanZezhongQY's avatar
PanZezhongQY committed
146
147
        )

xgqdut2016's avatar
xgqdut2016 committed
148
149
    lib_random_sample()

xgqdut2016's avatar
xgqdut2016 committed
150
151
152
153
154
155
    if torch_device == "npu":
        synchronize_device(torch_device)

    atol, rtol = get_tolerance(_TOLERANCE_MAP, dtype)
    if DEBUG:
        debug_all(
156
            (indices.type(ans.dtype), data[indices]),
xgqdut2016's avatar
xgqdut2016 committed
157
158
159
160
161
            (ans, data[ans]),
            "or",
            atol=atol,
            rtol=rtol,
        )
162
    assert indices.type(ans.dtype) == ans or data[ans] == data[indices]
xgqdut2016's avatar
xgqdut2016 committed
163

xgqdut2016's avatar
xgqdut2016 committed
164
165
166
    # Profiling workflow
    if PROFILE:
        # fmt: off
167
        profile_operation("PyTorch", lambda: random_sample(
168
                data, random_val, topp, topk, voc, temperature
xgqdut2016's avatar
xgqdut2016 committed
169
170
171
            ), torch_device, NUM_PRERUN, NUM_ITERATIONS)
        profile_operation("    lib", lambda: lib_random_sample(), torch_device, NUM_PRERUN, NUM_ITERATIONS)
        # fmt: on
PanZezhongQY's avatar
PanZezhongQY committed
172
173
    check_error(lib.infiniopDestroyRandomSampleDescriptor(descriptor))

174

PanZezhongQY's avatar
PanZezhongQY committed
175
176
177
if __name__ == "__main__":
    args = get_args()
    lib = open_lib()
xgqdut2016's avatar
xgqdut2016 committed
178

PanZezhongQY's avatar
PanZezhongQY committed
179
180
181
182
183
184
    lib.infiniopCreateRandomSampleDescriptor.restype = c_int32
    lib.infiniopCreateRandomSampleDescriptor.argtypes = [
        infiniopHandle_t,
        POINTER(infiniopRandomSampleDescriptor_t),
        infiniopTensorDescriptor_t,
    ]
xgqdut2016's avatar
xgqdut2016 committed
185

PanZezhongQY's avatar
PanZezhongQY committed
186
187
188
189
190
    lib.infiniopGetRandomSampleWorkspaceSize.restype = c_int32
    lib.infiniopGetRandomSampleWorkspaceSize.argtypes = [
        infiniopRandomSampleDescriptor_t,
        POINTER(c_uint64),
    ]
xgqdut2016's avatar
xgqdut2016 committed
191

PanZezhongQY's avatar
PanZezhongQY committed
192
193
194
195
196
197
198
199
200
201
202
203
204
    lib.infiniopRandomSample.restype = c_int32
    lib.infiniopRandomSample.argtypes = [
        infiniopRandomSampleDescriptor_t,
        c_void_p,
        c_uint64,
        c_uint64,
        c_void_p,
        c_float,
        c_float,
        c_int32,
        c_float,
        c_void_p,
    ]
xgqdut2016's avatar
xgqdut2016 committed
205

PanZezhongQY's avatar
PanZezhongQY committed
206
207
208
209
210
    lib.infiniopDestroyRandomSampleDescriptor.restype = c_int32
    lib.infiniopDestroyRandomSampleDescriptor.argtypes = [
        infiniopRandomSampleDescriptor_t,
    ]

xgqdut2016's avatar
xgqdut2016 committed
211
    DEBUG = args.debug
xgqdut2016's avatar
xgqdut2016 committed
212
213
214
215
216
217
218
    PROFILE = args.profile
    NUM_PRERUN = args.num_prerun
    NUM_ITERATIONS = args.num_iterations

    # Execute tests
    for device in get_test_devices(args):
        test_operator(lib, device, test, _TEST_CASES, _TENSOR_DTYPES)
219

PanZezhongQY's avatar
PanZezhongQY committed
220
    print("\033[92mTest passed!\033[0m")