@tduncan / Nvidia-Optimization-Stack-VLLM
Code Commits Issues Pulls Cloud Agent CI/CD
main
main
rs 139 lines 4.34 KB
Raw
1 // SPDX-License-Identifier: Apache-2.0
2 // SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4 use std::time::Duration;
5
6 use criterion::{BatchSize, Criterion, Throughput, black_box, criterion_group, criterion_main};
7 use tool_parser::parsers::MinimaxM2Parser as ExternalMinimaxM2Parser;
8 use vllm_parser::tool::test_utils::{split_by_chars, test_tools};
9 use vllm_parser::tool::{MinimaxM2ToolParser, Tool, ToolParser};
10
11 mod utils;
12 use utils::{feed_external_parser, feed_parser, openai_tools};
13
14 const CHUNK_CHARS: usize = 7;
15 const LONG_NORMAL_TEXT_REPEATS: usize = 2048;
16
17 fn mixed_fixture() -> String {
18 concat!(
19 "I will check two cities before answering.\n",
20 "<minimax:tool_call>",
21 "<invoke name=\"get_weather\">",
22 "<parameter name=\"city\">Hangzhou</parameter>",
23 "<parameter name=\"date\">2026-04-30</parameter>",
24 "<parameter name=\"unit\">celsius</parameter>",
25 "<parameter name=\"days\">3</parameter>",
26 "</invoke>",
27 "<invoke name=\"get_weather\">",
28 "<parameter name=\"city\">San Francisco</parameter>",
29 "<parameter name=\"date\">2026-04-30</parameter>",
30 "<parameter name=\"unit\">fahrenheit</parameter>",
31 "<parameter name=\"days\">2</parameter>",
32 "</invoke>",
33 "</minimax:tool_call>",
34 )
35 .to_string()
36 }
37
38 fn long_normal_text_fixture() -> String {
39 let line = "This is ordinary assistant text with no MiniMax M2 tool markers at all.\n";
40 line.repeat(LONG_NORMAL_TEXT_REPEATS)
41 }
42
43 fn native_parser(tools: &[Tool]) -> Box<dyn ToolParser> {
44 MinimaxM2ToolParser::create(tools).expect("MiniMax M2 parser should initialize")
45 }
46
47 fn run_stream_group(
48 c: &mut Criterion,
49 name: &str,
50 tools: &[Tool],
51 text: &str,
52 chunk_chars: usize,
53 expected_normal_text: &str,
54 expected_native_calls_len: usize,
55 ) {
56 let chunks = split_by_chars(text, chunk_chars);
57 let openai_tools = openai_tools(tools);
58
59 let mut group = c.benchmark_group(name);
60 group.sample_size(50);
61 group.warm_up_time(Duration::from_millis(300));
62 group.measurement_time(Duration::from_secs(2));
63 group.throughput(Throughput::Bytes(text.len() as u64));
64
65 group.bench_function("native_reuse_parser", |b| {
66 let mut parser = native_parser(tools);
67 b.iter(|| {
68 let result = feed_parser(&mut *parser, black_box(&chunks));
69 debug_assert_eq!(result.0, expected_normal_text);
70 debug_assert_eq!(result.1, expected_native_calls_len);
71 black_box(result);
72 })
73 });
74
75 group.bench_function("native_create_parser", |b| {
76 b.iter_batched(
77 || native_parser(tools),
78 |mut parser| {
79 let result = feed_parser(&mut *parser, black_box(&chunks));
80 debug_assert_eq!(result.0, expected_normal_text);
81 debug_assert_eq!(result.1, expected_native_calls_len);
82 black_box(result);
83 },
84 BatchSize::SmallInput,
85 )
86 });
87
88 group.bench_function("external_reuse_parser", |b| {
89 let mut parser = ExternalMinimaxM2Parser::new();
90 b.iter(|| {
91 let result = feed_external_parser(&mut parser, &openai_tools, black_box(&chunks));
92 debug_assert_eq!(result.0, expected_normal_text);
93 black_box(result);
94 })
95 });
96
97 group.bench_function("external_create_parser", |b| {
98 b.iter_batched(
99 ExternalMinimaxM2Parser::new,
100 |mut parser| {
101 let result = feed_external_parser(&mut parser, &openai_tools, black_box(&chunks));
102 debug_assert_eq!(result.0, expected_normal_text);
103 black_box(result);
104 },
105 BatchSize::SmallInput,
106 )
107 });
108
109 group.finish();
110 }
111
112 fn bench_minimax_m2(c: &mut Criterion) {
113 let tools = test_tools();
114 let mixed_text = mixed_fixture();
115 let long_normal_text = long_normal_text_fixture();
116
117 run_stream_group(
118 c,
119 "minimax_m2/mixed_text_tool_call",
120 &tools,
121 &mixed_text,
122 CHUNK_CHARS,
123 "I will check two cities before answering.\n",
124 2,
125 );
126
127 run_stream_group(
128 c,
129 "minimax_m2/long_normal_text",
130 &tools,
131 &long_normal_text,
132 CHUNK_CHARS,
133 &long_normal_text,
134 0,
135 );
136 }
137
138 criterion_group!(benches, bench_minimax_m2);
139 criterion_main!(benches);