Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
benchmark_mhc_post.py 3.92 KB
Newer Older
xiaowei.zhang's avatar
1. Update MOE; 2. Update sglang mHC; 3. Update test scripts; 4 Add new  
xiaowei.zhang 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 
from typing import Callable
import torch
import pytest

from aiter.ops.tilelang import mhc_post_fwd as mhc_post


def generate_mhc_post_test_data(
    num_tokens: int,
    h: int,
    mhc_mult: int,
    device: str = 'cuda',
) -> dict[str, torch.Tensor]:
    x = torch.randn((num_tokens, h), dtype=torch.bfloat16, device=device)
    residual = torch.randn((num_tokens, mhc_mult, h), dtype=torch.bfloat16, device=device)
    post_layer_mix = torch.randn((num_tokens, mhc_mult), dtype=torch.float32, device=device)
    comb_res_mix = torch.randn((num_tokens, mhc_mult, mhc_mult), dtype=torch.float32, device=device)

    o_grad = torch.randn((num_tokens, mhc_mult, h), dtype=torch.bfloat16, device=device)

    return {
        'x': x,
        'residual': residual,
        'post_layer_mix': post_layer_mix,
        'comb_res_mix': comb_res_mix,
        'o_grad': o_grad,
    }


def _tester(
    impl: Callable[[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor], torch.Tensor],
    test_data: dict[str, torch.Tensor],
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    x_ = test_data['x'].clone().requires_grad_()
    residual_ = test_data['residual'].clone().requires_grad_()
    post_layer_mix_ = test_data['post_layer_mix'].clone().requires_grad_()
    comb_res_mix_ = test_data['comb_res_mix'].clone().requires_grad_()
    out_ = impl(x_, residual_, post_layer_mix_, comb_res_mix_)
    torch.autograd.backward([out_], [test_data['o_grad']])
    return out_, x_.grad, residual_.grad, post_layer_mix_.grad, comb_res_mix_.grad


def _estimate_fwd_io_bytes(num_tokens: int, h: int, mhc_mult: int) -> int:
    n = num_tokens
    read_bytes = (
        n * h * 2
        + n * mhc_mult * h * 2
        + n * mhc_mult * 4
        + n * mhc_mult * mhc_mult * 4
    )
    write_bytes = n * mhc_mult * h * 2
    return read_bytes + write_bytes


def _estimate_bwd_io_bytes(num_tokens: int, h: int, mhc_mult: int) -> int:
    n = num_tokens
    read_bytes = (
        n * mhc_mult * h * 2
        + n * h * 2
        + n * mhc_mult * h * 2
        + n * mhc_mult * 4
        + n * mhc_mult * mhc_mult * 4
    )
    write_bytes = n * h * 2 + n * mhc_mult * h * 2 + n * mhc_mult * 4 + n * mhc_mult * mhc_mult * 4
    return read_bytes + write_bytes


@pytest.mark.skipif(not torch.cuda.is_available(), reason='CUDA is required')
@pytest.mark.benchmark
@pytest.mark.parametrize(
    'num_tokens,h,mhc_mult',
    [
        # (1, 4096, 1280, 4),
        # (1, 4096, 2560, 4),
        # (1, 4096, 7168, 4),
        # (2, 4096, 2560, 4),
        
        # h = 7168
        (1, 7168, 4),
        (32, 7168, 4),
        (64, 7168, 4),
        (128, 7168, 4),
        (256, 7168, 4),
        (512, 7168, 4),
        (1024, 7168, 4),
        (2048, 7168, 4),
        (8192, 7168, 4),
        (65536, 7168, 4),
        
        # h = 4096
        (1, 4096, 4),
        (32, 4096, 4),
        (64, 4096, 4),
        (128, 4096, 4),
        (256, 4096, 4),
        (512, 4096, 4),
        (1024, 4096, 4),
        (2048, 4096, 4),
        (8192, 4096, 4),
        (65536, 4096, 4),
        
    ],
)
def test_mhc_post_fwd_benchmark(
    num_tokens: int,
    h: int,
    mhc_mult: int,
    benchmark_timer,
    benchmark_record,
) -> None:
    test_data = generate_mhc_post_test_data(num_tokens=num_tokens, h=h, mhc_mult=mhc_mult)
    x = test_data['x']
    residual = test_data['residual']
    post_layer_mix = test_data['post_layer_mix']
    comb_res_mix = test_data['comb_res_mix']

    def fn_fwd() -> torch.Tensor:
        return mhc_post(x, residual, post_layer_mix, comb_res_mix)

    fn_fwd()
    t_tl_us = benchmark_timer(fn_fwd)
    io_bytes = _estimate_fwd_io_bytes(num_tokens, h, mhc_mult)
    bw_tl_gbs = io_bytes / t_tl_us / 1e3

    benchmark_record(
        kernel='mhc_post',
        operation='fwd',
        params={'num_tokens': num_tokens, 'h': h, 'mhc_mult': mhc_mult},
        time_us=t_tl_us,
        bandwidth_gbs=bw_tl_gbs,
        extras={'num_tokens': num_tokens, 'io_bytes': io_bytes},
    )