met 1.0.0-beta-preview

Metrale Engine: pure Rust LLM inference server (HTTP API, scheduler, TUI, CLI).
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101

// SPDX-License-Identifier: MIT OR Apache-2.0

use crate::tokenizer::ChatTokenizer;
use crate::tokenizer::HeapTrimPolicy;
use crate::tokenizer::TemplateSource::OverrideDir;
use serde_json::json;

#[test]
fn official_k3_chat_refuses_unimplemented_xtml_before_template_fallback() {
    let dir = tempfile::tempdir().unwrap();
    tokenizers::Tokenizer::new(tokenizers::models::wordlevel::WordLevel::default())
        .save(dir.path().join("tokenizer.json"), false)
        .unwrap();
    std::fs::write(
        dir.path().join("tokenizer_config.json"),
        json!({"tokenizer_class":"TikTokenTokenizer"}).to_string(),
    )
    .unwrap();
    let tokenizer = ChatTokenizer::from_model_dir(
        dir.path(),
        163585,
        true,
        "kimi_k3",
        None,
        OverrideDir,
        HeapTrimPolicy::off(),
    )
    .unwrap();
    assert!(tokenizer.uses_kimi_k3_xtml());
    let messages = [json!({"role":"user","content":"Hello"})];
    for thinking in [false, true] {
        for result in [
            tokenizer.apply_chat_template_jinja(&messages, None, thinking, false),
            tokenizer.apply_chat_template_openai(&messages, None, thinking, false),
        ] {
            let error = result.unwrap_err().to_string();
            assert!(error.contains("Kimi K3 XTML"), "unexpected error: {error}");
        }
    }
}

#[test]
fn official_k3_xtml_detection_does_not_confuse_the_twin() {
    let dir = tempfile::tempdir().unwrap();
    assert!(!super::super::kimi_k3::uses_xtml(dir.path(), "kimi_k3").unwrap());
    std::fs::write(
        dir.path().join("tokenizer_config.json"),
        json!({"tokenizer_class":"PreTrainedTokenizerFast", "chat_template":"{{ messages }}"})
            .to_string(),
    )
    .unwrap();
    assert!(!super::super::kimi_k3::uses_xtml(dir.path(), "kimi_k3").unwrap());
    tokenizers::Tokenizer::new(tokenizers::models::wordlevel::WordLevel::default())
        .save(dir.path().join("tokenizer.json"), false)
        .unwrap();
    let twin = ChatTokenizer::from_model_dir(
        dir.path(),
        0,
        false,
        "kimi_k3",
        None,
        OverrideDir,
        HeapTrimPolicy::off(),
    )
    .unwrap();
    assert!(!twin.uses_kimi_k3_xtml());
    std::fs::write(dir.path().join("tiktoken.model"), "fixture").unwrap();
    assert!(super::super::kimi_k3::uses_xtml(dir.path(), "kimi_k3").unwrap());
    assert!(!super::super::kimi_k3::uses_xtml(dir.path(), "qwen3").unwrap());
}

#[test]
fn official_k3_guard_also_covers_tool_requests_and_template_overrides() {
    let dir = tempfile::tempdir().unwrap();
    let templates = dir.path().join("jinja-templates/openai");
    std::fs::create_dir_all(&templates).unwrap();
    std::fs::write(templates.join("kimi_k3.jinja"), "wrong template").unwrap();
    std::fs::write(dir.path().join("tiktoken.model"), "fixture").unwrap();
    tokenizers::Tokenizer::new(tokenizers::models::wordlevel::WordLevel::default())
        .save(dir.path().join("tokenizer.json"), false)
        .unwrap();
    let tokenizer = ChatTokenizer::from_model_dir(
        dir.path(),
        163585,
        true,
        "kimi_k3",
        Some(dir.path()),
        OverrideDir,
        HeapTrimPolicy::off(),
    )
    .unwrap();
    let error = tokenizer
        .apply_chat_template_openai(
            &[json!({"role":"user","content":"Check weather"})],
            Some(&[json!({"type":"function","function":{"name":"weather"}})]),
            false,
            false,
        )
        .unwrap_err();
    assert!(error.to_string().contains("Kimi K3 XTML"));
}