File size: 7,534 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
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
import itertools
import os
import time
from functools import partial
from pathlib import Path

import torch
import triton
from sgl_kernel.test_utils import create_per_token_group_quant_test_data

from sglang.srt.layers.quantization.fp8_kernel import (
    create_per_token_group_quant_fp8_output_scale,
)
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 is_hip
from sglang.srt.utils.bench_utils import bench_kineto

# CI environment detection
IS_CI = (
    os.getenv("CI", "false").lower() == "true"
    or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
)

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


mode_concentrated = IS_CI or (os.environ.get("SGLANG_BENCH_MODE", "") == "concentrated")

if int(os.environ.get("SGLANG_NSYS_PROFILING", "0")):
    configs = [
        [
            768 * 8,
            2048,
            128,
            48,
            fp8_type_,
            dict(
                column_major_scales=True,
                scale_tma_aligned=True,
                scale_ue8m0=True,
                fuse_silu_and_mul=True,
                # masked_layout_mode=None,
                masked_layout_mode="balanced",
                # masked_layout_mode="extreme",
            ),
        ]
    ]
elif mode_concentrated:
    configs = list(
        itertools.product(
            [768],
            [1536, 7168, 16384],
            [128],
            [None],
            [fp8_type_],
            [
                dict(
                    column_major_scales=True,
                    scale_tma_aligned=True,
                    scale_ue8m0=True,
                    fuse_silu_and_mul=False,
                    masked_layout_mode=None,
                ),
            ],
        )
    ) + list(
        itertools.product(
            [768 * 8],
            [2048],
            [128],
            [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",
                ),
            ],
        )
    )
else:
    configs = list(
        itertools.product(
            [1, 4, 16, 64, 256, 768, 2048, 8192, 16384],
            [1536, 7168, 16384],
            [128],
            [None],
            [fp8_type_],
            [
                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 * 8, 4 * 8, 64 * 8, 256 * 8, 768 * 8],
            [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",
                ),
            ],
        )
    )


@triton.testing.perf_report(
    triton.testing.Benchmark(
        x_names=[
            "num_tokens",
            "hidden_dim",
            "group_size",
            "num_ranks",
            "dst_dtype",
            "flags",
        ],
        x_vals=configs,
        line_arg="provider",
        line_vals=["triton", "sglang"],
        # Triton has multi kernels and we only report the time for the core one
        line_names=["Triton (Inaccurate)", "SGL Kernel"],
        styles=[("blue", "-"), ("green", "-")],
        ylabel="us",
        plot_name="per-token-group-quant-8bit-performance",
        args={},
    )
)
def benchmark(
    num_tokens, hidden_dim, group_size, num_ranks, dst_dtype, flags, provider
):
    print(
        f"Testing: {num_tokens=} {hidden_dim=} {group_size=} {num_ranks=} {dst_dtype=} {flags=} {provider=}"
    )

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

    fn, kernel_names = {
        "triton": (
            triton_per_token_group_quant_8bit,
            "_per_token_group_quant_8bit|_silu_and_mul_post_quant_kernel",
        ),
        "sglang": (
            partial(sglang_per_token_group_quant_8bit, enable_v2=True),
            "per_token_group_quant_8bit_kernel",
        ),
    }[provider]
    bench_fn = lambda: fn(
        x=x,
        masked_m=masked_m,
        group_size=group_size,
        dst_dtype=dst_dtype,
        **{k: v for k, v in flags.items() if k not in ["masked_layout_mode"]},
    )

    time_s = bench_kineto(
        bench_fn, kernel_names=kernel_names, num_tests=300 if mode_concentrated else 30
    )
    return time_s * 1e6


if __name__ == "__main__":
    benchmark.run(print_data=True)