A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 18:15 UTC
[Misc] Enhance prefix-caching benchmark tool (#6568)
Jiaxin Shan committed
Aug 23, 2024 at 00:32 UTC
d3b5b98021ca2030a0056121122a8965f2328fa2
1 file changed
+140
-4
benchmarks/benchmark_prefix_caching.py
+140
-4
@@ -1,8 +1,45 @@
1
+"""
2
+Benchmark the efficiency of prefix caching.
3
+
4
+This script allows you to benchmark the performance of
5
+a model with and without prefix caching using either fixed prompts
6
+or prompts sampled from the ShareGPT dataset.
7
+
8
+Fixed example usage:
9
+ python benchmark_prefix_caching.py \
10
+ --model meta-llama/Llama-2-7b-chat-hf \
11
+ --enable-prefix-caching \
12
+ --num-prompts 1 \
13
+ --repeat-count 100
14
+
15
+ShareGPT example usage:
16
+ # This command samples 20 prompts with input lengths
17
+ # between 128 and 256 tokens from the ShareGPT dataset,
18
+ # then replicates each prompt 5 times.
19
+ python benchmark_prefix_caching.py \
20
+ --model meta-llama/Llama-2-7b-chat-hf \
21
+ --dataset-path /path/to/ShareGPT_V3_unfiltered_cleaned_split.json \
22
+ --enable-prefix-caching \
23
+ --num-prompts 20 \
24
+ --repeat-count 5 \
25
+ --input-length-range 128:256
26
+"""
27
+
28
+import json
29
+import random
30
import time
31
+from typing import List, Optional, Tuple
32
+
33
+from transformers import PreTrainedTokenizerBase
34
35
from vllm import LLM, SamplingParams
36
from vllm.utils import FlexibleArgumentParser
37
38
+try:
39
+ from vllm.transformers_utils.tokenizer import get_tokenizer
40
+except ImportError:
41
+ from backend_request_func import get_tokenizer
42
+
43
PROMPT = "You are a helpful assistant in recognizes the content of tables in markdown format. Here is a table as fellows. You need to answer my question about the table.\n# Table\n|Opening|Opening|Sl. No.|Film|Cast|Director|Music Director|Notes|\n|----|----|----|----|----|----|----|----|\n|J A N|9|1|Agni Pushpam|Jayabharathi, Kamalahasan|Jeassy|M. K. Arjunan||\n|J A N|16|2|Priyamvada|Mohan Sharma, Lakshmi, KPAC Lalitha|K. S. Sethumadhavan|V. Dakshinamoorthy||\n|J A N|23|3|Yakshagaanam|Madhu, Sheela|Sheela|M. S. Viswanathan||\n|J A N|30|4|Paalkkadal|Sheela, Sharada|T. K. Prasad|A. T. Ummer||\n|F E B|5|5|Amma|Madhu, Srividya|M. Krishnan Nair|M. K. Arjunan||\n|F E B|13|6|Appooppan|Thikkurissi Sukumaran Nair, Kamal Haasan|P. Bhaskaran|M. S. Baburaj||\n|F E B|20|7|Srishti|Chowalloor Krishnankutty, Ravi Alummoodu|K. T. Muhammad|M. S. Baburaj||\n|F E B|20|8|Vanadevatha|Prem Nazir, Madhubala|Yusufali Kechery|G. Devarajan||\n|F E B|27|9|Samasya|Madhu, Kamalahaasan|K. Thankappan|Shyam||\n|F E B|27|10|Yudhabhoomi|K. P. Ummer, Vidhubala|Crossbelt Mani|R. K. Shekhar||\n|M A R|5|11|Seemantha Puthran|Prem Nazir, Jayabharathi|A. B. Raj|M. K. Arjunan||\n|M A R|12|12|Swapnadanam|Rani Chandra, Dr. Mohandas|K. G. George|Bhaskar Chandavarkar||\n|M A R|19|13|Thulavarsham|Prem Nazir, sreedevi, Sudheer|N. Sankaran Nair|V. Dakshinamoorthy||\n|M A R|20|14|Aruthu|Kaviyoor Ponnamma, Kamalahasan|Ravi|G. Devarajan||\n|M A R|26|15|Swimming Pool|Kamal Haasan, M. G. Soman|J. Sasikumar|M. K. Arjunan||\n\n# Question\nWhat' s the content in the (1,1) cells\n" # noqa: E501
44
45
@@ -15,7 +52,83 @@ def test_prefix(llm=None, sampling_params=None, prompts=None):
52
print(f"cost time {end_time - start_time}")
53
54
55
+def sample_requests(
56
+ dataset_path: str,
57
+ num_requests: int,
58
+ tokenizer: PreTrainedTokenizerBase,
59
+ input_length_range: Tuple[int, int],
60
+ fixed_output_len: Optional[int],
61
+) -> List[Tuple[str, int, int]]:
62
+ if fixed_output_len is not None and fixed_output_len < 4:
63
+ raise ValueError("output_len too small")
64
+
65
+ # Load the dataset.
66
+ with open(dataset_path) as f:
67
+ dataset = json.load(f)
68
+ # Filter out the conversations with less than 2 turns.
69
+ dataset = [data for data in dataset if len(data["conversations"]) >= 2]
70
+ # Only keep the first two turns of each conversation.
71
+ dataset = [(data["conversations"][0]["value"],
72
+ data["conversations"][1]["value"]) for data in dataset]
73
+
74
+ # Shuffle the dataset.
75
+ random.shuffle(dataset)
76
+
77
+ min_len, max_len = input_length_range
78
+
79
+ # Filter out sequences that are too long or too short
80
+ filtered_dataset: List[Tuple[str, int, int]] = []
81
+ for i in range(len(dataset)):
82
+ if len(filtered_dataset) == num_requests:
83
+ break
84
+
85
+ # Tokenize the prompts and completions.
86
+ prompt = dataset[i][0]
87
+ prompt_token_ids = tokenizer(prompt).input_ids
88
+ completion = dataset[i][1]
89
+ completion_token_ids = tokenizer(completion).input_ids
90
+ prompt_len = len(prompt_token_ids)
91
+ output_len = len(completion_token_ids
92
+ ) if fixed_output_len is None else fixed_output_len
93
+ if prompt_len < 4 or output_len < 4:
94
+ # Prune too short sequences.
95
+ continue
96
+ if min_len <= prompt_len <= max_len:
97
+ filtered_dataset.append((prompt, prompt_len, output_len))
98
+
99
+ return filtered_dataset
100
+
101
+
102
+def repeat_and_sort_requests(requests: List[Tuple[str, int, int]],
103
+ repeat_count: int,
104
+ sort: bool = False) -> List[str]:
105
+ repeated_requests = requests * repeat_count
106
+ if sort:
107
+ repeated_requests.sort(key=lambda x: x[1])
108
+ else:
109
+ random.shuffle(repeated_requests)
110
+ return [req[0] for req in repeated_requests]
111
+
112
+
113
def main(args):
114
+ tokenizer = get_tokenizer(args.model, trust_remote_code=True)
115
+ input_length_range = tuple(map(int, args.input_length_range.split(':')))
116
+
117
+ if args.dataset_path is not None:
118
+ print(f"Start to sample {args.num_prompts} prompts"
119
+ "from {args.dataset_path}")
120
+ filtered_datasets = sample_requests(
121
+ dataset_path=args.dataset_path,
122
+ num_requests=args.num_prompts,
123
+ tokenizer=tokenizer,
124
+ input_length_range=input_length_range,
125
+ fixed_output_len=args.output_len,
126
+ )
127
+ else:
128
+ prompt_len = len(tokenizer(PROMPT).input_ids)
129
+ filtered_datasets = [(PROMPT, prompt_len, args.output_len)
130
+ ] * args.num_prompts
131
+
132
llm = LLM(model=args.model,
133
tokenizer_mode='auto',
134
trust_remote_code=True,
@@ -24,10 +137,13 @@ def main(args):
137
tensor_parallel_size=args.tensor_parallel_size,
138
enable_prefix_caching=args.enable_prefix_caching)
139
27
- num_prompts = 100
28
- prompts = [PROMPT] * num_prompts
140
sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
141
142
+ print("Testing filtered datasets")
143
+ prompts = repeat_and_sort_requests(filtered_datasets,
144
+ repeat_count=args.repeat_count,
145
+ sort=args.sort)
146
+
147
print("------warm up------")
148
test_prefix(
149
llm=llm,
@@ -45,11 +161,15 @@ def main(args):
161
162
if __name__ == "__main__":
163
parser = FlexibleArgumentParser(
48
- description='Benchmark the performance with or without automatic '
49
- 'prefix caching.')
164
+ description=
165
+ 'Benchmark the performance with or without automatic prefix caching.')
166
parser.add_argument('--model',
167
type=str,
168
default='baichuan-inc/Baichuan2-13B-Chat')
169
+ parser.add_argument("--dataset-path",
170
+ type=str,
171
+ default=None,
172
+ help="Path to the dataset.")
173
parser.add_argument('--tensor-parallel-size', '-tp', type=int, default=1)
174
parser.add_argument('--output-len', type=int, default=10)
175
parser.add_argument('--enable-prefix-caching',
@@ -58,5 +178,21 @@ if __name__ == "__main__":
178
parser.add_argument('--use-v2-block-manager',
179
action='store_true',
180
help='Use BlockSpaceMangerV2')
181
+ parser.add_argument('--num-prompts',
182
+ type=int,
183
+ default=1,
184
+ help="Number of the prompts sampled from dataset")
185
+ parser.add_argument('--repeat-count',
186
+ type=int,
187
+ default=100,
188
+ help='Number of times to repeat each prompt')
189
+ parser.add_argument('--sort',
190
+ action='store_true',
191
+ help='Sort prompts by input length')
192
+ parser.add_argument('--input-length-range',
193
+ type=str,
194
+ default='128:256',
195
+ help='Range of input lengths for sampling prompts,'
196
+ 'specified as "min:max" (e.g., "128:256").')
197
args = parser.parse_args()
198
main(args)