File size: 6,232 Bytes
072b453
 
 
 
 
 
 
 
 
 
 
 
 
 
37c352a
 
 
 
 
 
072b453
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
37c352a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
072b453
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Benchmark transformer layout primitives."""

from __future__ import annotations

import argparse
import importlib
import sys
from pathlib import Path

import torch

ROOT = Path(__file__).resolve().parents[2]
sys.path.insert(0, str(ROOT / "transformer-layout-primitives" / "tests"))
from test_transformer_layout_primitives import (  # noqa: E402
    load_source_ops,
    qk_pair_rmsnorm_rope_ref,
    qk_rmsnorm_rope_ref,
    rotate_half_ref,
)


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("transformer_layout_primitives")
    finally:
        if artifact:
            sys.path.remove(artifact)


def time_us(fn, warmup: int, iters: int) -> 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(iters):
        fn()
    end.record()
    torch.cuda.synchronize()
    return start.elapsed_time(end) * 1000.0 / iters


def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument("--backend", choices=["source", "installed"], default="source")
    parser.add_argument("--artifact", default=None)
    parser.add_argument("--mode", choices=["headline", "full"], default="headline")
    parser.add_argument("--warmup", type=int, default=20)
    parser.add_argument("--iters", type=int, default=100)
    args = parser.parse_args()
    ops = load_ops(args.backend, args.artifact)

    print("workload,shape,op,flashrt_us,torch_eager_us,speedup")

    repeat_shapes = [("gqa_prefill", 2520, 8, 128, 4), ("decode_gqa", 1, 8, 128, 4)]
    if args.mode == "full":
        repeat_shapes += [("short_prefill", 128, 8, 128, 4), ("vl_prefill", 4096, 8, 128, 4)]
    for name, seq, heads, dim, repeat in repeat_shapes:
        src = torch.randn((seq, heads, dim), device="cuda", dtype=torch.bfloat16)
        out = torch.empty((seq, heads * repeat, dim), device="cuda", dtype=torch.bfloat16)

        def flash():
            ops.repeat_interleave_heads_bf16(src, repeat, out=out)

        def eager():
            src.repeat_interleave(repeat, dim=1)

        fu = time_us(flash, args.warmup, args.iters)
        eu = time_us(eager, max(5, args.warmup // 2), max(20, args.iters // 2))
        print(f"{name},{seq}x{heads}x{dim}x{repeat},repeat_interleave_heads_bf16,{fu:.3f},{eu:.3f},{eu/fu:.2f}x")

    rope_shapes = [("qwen_prefill", 4096, 32, 128), ("video_prefill", 2520, 24, 128)]
    if args.mode == "full":
        rope_shapes += [("decode", 1, 32, 128), ("short_prefill", 128, 32, 128)]
    for name, seq, heads, dim in rope_shapes:
        x = torch.randn((seq, heads, dim), device="cuda", dtype=torch.bfloat16)
        weight = torch.randn((dim,), device="cuda", dtype=torch.bfloat16)
        cos = torch.randn((seq, dim), device="cuda", dtype=torch.bfloat16)
        sin = torch.randn((seq, dim), device="cuda", dtype=torch.bfloat16)
        out = x.clone()

        def flash():
            out.copy_(x)
            ops.qk_rmsnorm_rope_bf16_(out, weight, cos, sin)

        def eager():
            qk_rmsnorm_rope_ref(x, weight, cos, sin)

        fu = time_us(flash, args.warmup, args.iters)
        eu = time_us(eager, max(5, args.warmup // 2), max(20, args.iters // 2))
        print(f"{name},{seq}x{heads}x{dim},qk_rmsnorm_rope_bf16_,{fu:.3f},{eu:.3f},{eu/fu:.2f}x")

        def flash_rope():
            out.copy_(x)
            ops.rope_rotate_half_bf16_(out, cos, sin)

        def eager_rope():
            rotate_half_ref(x, cos, sin)

        fu = time_us(flash_rope, args.warmup, args.iters)
        eu = time_us(eager_rope, max(5, args.warmup // 2), max(20, args.iters // 2))
        print(f"{name},{seq}x{heads}x{dim},rope_rotate_half_bf16_,{fu:.3f},{eu:.3f},{eu/fu:.2f}x")

    pair_shapes = [
        ("groot_n17_llm", 277, 16, 8, 128),
        ("qwen3vl_vision", 1024, 16, 16, 72),
        ("lingbot_vision", 1024, 16, 16, 80),
        ("video_transformer", 2520, 24, 24, 128),
    ]
    if args.mode == "full":
        pair_shapes += [
            ("action_boundary", 51, 16, 16, 64),
            ("wan_partial_tile", 5070, 24, 24, 128),
        ]
    for name, rows, q_heads, k_heads, dim in pair_shapes:
        q = torch.randn((rows, q_heads, dim), device="cuda", dtype=torch.bfloat16)
        k = torch.randn((rows, k_heads, dim), device="cuda", dtype=torch.bfloat16)
        q_weight = torch.randn((dim,), device="cuda", dtype=torch.bfloat16)
        k_weight = torch.randn((dim,), device="cuda", dtype=torch.bfloat16)
        cos = torch.randn((rows, dim), device="cuda", dtype=torch.bfloat16)
        sin = torch.randn((rows, dim), device="cuda", dtype=torch.bfloat16)
        q_out = torch.empty_like(q)
        k_out = torch.empty_like(k)

        def flash_pair():
            ops.qk_pair_rmsnorm_rope_bf16(
                q, k, q_weight, k_weight, cos, sin, q_out=q_out, k_out=k_out
            )

        def eager_pair():
            qk_pair_rmsnorm_rope_ref(
                q, k, q_weight, k_weight, cos, sin
            )

        fu = time_us(flash_pair, args.warmup, args.iters)
        eu = time_us(eager_pair, max(5, args.warmup // 2), max(20, args.iters // 2))
        print(
            f"{name},{rows}x{q_heads}+{k_heads}x{dim},"
            f"qk_pair_rmsnorm_rope_bf16,{fu:.3f},{eu:.3f},{eu/fu:.2f}x"
        )

    batch, seq, dim = 8, 2048, 2048
    x = torch.randn((batch * seq, dim), device="cuda", dtype=torch.bfloat16)
    gathered = torch.empty((2 * batch, dim), device="cuda", dtype=torch.bfloat16)

    def flash_gather():
        ops.text_gather_bf16(x, batch, seq, out=gathered)

    def eager_gather():
        torch.stack([x[b * seq + offset] for b in range(batch) for offset in (0, seq - 1)], dim=0)

    fu = time_us(flash_gather, args.warmup, args.iters)
    eu = time_us(eager_gather, max(5, args.warmup // 2), max(20, args.iters // 2))
    print(f"text_tokens,{batch}x{seq}x{dim},text_gather_bf16,{fu:.3f},{eu:.3f},{eu/fu:.2f}x")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())