File size: 6,534 Bytes
48b17de
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Native/wrapper/eager/compile benchmark for padded FP8 producers."""

from __future__ import annotations

import argparse
import importlib
import os
import sys
from pathlib import Path

import torch
import torch.nn.functional as F

PACKAGE = Path(__file__).resolve().parents[1]
ROOT = PACKAGE.parent
sys.path.insert(0, str(PACKAGE / "tests"))
from test_padded_fp8_producers import load_source_ops  # noqa: E402

SHAPES = [
    ("decode", 1, 1, 1280, 16),
    ("groot-dit", 1, 40, 1536, 64),
    ("vla", 1, 51, 2048, 64),
    ("vision", 1, 105, 1280, 128),
    ("prefill", 2, 277, 2048, 320),
]


def load_ops(backend: str, artifact: str | None):
    if backend == "source":
        return load_source_ops()
    if artifact:
        sys.path.insert(0, artifact)
    try:
        return importlib.import_module("padded_fp8_producers")
    finally:
        if artifact:
            sys.path.remove(artifact)


def load_native():
    from torch.utils.cpp_extension import load

    major, minor = torch.cuda.get_device_capability()
    os.environ.setdefault(
        "TORCH_CUDA_ARCH_LIST", "12.0a" if major >= 12 else f"{major}.{minor}"
    )
    return load(
        name="padded_fp8_producers_native_bench",
        sources=[
            str(PACKAGE / "benchmarks" / "native_binding.cpp"),
            str(PACKAGE / "csrc" / "padded_fp8_producers.cu"),
        ],
        extra_include_paths=[str(PACKAGE / "csrc")],
        extra_cflags=["-O3"],
        extra_cuda_cflags=[
            "-O3",
            "--expt-relaxed-constexpr",
            "-U__CUDA_NO_BFLOAT16_CONVERSIONS__",
            "-U__CUDA_NO_BFLOAT16_OPERATORS__",
            "-U__CUDA_NO_BFLOAT162_OPERATORS__",
        ],
        verbose=False,
    )


def bench(fn, warmup=100, iterations=500) -> float:
    for _ in range(warmup):
        fn()
    torch.cuda.synchronize()
    start = torch.cuda.Event(enable_timing=True)
    end = torch.cuda.Event(enable_timing=True)
    start.record()
    for _ in range(iterations):
        fn()
    end.record()
    end.synchronize()
    return start.elapsed_time(end) * 1000.0 / iterations


def eager_adaptive(x, weight, gamma, beta, scale, output):
    batch, rows, _ = x.shape
    norm = x.float() * torch.rsqrt(
        x.float().square().mean(dim=-1, keepdim=True) + 1e-6
    )
    value = (
        (1.0 + gamma[:, None, :].float()) * norm * weight.float()
        + beta[:, None, :].float()
    ).bfloat16()
    output[:, :rows].copy_(
        torch.clamp(value.float() / scale, -448.0, 448.0).to(
            torch.float8_e4m3fn
        )
    )
    output[:, rows:].zero_()
    return output


def eager_swiglu(gate, up, scale, output):
    rows = gate.shape[0]
    value = (F.silu(gate.float()) * up.float()).bfloat16()
    output[:rows].copy_(
        torch.clamp(value.float() / scale, -448.0, 448.0).to(
            torch.float8_e4m3fn
        )
    )
    output[rows:].zero_()
    return output


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--backend", choices=("source", "installed"), default="source")
    parser.add_argument("--artifact")
    args = parser.parse_args()
    ops = load_ops(args.backend, args.artifact)
    native = load_native()
    print("op,shape,native_us,wrapper_us,eager_us,compile_us,wrapper/native")

    for label, batch, rows, dim, padded in SHAPES:
        x = (torch.randn((batch, rows, dim), device="cuda") * 0.4).bfloat16()
        weight = torch.ones((dim,), device="cuda", dtype=torch.bfloat16)
        gamma = torch.zeros((batch, dim), device="cuda", dtype=torch.bfloat16)
        beta = torch.zeros_like(gamma)
        scale = torch.tensor([0.01], device="cuda", dtype=torch.float32)
        native_out = torch.empty(
            (batch, padded, dim), device="cuda", dtype=torch.float8_e4m3fn
        )
        wrapper_out = torch.empty_like(native_out)
        eager_out = torch.empty_like(native_out)
        compile_out = torch.empty_like(native_out)

        native_fn = lambda: native.adaptive(
            x.data_ptr(), weight.data_ptr(), gamma.data_ptr(), beta.data_ptr(),
            scale.data_ptr(), native_out.data_ptr(), batch, rows, padded, dim,
            1e-6
        )
        wrapper_fn = lambda: ops.adaptive_rms_norm_quant_fp8_padded_bf16(
            x, weight, gamma, beta, scale, output=wrapper_out
        )
        eager_fn = lambda: eager_adaptive(
            x, weight, gamma, beta, scale, eager_out
        )
        compiled_call = torch.compile(eager_adaptive, fullgraph=True)
        compile_fn = lambda: compiled_call(
            x, weight, gamma, beta, scale, compile_out
        )
        native_fn()
        wrapper_fn()
        torch.testing.assert_close(
            native_out.float(), wrapper_out.float(), rtol=0, atol=0
        )
        times = [bench(fn) for fn in (native_fn, wrapper_fn, eager_fn, compile_fn)]
        print(
            f"adaptive,{label}:B{batch}S{rows}D{dim}P{padded},"
            f"{times[0]:.3f},{times[1]:.3f},{times[2]:.3f},{times[3]:.3f},"
            f"{times[1] / times[0]:.3f}"
        )

        gate = x.reshape(batch * rows, dim)
        up = torch.randn_like(gate)
        native_swiglu = torch.empty(
            (batch * padded, dim), device="cuda", dtype=torch.float8_e4m3fn
        )
        wrapper_swiglu = torch.empty_like(native_swiglu)
        eager_swiglu_out = torch.empty_like(native_swiglu)
        compile_swiglu_out = torch.empty_like(native_swiglu)
        native_fn = lambda: native.swiglu(
            gate.data_ptr(), up.data_ptr(), scale.data_ptr(),
            native_swiglu.data_ptr(), batch * rows, batch * padded, dim
        )
        wrapper_fn = lambda: ops.swiglu_quant_fp8_padded_bf16(
            gate, up, scale, output=wrapper_swiglu
        )
        eager_fn = lambda: eager_swiglu(
            gate, up, scale, eager_swiglu_out
        )
        compiled_call = torch.compile(eager_swiglu, fullgraph=True)
        compile_fn = lambda: compiled_call(
            gate, up, scale, compile_swiglu_out
        )
        native_fn()
        wrapper_fn()
        torch.testing.assert_close(
            native_swiglu.float(), wrapper_swiglu.float(), rtol=0, atol=0
        )
        times = [bench(fn) for fn in (native_fn, wrapper_fn, eager_fn, compile_fn)]
        print(
            f"swiglu,{label}:S{batch * rows}D{dim}P{batch * padded},"
            f"{times[0]:.3f},{times[1]:.3f},{times[2]:.3f},{times[3]:.3f},"
            f"{times[1] / times[0]:.3f}"
        )


if __name__ == "__main__":
    main()