File size: 5,443 Bytes
a402b9b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
173
174
175
176
177
178
179
180
181
182
183
184
185
186
import itertools
import os
import time
from pathlib import Path

import pytest
import torch
from sgl_kernel.test_utils import (
    assert_all_close_or_tiny_diff,
    create_per_token_group_quant_test_data,
)

from sglang.srt.layers.quantization.fp8_kernel import (
    per_token_group_quant_8bit as triton_per_token_group_quant_8bit,
)
from sglang.srt.layers.quantization.fp8_kernel import (
    sglang_per_token_group_quant_8bit,
)
from sglang.srt.utils import get_bool_env_var, is_hip

_is_hip = is_hip()
fp8_type_ = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn

configs = list(
    itertools.product(
        [1, 4, 16, 64, 127, 128, 512, 1024, 4096, 8192],  # num_tokens
        [128, 256, 384, 512, 1024, 1536, 1664, 2048, 4096, 7168, 16384],  # hidden_dim
        [16, 32, 64, 128],  # group_size
        [None],  # num_ranks
        [fp8_type_, torch.int8],  # dtype
        [
            dict(
                column_major_scales=False,
                scale_tma_aligned=False,
                scale_ue8m0=False,
                fuse_silu_and_mul=False,
                masked_layout_mode=None,
            ),
            dict(
                column_major_scales=True,
                scale_tma_aligned=False,
                scale_ue8m0=False,
                fuse_silu_and_mul=False,
                masked_layout_mode=None,
            ),
            dict(
                column_major_scales=True,
                scale_tma_aligned=True,
                scale_ue8m0=False,
                fuse_silu_and_mul=False,
                masked_layout_mode=None,
            ),
            dict(
                column_major_scales=True,
                scale_tma_aligned=True,
                scale_ue8m0=True,
                fuse_silu_and_mul=False,
                masked_layout_mode=None,
            ),
        ],
    )
) + list(
    itertools.product(
        [1, 4, 1 * 8, 4 * 8, 64 * 8, 256 * 8, 768 * 8],
        # TODO support more
        [2048],
        [128],
        [8, 16, 32, 48],
        [fp8_type_],
        [
            dict(
                column_major_scales=True,
                scale_tma_aligned=True,
                scale_ue8m0=True,
                fuse_silu_and_mul=True,
                masked_layout_mode=None,
            ),
            dict(
                column_major_scales=True,
                scale_tma_aligned=True,
                scale_ue8m0=True,
                fuse_silu_and_mul=True,
                masked_layout_mode="balanced",
            ),
            dict(
                column_major_scales=True,
                scale_tma_aligned=True,
                scale_ue8m0=True,
                fuse_silu_and_mul=True,
                masked_layout_mode="imbalanced",
            ),
            dict(
                column_major_scales=True,
                scale_tma_aligned=True,
                scale_ue8m0=True,
                fuse_silu_and_mul=True,
                masked_layout_mode="extreme",
            ),
        ],
    )
)


@pytest.mark.parametrize(
    "num_tokens, hidden_dim, group_size, num_ranks, dst_dtype, flags", configs
)
def test_per_token_group_quant_with_column_major(
    num_tokens,
    hidden_dim,
    group_size,
    num_ranks,
    dst_dtype,
    flags,
):
    print(
        f"{num_tokens=} {hidden_dim=} {group_size=} {num_ranks=} {dst_dtype=} {flags=}"
    )

    arch_major, _ = torch.cuda.get_device_capability(torch.cuda.current_device())
    if flags["scale_ue8m0"] and (arch_major <= 9):
        pytest.skip("Only Blackwell need ue8m0 fusion")
        return

    if (flags["scale_ue8m0"] and (group_size != 128)) or (
        (dst_dtype == torch.int8) and flags["column_major_scales"]
    ):
        pytest.skip()
        return

    x, masked_m = create_per_token_group_quant_test_data(
        num_tokens=num_tokens, hidden_dim=hidden_dim, num_ranks=num_ranks, flags=flags
    )

    # print("hack data!!!")
    # x = torch.full_like(x, fill_value=100)

    execute_kwargs = dict(
        x=x,
        masked_m=masked_m,
        group_size=group_size,
        eps=1e-10,
        dst_dtype=dst_dtype,
        **{k: v for k, v in flags.items() if k not in ["masked_layout_mode"]},
    )

    def _postprocess(x_q, x_s):
        if masked_m is not None:
            print(f"Mask tokens after {masked_m} to be zero")
            for i in range(len(masked_m)):
                x_q[i, masked_m[i] :, :] = 0
                x_s[i, masked_m[i] :, :] = 0
        return x_q, x_s

    x_q_triton, x_s_triton = _postprocess(
        *triton_per_token_group_quant_8bit(**execute_kwargs)
    )
    x_q_sglang, x_s_sglang = _postprocess(
        *sglang_per_token_group_quant_8bit(**execute_kwargs, enable_v2=True)
    )

    try:
        assert_all_close_or_tiny_diff(x_q_triton, x_q_sglang)
        torch.testing.assert_close(
            x_s_triton.contiguous(),
            x_s_sglang.contiguous(),
            rtol=1e-3,
            atol=1e-5,
            msg=lambda message: message + f" {x_s_triton=} {x_s_sglang=}",
        )
    except AssertionError:
        print(
            f"{x.shape=} {x_q_triton.shape=} {x_s_triton.shape=} {x_q_sglang.shape=} {x_s_sglang.shape=}"
        )
        print(f"{x=}")
        print(f"{masked_m=}")
        print(f"{x_q_triton=}")
        print(f"{x_s_triton=}")
        print(f"{x_q_sglang=}")
        print(f"{x_s_sglang=}")

        raise


if __name__ == "__main__":
    pytest.main([__file__])