Update benchmark scripts (#8)
This commit is contained in:
Lianmin Zheng
60
benchmark/json_decode_regex/README.md
Normal file
60
benchmark/json_decode_regex/README.md
Normal file
@@ -0,0 +1,60 @@
|
||||
## Run benchmark
|
||||
|
||||
### Build dataset
|
||||
```
|
||||
pip install wikipedia
|
||||
python3 build_dataset.py
|
||||
```
|
||||
|
||||
### Dependencies
|
||||
|
||||
```
|
||||
llama_cpp_python 0.2.19
|
||||
guidance 0.1.10
|
||||
vllm 0.2.5
|
||||
outlines 0.0.22
|
||||
```
|
||||
|
||||
### Benchmark sglang
|
||||
|
||||
Run Llama-7B
|
||||
|
||||
```
|
||||
python3 -m sglang.launch_server --model-path meta-llama/Llama-2-7b-chat-hf --port 30000
|
||||
```
|
||||
|
||||
Run Mixtral-8x7B
|
||||
|
||||
```
|
||||
python3 -m sglang.launch_server --model-path mistralai/Mixtral-8x7B-Instruct-v0.1 --port 30000 --tp-size 8
|
||||
```
|
||||
|
||||
Benchmark
|
||||
|
||||
```
|
||||
python3 bench_sglang.py --num-questions 10
|
||||
```
|
||||
|
||||
|
||||
### Benchmark vllm
|
||||
|
||||
Run Llama-7B
|
||||
|
||||
```
|
||||
python3 -m outlines.serve.serve --tokenizer-mode auto --model meta-llama/Llama-2-7b-chat-hf --disable-log-requests --port 21000
|
||||
```
|
||||
|
||||
Benchmark
|
||||
|
||||
```
|
||||
python3 bench_other.py --backend vllm --num-questions 10
|
||||
```
|
||||
|
||||
|
||||
### Benchmark guidance
|
||||
|
||||
Run Llama-7B and benchmark
|
||||
|
||||
```
|
||||
python3 bench_other.py --backend guidance --num-questions 10 --parallel 1
|
||||
```
|
||||
125
benchmark/json_decode_regex/bench_other.py
Normal file
125
benchmark/json_decode_regex/bench_other.py
Normal file
@@ -0,0 +1,125 @@
|
||||
import argparse
|
||||
import json
|
||||
import time
|
||||
from concurrent.futures import ThreadPoolExecutor
|
||||
from functools import partial
|
||||
|
||||
from sglang.test.test_utils import (
|
||||
add_common_other_args_and_parse,
|
||||
call_generate_outlines,
|
||||
)
|
||||
from sglang.utils import dump_state_text, read_jsonl
|
||||
from sglang.lang.ir import REGEX_INT, REGEX_STRING, REGEX_FLOAT
|
||||
from tqdm import tqdm
|
||||
|
||||
REGEX_LIST = r"\[(" + REGEX_STRING + ", )*" + REGEX_STRING + r"\]"
|
||||
|
||||
|
||||
# fmt: off
|
||||
def json_decode(document, generate):
|
||||
s = "Please extract the information of a city from the following wikipedia page.\n"
|
||||
s += "Page begin.\n" + document + "Page end.\n"
|
||||
s += "Here is the name, country, and symbol of the city in JSON format.\n"
|
||||
s += "{\n"
|
||||
s += ' "name": '
|
||||
s += generate(s, max_tokens=8, regex=REGEX_STRING + ",") + "\n"
|
||||
s += ' "country": '
|
||||
s += generate(s, max_tokens=8, regex=REGEX_STRING + ",") + "\n"
|
||||
s += ' "latitude": '
|
||||
s += generate(s, max_tokens=8, regex=REGEX_FLOAT + ",") + "\n"
|
||||
s += ' "population": '
|
||||
s += generate(s, max_tokens=8, regex=REGEX_INT + ",") + "\n"
|
||||
s += ' "top 3 landmarks": '
|
||||
s += generate(s, max_tokens=24, regex=REGEX_LIST) + "\n"
|
||||
s += "}\n"
|
||||
|
||||
return s
|
||||
# fmt: on
|
||||
|
||||
|
||||
def main(args):
|
||||
lines = read_jsonl(args.data_path)
|
||||
arguments = []
|
||||
for i in range(len(lines[: args.num_questions])):
|
||||
arguments.append(
|
||||
{
|
||||
"document": lines[i]["document"],
|
||||
}
|
||||
)
|
||||
states = [None] * len(arguments)
|
||||
|
||||
# Select backend
|
||||
if args.backend == "vllm":
|
||||
url = f"{args.host}:{args.port}/generate"
|
||||
generate = partial(call_generate_outlines, url=url, temperature=0)
|
||||
elif args.backend == "guidance":
|
||||
from guidance import gen, models
|
||||
|
||||
model = models.LlamaCpp(
|
||||
"/home/ubuntu/model_weights/Llama-2-7b-chat-hf/ggml-model-f16.gguf",
|
||||
n_gpu_layers=-1,
|
||||
n_ctx=4096,
|
||||
)
|
||||
|
||||
def generate(prompt, max_tokens, stop=None, regex=None):
|
||||
out = (
|
||||
model
|
||||
+ prompt
|
||||
+ gen(
|
||||
name="answer",
|
||||
max_tokens=max_tokens,
|
||||
temperature=0,
|
||||
stop=stop,
|
||||
regex=regex,
|
||||
)
|
||||
)
|
||||
return out["answer"]
|
||||
|
||||
# warmup
|
||||
for _ in range(3):
|
||||
generate("Hello!" * 10, max_tokens=64, stop=None)
|
||||
else:
|
||||
raise ValueError(f"Invalid backend: {args.backend}")
|
||||
|
||||
# Run requests
|
||||
def get_one_answer(i):
|
||||
states[i] = json_decode(generate=generate, **arguments[i])
|
||||
|
||||
tic = time.time()
|
||||
if args.parallel == 1:
|
||||
for i in tqdm(range(len(arguments))):
|
||||
get_one_answer(i)
|
||||
else:
|
||||
with ThreadPoolExecutor(args.parallel) as executor:
|
||||
rets = executor.map(get_one_answer, list(range(len(arguments))))
|
||||
for _ in rets:
|
||||
pass
|
||||
|
||||
latency = time.time() - tic
|
||||
|
||||
# Compute accuracy
|
||||
print(f"Latency: {latency:.3f}")
|
||||
|
||||
# Write results
|
||||
dump_state_text(f"tmp_output_{args.backend}.txt", states)
|
||||
|
||||
with open(args.result_file, "a") as fout:
|
||||
value = {
|
||||
"task": "json_decode_regex",
|
||||
"backend": args.backend,
|
||||
"num_gpus": 1,
|
||||
"latency": round(latency, 3),
|
||||
"num_requests": args.num_questions,
|
||||
"other": {
|
||||
"parallel": args.parallel,
|
||||
},
|
||||
}
|
||||
fout.write(json.dumps(value) + "\n")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("--data-path", type=str, default="questions.jsonl")
|
||||
parser.add_argument("--num-questions", type=int, default=20)
|
||||
args = add_common_other_args_and_parse(parser)
|
||||
main(args)
|
||||
98
benchmark/json_decode_regex/bench_sglang.py
Normal file
98
benchmark/json_decode_regex/bench_sglang.py
Normal file
@@ -0,0 +1,98 @@
|
||||
import argparse
|
||||
import json
|
||||
import time
|
||||
|
||||
import sglang as sgl
|
||||
from sglang.lang.ir import REGEX_INT, REGEX_STRING, REGEX_FLOAT
|
||||
from sglang.test.test_utils import (
|
||||
add_common_sglang_args_and_parse,
|
||||
select_sglang_backend,
|
||||
)
|
||||
from sglang.utils import dump_state_text, read_jsonl
|
||||
|
||||
REGEX_LIST = r"\[(" + REGEX_STRING + ", )*" + REGEX_STRING + r"\]"
|
||||
|
||||
# fmt: off
|
||||
@sgl.function
|
||||
def json_warm_up(s):
|
||||
s += "The information about Hogwarts is in the following JSON format.\n"
|
||||
with s.var_scope("json_output"):
|
||||
s += "{\n"
|
||||
s += ' "name": ' + sgl.gen("name", max_tokens=8, regex=REGEX_STRING + ",") + "\n"
|
||||
s += ' "country": ' + sgl.gen("country", max_tokens=8, regex=REGEX_STRING + ",") + "\n"
|
||||
s += ' "latitude": ' + sgl.gen("latitude", max_tokens=8, regex=REGEX_FLOAT + ",") + "\n"
|
||||
s += ' "population": ' + sgl.gen("population", max_tokens=8, regex=REGEX_INT + ",") + "\n"
|
||||
s += ' "top 3 landmarks": ' + sgl.gen( "landmarks", max_tokens=24, regex=REGEX_LIST) + "\n"
|
||||
s += "}\n"
|
||||
print(f'The warmp up json result is:\n{s["json_output"]}')
|
||||
# fmt: on
|
||||
|
||||
# fmt: off
|
||||
@sgl.function
|
||||
def json_decode(s, document):
|
||||
s += "Please extract the information of a city from the following wikipedia page.\n"
|
||||
s += "Page begin.\n" + document + "Page end.\n"
|
||||
s += "Here is the name, country, and symbol of the city in JSON format.\n"
|
||||
with s.var_scope("json_output"):
|
||||
s += "{\n"
|
||||
s += ' "name": ' + sgl.gen("name", max_tokens=8, regex=REGEX_STRING + ",") + "\n"
|
||||
s += ' "country": ' + sgl.gen("country", max_tokens=8, regex=REGEX_STRING + ",") + "\n"
|
||||
s += ' "latitude": ' + sgl.gen("latitude", max_tokens=8, regex=REGEX_FLOAT + ",") + "\n"
|
||||
s += ' "population": ' + sgl.gen("population", max_tokens=8, regex=REGEX_INT + ",") + "\n"
|
||||
s += ' "top 3 landmarks": ' + sgl.gen( "landmarks", max_tokens=24, regex=REGEX_LIST) + "\n"
|
||||
s += "}\n"
|
||||
# fmt: on
|
||||
|
||||
|
||||
def main(args):
|
||||
lines = read_jsonl(args.data_path)
|
||||
arguments = []
|
||||
for i in range(len(lines[: args.num_questions])):
|
||||
arguments.append(
|
||||
{
|
||||
"document": lines[i]["document"],
|
||||
}
|
||||
)
|
||||
|
||||
# Select backend
|
||||
backend = select_sglang_backend(args)
|
||||
sgl.set_default_backend(backend)
|
||||
|
||||
# Warm up
|
||||
json_warm_up.run().sync()
|
||||
|
||||
# Run requests
|
||||
tic = time.time()
|
||||
states = json_decode.run_batch(arguments, temperature=0, num_threads=args.parallel)
|
||||
latency = time.time() - tic
|
||||
|
||||
# Compute accuracy
|
||||
print(f"Latency: {latency:.3f}")
|
||||
|
||||
# Write results
|
||||
dump_state_text(f"tmp_output_{args.backend}.txt", states)
|
||||
|
||||
with open(f"tmp_{args.backend}_json_results.txt", "w") as fout:
|
||||
for state in states:
|
||||
fout.write(state["json_output"] + "\n")
|
||||
|
||||
with open(args.result_file, "a") as fout:
|
||||
value = {
|
||||
"task": "json_decode_regex",
|
||||
"backend": args.backend,
|
||||
"num_gpus": 1,
|
||||
"latency": round(latency, 3),
|
||||
"num_requests": args.num_questions,
|
||||
"other": {
|
||||
"parallel": args.parallel,
|
||||
},
|
||||
}
|
||||
fout.write(json.dumps(value) + "\n")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("--data-path", type=str, default="questions.jsonl")
|
||||
parser.add_argument("--num-questions", type=int, default=20)
|
||||
args = add_common_sglang_args_and_parse(parser)
|
||||
main(args)
|
||||
58
benchmark/json_decode_regex/build_dataset.py
Normal file
58
benchmark/json_decode_regex/build_dataset.py
Normal file
@@ -0,0 +1,58 @@
|
||||
import json
|
||||
|
||||
import transformers
|
||||
import wikipedia
|
||||
|
||||
model_path = "meta-llama/Llama-2-7b-chat-hf"
|
||||
t = transformers.AutoTokenizer.from_pretrained(model_path)
|
||||
city_names = [
|
||||
"los angles",
|
||||
"london",
|
||||
"tokyo",
|
||||
"beijing",
|
||||
"singapore",
|
||||
"paris",
|
||||
"dubai",
|
||||
"sydney",
|
||||
"moscow",
|
||||
"rome",
|
||||
"toronto",
|
||||
"rio de janeiro",
|
||||
"istanbul",
|
||||
"berlin",
|
||||
"auckland",
|
||||
"buenos aires",
|
||||
"mexico city",
|
||||
"mumbai",
|
||||
"seoul",
|
||||
"bangkok",
|
||||
"cairo",
|
||||
"athens",
|
||||
"jerusalem",
|
||||
]
|
||||
|
||||
|
||||
def get_content(city_name):
|
||||
content = str(wikipedia.page(city_name).content)
|
||||
content = content.replace("\n\n", "\n")
|
||||
|
||||
tokens = t.encode(content)
|
||||
|
||||
expected_tokens = 3000
|
||||
truncate_len = int((expected_tokens / len(tokens)) * len(content))
|
||||
truncate_content = content[:truncate_len]
|
||||
truncate_tokens = t.encode(truncate_content)
|
||||
|
||||
# Count token
|
||||
print(
|
||||
f"city_name: {city_name}, #tokens: {len(tokens)}, #truncate tokens: {len(truncate_tokens)}"
|
||||
)
|
||||
|
||||
return truncate_content
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
with open("questions.jsonl", "w") as fout:
|
||||
for city_name in city_names:
|
||||
truncate_content = get_content(city_name)
|
||||
fout.write(json.dumps({"document": truncate_content}) + "\n")
|
||||
Reference in New Issue
Block a user