Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
benchmark_sparsemax.py 4.94 KB
Newer Older
cmx's avatar
first commit  
cmx committed
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
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172 
import torch
import triton

from utils import QUANTILES
from utils import SingleBenchmarkRunInput
from utils import SingleBenchmarkRunOutput
from utils import _test_memory
from utils import parse_benchmark_script_args
from utils import run_benchmarks

from liger_kernel.transformers.sparsemax import LigerSparsemax
from liger_kernel.utils import infer_device

device = infer_device()


def torch_sparsemax(input_tensor: torch.Tensor, dim: int = -1) -> torch.Tensor:
    input_dims = input_tensor.dim()
    if dim < 0:
        dim = input_dims + dim
    input_sorted, _ = torch.sort(input_tensor, dim=dim, descending=True)
    cumsum_input = torch.cumsum(input_sorted, dim=dim)
    input_size = input_tensor.size(dim)
    range_tensor = torch.arange(1, input_size + 1, device=input_tensor.device, dtype=input_tensor.dtype)
    shape = [1] * input_dims
    shape[dim] = input_size
    range_tensor = range_tensor.view(shape)
    k_bound = 1 + range_tensor * input_sorted
    support = k_bound > cumsum_input
    k = support.sum(dim=dim, keepdim=True).clamp(min=1)
    support_sum = (input_sorted * support).sum(dim=dim, keepdim=True)
    tau = (support_sum - 1) / k
    return torch.clamp(input_tensor - tau, min=0)


class TorchSparsemax(torch.nn.Module):
    def __init__(self, dim: int = -1):
        super().__init__()
        self.dim = dim

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return torch_sparsemax(x, dim=self.dim)


def bench_speed_sparsemax(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunOutput:
    V = input.x
    provider = input.kernel_provider
    mode = input.kernel_operation_mode

    extra_benchmark_config = input.extra_benchmark_config
    B = extra_benchmark_config["B"]
    T = extra_benchmark_config["T"]
    dim = extra_benchmark_config["dim"]
    dtype = extra_benchmark_config["dtype"]

    x_shape = (B * T, V)

    torch_sparsemax_module = TorchSparsemax(dim=dim).to(device)
    liger_sparsemax_module = LigerSparsemax(dim=dim).to(device)

    x = torch.randn(x_shape, dtype=dtype, device=device)
    dy = torch.randn_like(x)
    x.requires_grad_(True)

    # utility functions
    def y_fwd():
        if provider == "liger":
            return liger_sparsemax_module(x)
        elif provider == "torch":
            return torch_sparsemax_module(x)

    if mode == "forward":
        ms_50, ms_20, ms_80 = triton.testing.do_bench(
            y_fwd,
            grad_to_none=[x],
            rep=500,
            quantiles=QUANTILES,
        )
    elif mode == "backward":
        y = y_fwd()
        ms_50, ms_20, ms_80 = triton.testing.do_bench(
            lambda: y.backward(dy, retain_graph=True),
            grad_to_none=[x],
            rep=500,
            quantiles=QUANTILES,
        )
    elif mode == "full":

        def full():
            y = y_fwd()
            y.backward(dy, retain_graph=True)

        ms_50, ms_20, ms_80 = triton.testing.do_bench(
            full,
            grad_to_none=[x],
            rep=500,
            quantiles=QUANTILES,
        )

    return SingleBenchmarkRunOutput(
        y_20=ms_20,
        y_50=ms_50,
        y_80=ms_80,
    )


def bench_memory_sparsemax(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunOutput:
    V = input.x
    provider = input.kernel_provider

    extra_benchmark_config = input.extra_benchmark_config
    B = extra_benchmark_config["B"]
    T = extra_benchmark_config["T"]
    dim = extra_benchmark_config["dim"]
    dtype = extra_benchmark_config["dtype"]

    x_shape = (B * T, V)

    torch_sparsemax_module = TorchSparsemax(dim=dim).to(device)
    liger_sparsemax_module = LigerSparsemax(dim=dim).to(device)

    x = torch.randn(x_shape, dtype=dtype, device=device)
    dy = torch.randn_like(x)
    x.requires_grad_(True)

    # utility functions
    def y_fwd():
        if provider == "liger":
            return liger_sparsemax_module(x)
        elif provider == "torch":
            return torch_sparsemax_module(x)

    def full():
        y = y_fwd()
        y.backward(dy, retain_graph=True)

    mem_50, mem_20, mem_80 = _test_memory(full, quantiles=QUANTILES)

    return SingleBenchmarkRunOutput(
        y_20=mem_20,
        y_50=mem_50,
        y_80=mem_80,
    )


if __name__ == "__main__":
    args = parse_benchmark_script_args()

    common_configs = {
        "kernel_name": "sparsemax",
        "x_name": "V",
        "x_label": "feature size",
        "x_values": [2**i for i in range(10, 16)],
        "kernel_providers": ["liger", "torch"],
        "extra_benchmark_configs": [{"B": 4, "T": 512, "dim": -1, "dtype": torch.float32}],
        "overwrite": args.overwrite,
    }

    run_benchmarks(
        bench_test_fn=bench_speed_sparsemax,
        kernel_operation_modes=["forward", "full", "backward"],
        metric_name="speed",
        metric_unit="ms",
        **common_configs,
    )
    run_benchmarks(
        bench_test_fn=bench_memory_sparsemax,
        kernel_operation_modes=["full"],
        metric_name="memory",
        metric_unit="MB",
        **common_configs,
    )