benchmark_paged_attention.py 8.03 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
5
import random
import time
6
from typing import Optional
7
8
9

import torch

10
from vllm import _custom_ops as ops
11
from vllm.logger import init_logger
12
from vllm.platforms import current_platform
13
14
15
16
17
from vllm.utils import (
    STR_DTYPE_TO_TORCH_DTYPE,
    FlexibleArgumentParser,
    create_kv_caches_with_random,
)
18

19
20
logger = init_logger(__name__)

21
NUM_BLOCKS = 128 * 1024
22
PARTITION_SIZE = 512
23
PARTITION_SIZE_ROCM = 256
24
25
26
27
28
29


@torch.inference_mode()
def main(
    version: str,
    num_seqs: int,
30
    seq_len: int,
31
32
33
34
35
36
37
38
    num_query_heads: int,
    num_kv_heads: int,
    head_size: int,
    use_alibi: bool,
    block_size: int,
    dtype: torch.dtype,
    seed: int,
    do_profile: bool,
39
    device: str = "cuda",
40
    kv_cache_dtype: Optional[str] = None,
41
) -> None:
42
    current_platform.seed_everything(seed)
43
44

    scale = float(1.0 / (head_size**0.5))
45
46
47
    query = torch.empty(
        num_seqs, num_query_heads, head_size, dtype=dtype, device=device
    )
48
49
50
51
52
    query.uniform_(-scale, scale)

    assert num_query_heads % num_kv_heads == 0
    alibi_slopes = None
    if use_alibi:
53
        alibi_slopes = torch.randn(num_query_heads, dtype=torch.float, device=device)
54

55
56
57
    seq_lens = [seq_len for _ in range(num_seqs)]
    max_seq_len = max(seq_lens)
    seq_lens = torch.tensor(seq_lens, dtype=torch.int, device=device)
58
59

    # Create the block tables.
60
    max_num_blocks_per_seq = (max_seq_len + block_size - 1) // block_size
61
    block_tables_lst: list[list[int]] = []
62
63
    for _ in range(num_seqs):
        block_table = [
64
            random.randint(0, NUM_BLOCKS - 1) for _ in range(max_num_blocks_per_seq)
65
        ]
66
67
        block_tables_lst.append(block_table)

68
    block_tables = torch.tensor(block_tables_lst, dtype=torch.int, device=device)
69
70

    # Create the KV cache.
71
72
73
74
75
76
77
78
79
80
    key_caches, value_caches = create_kv_caches_with_random(
        NUM_BLOCKS,
        block_size,
        1,
        num_kv_heads,
        head_size,
        kv_cache_dtype,
        dtype,
        device=device,
    )
81
    key_cache, value_cache = key_caches[0], value_caches[0]
82
83
84
85

    # Prepare for the paged attention kernel.
    output = torch.empty_like(query)
    if version == "v2":
86
87
        if current_platform.is_rocm():
            global PARTITION_SIZE
88
89
90
91
            if not args.custom_paged_attn and not current_platform.is_navi():
                PARTITION_SIZE = 1024
            else:
                PARTITION_SIZE = PARTITION_SIZE_ROCM
92
        num_partitions = (max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE
93
94
95
96
97
98
99
100
101
102
103
104
        tmp_output = torch.empty(
            size=(num_seqs, num_query_heads, num_partitions, head_size),
            dtype=output.dtype,
            device=output.device,
        )
        exp_sums = torch.empty(
            size=(num_seqs, num_query_heads, num_partitions),
            dtype=torch.float32,
            device=output.device,
        )
        max_logits = torch.empty_like(exp_sums)

105
    def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float:
106
107
108
109
110
        torch.cuda.synchronize()
        if profile:
            torch.cuda.cudart().cudaProfilerStart()
        start_time = time.perf_counter()

111
        # Using default kv_scale
112
        k_scale = v_scale = torch.tensor(1.0, dtype=torch.float32, device=device)
113

114
115
        for _ in range(num_iters):
            if version == "v1":
116
                ops.paged_attention_v1(
117
118
119
120
                    output,
                    query,
                    key_cache,
                    value_cache,
121
                    num_kv_heads,
122
123
                    scale,
                    block_tables,
124
                    seq_lens,
125
                    block_size,
126
                    max_seq_len,
127
                    alibi_slopes,
128
                    kv_cache_dtype,
129
130
                    k_scale,
                    v_scale,
131
132
                )
            elif version == "v2":
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
                if not args.custom_paged_attn:
                    ops.paged_attention_v2(
                        output,
                        exp_sums,
                        max_logits,
                        tmp_output,
                        query,
                        key_cache,
                        value_cache,
                        num_kv_heads,
                        scale,
                        block_tables,
                        seq_lens,
                        block_size,
                        max_seq_len,
                        alibi_slopes,
                        kv_cache_dtype,
                        k_scale,
                        v_scale,
                    )
                else:
                    ops.paged_attention_rocm(
                        output,
                        exp_sums,
                        max_logits,
                        tmp_output,
                        query,
                        key_cache,
                        value_cache,
                        num_kv_heads,
                        scale,
                        block_tables,
                        seq_lens,
166
                        None,
167
168
169
170
171
172
173
                        block_size,
                        max_seq_len,
                        alibi_slopes,
                        kv_cache_dtype,
                        k_scale,
                        v_scale,
                    )
174
175
176
177
178
179
            else:
                raise ValueError(f"Invalid version: {version}")
        torch.cuda.synchronize()

        end_time = time.perf_counter()
        if profile:
180
            torch.cuda.cudart().cudaProfilerStop()
181
182
183
184
        return (end_time - start_time) / num_iters

    # Warmup.
    print("Warming up...")
185
    run_benchmark = run_cuda_benchmark
186
187
188
189
190
191
192
193
194
195
    run_benchmark(num_iters=3, profile=False)

    # Benchmark.
    if do_profile:
        latency = run_benchmark(num_iters=1, profile=True)
    else:
        latency = run_benchmark(num_iters=100, profile=False)
    print(f"Kernel running time: {latency * 1000000:.3f} us")


196
197
198
199
200
if __name__ == "__main__":
    logger.warning(
        "This script benchmarks the paged attention kernel. "
        "By default this is no longer used in vLLM inference."
    )
201

202
203
    parser = FlexibleArgumentParser(description="Benchmark the paged attention kernel.")
    parser.add_argument("--version", type=str, choices=["v1", "v2"], default="v2")
204
    parser.add_argument("--batch-size", type=int, default=8)
Allen.Dou's avatar
Allen.Dou committed
205
    parser.add_argument("--seq-len", type=int, default=4096)
206
207
    parser.add_argument("--num-query-heads", type=int, default=64)
    parser.add_argument("--num-kv-heads", type=int, default=8)
208
209
210
211
212
213
    parser.add_argument(
        "--head-size",
        type=int,
        choices=[64, 80, 96, 112, 120, 128, 192, 256],
        default=128,
    )
214
215
    parser.add_argument("--block-size", type=int, choices=[16, 32], default=16)
    parser.add_argument("--use-alibi", action="store_true")
216
217
218
    parser.add_argument(
        "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half"
    )
219
220
    parser.add_argument("--seed", type=int, default=0)
    parser.add_argument("--profile", action="store_true")
221
222
223
    parser.add_argument(
        "--kv-cache-dtype",
        type=str,
224
        choices=["auto", "fp8", "fp8_e5m2", "fp8_e4m3"],
225
        default="auto",
226
227
        help="Data type for kv cache storage. If 'auto', will use model "
        "data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. "
228
229
230
231
232
        "ROCm (AMD GPU) supports fp8 (=fp8_e4m3)",
    )
    parser.add_argument(
        "--custom-paged-attn", action="store_true", help="Use custom paged attention"
    )
233
234
235
236
237
238
239
240
    args = parser.parse_args()
    print(args)

    if args.num_query_heads % args.num_kv_heads != 0:
        raise ValueError("num_query_heads must be divisible by num_kv_heads")
    main(
        version=args.version,
        num_seqs=args.batch_size,
241
        seq_len=args.seq_len,
242
243
244
245
246
        num_query_heads=args.num_query_heads,
        num_kv_heads=args.num_kv_heads,
        head_size=args.head_size,
        block_size=args.block_size,
        use_alibi=args.use_alibi,
247
        dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
248
249
        seed=args.seed,
        do_profile=args.profile,
250
        kv_cache_dtype=args.kv_cache_dtype,
251
    )