Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
benchmark_cpo_loss.py 5.06 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 
import os
import sys

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.utils import infer_device

device = infer_device()

sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../..")))


#############################################################################
# Test the memory consumption of the linear fused cross entropy loss
#############################################################################


def bench_memory_fused_linear_cpo_loss(
    input: SingleBenchmarkRunInput,
) -> SingleBenchmarkRunOutput:
    from test.chunked_loss.test_cpo_loss import LigerLMHeadCPO
    from test.chunked_loss.test_cpo_loss import TorchLMHeadCPO

    B = input.x
    T = input.extra_benchmark_config["T"]
    H = input.extra_benchmark_config["H"]
    V = input.extra_benchmark_config["V"]
    dtype = input.extra_benchmark_config["dtype"]
    provider = input.kernel_provider

    # Instantiate once and retrieve the first output only
    torch_lm_head_cpo = TorchLMHeadCPO(H=H, V=V, dtype=dtype).to(device)
    liger_lm_head_cpo = LigerLMHeadCPO(H=H, V=V, dtype=dtype).to(device)
    torch_fwd = lambda x, target: torch_lm_head_cpo(x, target)[0]
    liger_fwd = lambda x, target: liger_lm_head_cpo(x, target)[0]

    _input = torch.randn(B, T, H, requires_grad=True, dtype=dtype, device=device)
    target = torch.randint(V, (B, T), dtype=torch.long, device=device)

    def fwd():
        if provider == "liger":
            return liger_fwd(_input, target)
        elif provider == "huggingface":
            return torch_fwd(_input, target)

    def full():
        y = fwd()
        y.backward()

    mem_50, mem_20, mem_80 = _test_memory(full, _iter=10, quantiles=QUANTILES)
    return SingleBenchmarkRunOutput(
        y_20=mem_20,
        y_50=mem_50,
        y_80=mem_80,
    )


# #############################################################################
# # Test the speed of the fused linear cross entropy loss
# #############################################################################


def bench_speed_fused_linear_cpo_loss(
    input: SingleBenchmarkRunInput,
) -> SingleBenchmarkRunOutput:
    from test.chunked_loss.test_cpo_loss import LigerLMHeadCPO
    from test.chunked_loss.test_cpo_loss import TorchLMHeadCPO

    B = input.x
    T = input.extra_benchmark_config["T"]
    H = input.extra_benchmark_config["H"]
    V = input.extra_benchmark_config["V"]
    dtype = input.extra_benchmark_config["dtype"]
    provider = input.kernel_provider
    mode = input.kernel_operation_mode

    # Instantiate once and retrieve the first output only
    torch_lm_head_cpo = TorchLMHeadCPO(H=H, V=V, dtype=dtype).to(device)
    liger_lm_head_cpo = LigerLMHeadCPO(H=H, V=V, dtype=dtype).to(device)
    torch_fwd = lambda x, target: torch_lm_head_cpo(x, target)[0]
    liger_fwd = lambda x, target: liger_lm_head_cpo(x, target)[0]

    _input = torch.randn(B, T, H, requires_grad=True, dtype=dtype, device=device)
    target = torch.randint(V, (B, T), dtype=torch.long, device=device)

    def fwd():
        if provider == "liger":
            return liger_fwd(_input, target)
        elif provider == "huggingface":
            return torch_fwd(_input, target)

    if mode == "forward":
        ms_50, ms_20, ms_80 = triton.testing.do_bench(
            fwd,
            rep=100,
            quantiles=QUANTILES,
        )
    elif mode == "backward":
        y = fwd()

        ms_50, ms_20, ms_80 = triton.testing.do_bench(
            lambda: y.backward(retain_graph=True),
            grad_to_none=[_input],
            rep=100,
            quantiles=QUANTILES,
        )
    elif mode == "full":

        def full():
            y = fwd()
            y.backward()

        ms_50, ms_20, ms_80 = triton.testing.do_bench(
            full,
            rep=100,
            quantiles=QUANTILES,
        )
    return SingleBenchmarkRunOutput(
        y_20=ms_20,
        y_50=ms_50,
        y_80=ms_80,
    )


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

    common_configs = {
        "kernel_name": "fused_linear_cpo_loss",
        "x_name": "B",
        "x_label": "B",
        "x_values": [2**i for i in range(1, 5)],
        "kernel_providers": ["liger", "huggingface"],
        "extra_benchmark_configs": [
            {
                "T": 1024,
                "H": 4096,
                "V": 128256,
                "mode": "forward",
                "dtype": torch.bfloat16,
            }
        ],
        "overwrite": args.overwrite,
    }

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