| 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::Glm4MoeParser as ExternalGlm4MoeParser; |
| 8 | use vllm_parser::tool::test_utils::{split_by_chars, test_tools}; |
| 9 | use vllm_parser::tool::{Glm45MoeToolParser, Glm47MoeToolParser, Tool, ToolParser}; |
| 10 | |
| 11 | mod utils; |
| 12 | use utils::{feed_external_parser, feed_parser, openai_tools}; |
| 13 | |
| 14 | const GLM45_PARSER_NAME: &str = "glm45"; |
| 15 | const GLM47_PARSER_NAME: &str = "glm47"; |
| 16 | const CHUNK_CHARS: usize = 7; |
| 17 | const LONG_NORMAL_TEXT_REPEATS: usize = 2048; |
| 18 | |
| 19 | fn glm45_mixed_fixture() -> String { |
| 20 | concat!( |
| 21 | "I will check two cities before answering.\n", |
| 22 | "<tool_call>get_weather\n", |
| 23 | "<arg_key>city</arg_key>\n", |
| 24 | "<arg_value>Hangzhou</arg_value>\n", |
| 25 | "<arg_key>date</arg_key>\n", |
| 26 | "<arg_value>2026-05-07</arg_value>\n", |
| 27 | "<arg_key>unit</arg_key>\n", |
| 28 | "<arg_value>celsius</arg_value>\n", |
| 29 | "<arg_key>days</arg_key>\n", |
| 30 | "<arg_value>3</arg_value>\n", |
| 31 | "</tool_call>\n", |
| 32 | "<tool_call>get_weather\n", |
| 33 | "<arg_key>city</arg_key>\n", |
| 34 | "<arg_value>San Francisco</arg_value>\n", |
| 35 | "<arg_key>date</arg_key>\n", |
| 36 | "<arg_value>2026-05-07</arg_value>\n", |
| 37 | "<arg_key>unit</arg_key>\n", |
| 38 | "<arg_value>fahrenheit</arg_value>\n", |
| 39 | "<arg_key>days</arg_key>\n", |
| 40 | "<arg_value>2</arg_value>\n", |
| 41 | "</tool_call>", |
| 42 | ) |
| 43 | .to_string() |
| 44 | } |
| 45 | |
| 46 | fn glm47_mixed_fixture() -> String { |
| 47 | concat!( |
| 48 | "I will check two cities before answering.\n", |
| 49 | "<tool_call>get_weather", |
| 50 | "<arg_key>city</arg_key>", |
| 51 | "<arg_value>Hangzhou</arg_value>", |
| 52 | "<arg_key>date</arg_key>", |
| 53 | "<arg_value>2026-05-07</arg_value>", |
| 54 | "<arg_key>unit</arg_key>", |
| 55 | "<arg_value>celsius</arg_value>", |
| 56 | "<arg_key>days</arg_key>", |
| 57 | "<arg_value>3</arg_value>", |
| 58 | "</tool_call>", |
| 59 | "<tool_call>get_weather", |
| 60 | "<arg_key>city</arg_key>", |
| 61 | "<arg_value>San Francisco</arg_value>", |
| 62 | "<arg_key>date</arg_key>", |
| 63 | "<arg_value>2026-05-07</arg_value>", |
| 64 | "<arg_key>unit</arg_key>", |
| 65 | "<arg_value>fahrenheit</arg_value>", |
| 66 | "<arg_key>days</arg_key>", |
| 67 | "<arg_value>2</arg_value>", |
| 68 | "</tool_call>", |
| 69 | ) |
| 70 | .to_string() |
| 71 | } |
| 72 | |
| 73 | fn long_normal_text_fixture() -> String { |
| 74 | let line = "This is ordinary assistant text with no GLM MoE tool markers at all.\n"; |
| 75 | line.repeat(LONG_NORMAL_TEXT_REPEATS) |
| 76 | } |
| 77 | |
| 78 | fn native_parser(name: &str, tools: &[Tool]) -> Box<dyn ToolParser> { |
| 79 | match name { |
| 80 | GLM45_PARSER_NAME => Glm45MoeToolParser::create(tools), |
| 81 | GLM47_PARSER_NAME => Glm47MoeToolParser::create(tools), |
| 82 | _ => unreachable!("unexpected GLM parser name"), |
| 83 | } |
| 84 | .expect("GLM MoE parser should initialize") |
| 85 | } |
| 86 | |
| 87 | fn external_parser(name: &str) -> ExternalGlm4MoeParser { |
| 88 | match name { |
| 89 | GLM45_PARSER_NAME => ExternalGlm4MoeParser::glm45(), |
| 90 | GLM47_PARSER_NAME => ExternalGlm4MoeParser::glm47(), |
| 91 | _ => unreachable!("unexpected GLM parser name"), |
| 92 | } |
| 93 | } |
| 94 | |
| 95 | fn run_stream_group( |
| 96 | c: &mut Criterion, |
| 97 | name: &str, |
| 98 | parser_name: &str, |
| 99 | tools: &[Tool], |
| 100 | text: &str, |
| 101 | chunk_chars: usize, |
| 102 | expected_normal_text: &str, |
| 103 | expected_native_calls_len: usize, |
| 104 | ) { |
| 105 | let chunks = split_by_chars(text, chunk_chars); |
| 106 | let openai_tools = openai_tools(tools); |
| 107 | |
| 108 | let mut group = c.benchmark_group(name); |
| 109 | group.sample_size(50); |
| 110 | group.warm_up_time(Duration::from_millis(300)); |
| 111 | group.measurement_time(Duration::from_secs(2)); |
| 112 | group.throughput(Throughput::Bytes(text.len() as u64)); |
| 113 | |
| 114 | group.bench_function("native_reuse_parser", |b| { |
| 115 | let mut parser = native_parser(parser_name, tools); |
| 116 | b.iter(|| { |
| 117 | let result = feed_parser(&mut *parser, black_box(&chunks)); |
| 118 | debug_assert_eq!(result.0, expected_normal_text); |
| 119 | debug_assert_eq!(result.1, expected_native_calls_len); |
| 120 | black_box(result); |
| 121 | }) |
| 122 | }); |
| 123 | |
| 124 | group.bench_function("native_create_parser", |b| { |
| 125 | b.iter_batched( |
| 126 | || native_parser(parser_name, tools), |
| 127 | |mut parser| { |
| 128 | let result = feed_parser(&mut *parser, black_box(&chunks)); |
| 129 | debug_assert_eq!(result.0, expected_normal_text); |
| 130 | debug_assert_eq!(result.1, expected_native_calls_len); |
| 131 | black_box(result); |
| 132 | }, |
| 133 | BatchSize::SmallInput, |
| 134 | ) |
| 135 | }); |
| 136 | |
| 137 | group.bench_function("external_reuse_parser", |b| { |
| 138 | let mut parser = external_parser(parser_name); |
| 139 | b.iter(|| { |
| 140 | let result = feed_external_parser(&mut parser, &openai_tools, black_box(&chunks)); |
| 141 | debug_assert_eq!(result.0, expected_normal_text); |
| 142 | black_box(result); |
| 143 | }) |
| 144 | }); |
| 145 | |
| 146 | group.bench_function("external_create_parser", |b| { |
| 147 | b.iter_batched( |
| 148 | || external_parser(parser_name), |
| 149 | |mut parser| { |
| 150 | let result = feed_external_parser(&mut parser, &openai_tools, black_box(&chunks)); |
| 151 | debug_assert_eq!(result.0, expected_normal_text); |
| 152 | black_box(result); |
| 153 | }, |
| 154 | BatchSize::SmallInput, |
| 155 | ) |
| 156 | }); |
| 157 | |
| 158 | group.finish(); |
| 159 | } |
| 160 | |
| 161 | fn bench_glm45_moe(c: &mut Criterion) { |
| 162 | let tools = test_tools(); |
| 163 | let glm45_mixed_text = glm45_mixed_fixture(); |
| 164 | let glm47_mixed_text = glm47_mixed_fixture(); |
| 165 | let long_normal_text = long_normal_text_fixture(); |
| 166 | |
| 167 | run_stream_group( |
| 168 | c, |
| 169 | "glm45/mixed_text_tool_call", |
| 170 | GLM45_PARSER_NAME, |
| 171 | &tools, |
| 172 | &glm45_mixed_text, |
| 173 | CHUNK_CHARS, |
| 174 | "I will check two cities before answering.\n", |
| 175 | 2, |
| 176 | ); |
| 177 | |
| 178 | run_stream_group( |
| 179 | c, |
| 180 | "glm47/mixed_text_tool_call", |
| 181 | GLM47_PARSER_NAME, |
| 182 | &tools, |
| 183 | &glm47_mixed_text, |
| 184 | CHUNK_CHARS, |
| 185 | "I will check two cities before answering.\n", |
| 186 | 2, |
| 187 | ); |
| 188 | |
| 189 | run_stream_group( |
| 190 | c, |
| 191 | "glm45/long_normal_text", |
| 192 | GLM45_PARSER_NAME, |
| 193 | &tools, |
| 194 | &long_normal_text, |
| 195 | CHUNK_CHARS, |
| 196 | &long_normal_text, |
| 197 | 0, |
| 198 | ); |
| 199 | |
| 200 | run_stream_group( |
| 201 | c, |
| 202 | "glm47/long_normal_text", |
| 203 | GLM47_PARSER_NAME, |
| 204 | &tools, |
| 205 | &long_normal_text, |
| 206 | CHUNK_CHARS, |
| 207 | &long_normal_text, |
| 208 | 0, |
| 209 | ); |
| 210 | } |
| 211 | |
| 212 | criterion_group!(benches, bench_glm45_moe); |
| 213 | criterion_main!(benches); |