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
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142

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

//! 2026-09-26: Split a model's completed output into reasoning and answer, per reasoning-block format.
//!
//! | Format | Delimiters |
//! |---|---|
//! | [`ReasoningFormat::Qwen`], [`ReasoningFormat::DeepSeekR1`], [`ReasoningFormat::MiniMax`] | `<think>` / `</think>` |
//! | [`ReasoningFormat::Mistral`] | `[THINK]` / `[/THINK]` |
//! | [`ReasoningFormat::Gemma4`] | `<|channel>` / `<channel|>` |
//! | [`ReasoningFormat::Glm47`] | `<think>` / `</think>` or `<tool_call>` |
//!
//! The three `<think>` formats share one implementation and differ only in
//! name; they assume the prompt already opened the block. Mistral uses the
//! same implementation but expects the model to emit its own `[THINK]`.
//! Gemma 4 has its own parser, and so has GLM-4.7 (`glm47.rs`), whose block
//! a tool call also closes. At serve time the format comes from the
//! `[reasoning]` table of `tool_defaults.toml`, keyed by `model_type`, and a
//! thinking-capable model without an entry gets `qwen`
//! (`main_modules/serve_phases/tokenizer_runtime.rs`).
//!
//! Owner: server.
//! Invariants: none beyond the types.

mod glm47;
#[cfg(test)]
mod glm47_tests;
mod parsers;
#[cfg(test)]
mod tests;

use std::str::FromStr;

use crate::tokenizer::ChatTokenizer;

/// 2026-10-14: The chat-response field that carries the reasoning text.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum ReasoningWire {
    /// `reasoning_content`, left out when there is no reasoning.
    ReasoningContent,
    /// vLLM's (`ChatMessage.reasoning`, `DeltaMessage.reasoning`):
    /// `reasoning`. The blocking message always has it, null when there is no
    /// reasoning (`model_dump()`); a stream delta has it only when it carries
    /// reasoning (`exclude_unset=True`).
    Vllm,
}

/// 2026-09-26: Splits reasoning blocks out of completed model output.
pub trait ReasoningParser: Send + Sync {
    /// 2026-09-26: Parser name for logging (e.g. `"qwen"`, `"deepseek_r1"`).
    fn name(&self) -> &str;

    /// 2026-09-26: Opening delimiter (e.g. `"<think>"`, `"[THINK]"`, `"<|channel>"`).
    fn start_tag(&self) -> &str;

    /// 2026-09-26: Closing delimiter (e.g. `"</think>"`, `"[/THINK]"`, `"<channel|>"`).
    fn end_tag(&self) -> &str;

    /// 2026-09-26: The end tag's token id, or `None` unless the tag encodes to
    /// exactly one token.
    fn end_token_id(&self, tokenizer: &ChatTokenizer) -> Option<u32> {
        match tokenizer.encode(self.end_tag()) {
            Ok(ids) if ids.len() == 1 => Some(ids[0]),
            _ => None,
        }
    }

    /// 2026-10-08: Whether the format's tool-call opener (`<tool_call>`) also
    /// closes the reasoning block, the opener and what follows being the
    /// answer. Such a parser keeps marker text inside a tool call, where a
    /// call that swallowed a `</think>` must reach the tool parser intact, so
    /// callers split with its `extract_thinking` on text and do not run the
    /// shared marker scrub (`api/chat_blocking.rs`, `api/chat_stream`).
    /// Default false.
    fn tool_call_closes_reasoning(&self) -> bool {
        false
    }

    /// 2026-10-14: Whether the streamed deltas get vLLM's parser-engine
    /// whitespace shaping (`api/chat_stream/vllm_shape.rs`). Default false.
    fn shapes_stream_whitespace(&self) -> bool {
        false
    }

    /// 2026-10-14: How a chat response carries this format's reasoning text
    /// (`openai::WireReasoning`). Default `ReasoningContent`.
    fn reasoning_wire(&self) -> ReasoningWire {
        ReasoningWire::ReasoningContent
    }

    /// 2026-09-26: Split completed text into `(reasoning, content)`.
    ///
    /// `enable_thinking` is the request's resolved thinking state. When it is
    /// `false` the reasoning is `None`; the answer is always in `content`.
    fn extract_thinking(&self, text: &str, enable_thinking: bool) -> (Option<String>, String);
}

/// 2026-09-26: Supported reasoning-block formats; the module doc lists each
/// one's delimiters.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum ReasoningFormat {
    /// 2026-09-26: `<think>...</think>`, parsed as `qwen`.
    Qwen,
    /// 2026-09-26: `<think>...</think>`, parsed as `deepseek_r1`; also named
    /// `deepseek`, `nemotron`, `nemotron_h` and `nano_v3`.
    DeepSeekR1,
    /// 2026-09-26: `<think>...</think>`, parsed as `minimax`.
    MiniMax,
    /// 2026-09-26: `[THINK]...[/THINK]`, opened by the model.
    Mistral,
    /// 2026-09-26: `<|channel>thought ... <channel|> ...`, Gemma 4's channel format.
    Gemma4,
    /// 2026-10-08: `<think>...</think>`, which `<tool_call>` also closes, parsed
    /// as `glm47`; also named `glm45`.
    Glm47,
}

impl FromStr for ReasoningFormat {
    type Err = String;
    fn from_str(s: &str) -> Result<Self, Self::Err> {
        match s.to_lowercase().as_str() {
            "qwen" | "qwen3" => Ok(Self::Qwen),
            "deepseek_r1" | "deepseek" | "nemotron" | "nemotron_h" | "nano_v3" => {
                Ok(Self::DeepSeekR1)
            }
            "minimax" | "minimax_m2" => Ok(Self::MiniMax),
            "mistral" => Ok(Self::Mistral),
            "gemma4" | "gemma" => Ok(Self::Gemma4),
            "glm47" | "glm45" => Ok(Self::Glm47),
            other => Err(format!(
                "Unknown reasoning parser '{other}'. Supported: qwen, \
                 deepseek_r1, minimax, mistral, gemma4, glm47"
            )),
        }
    }
}

impl ReasoningFormat {
    /// 2026-09-26: A boxed parser for this format.
    pub fn into_parser(self) -> Box<dyn ReasoningParser> {
        parsers::build(self)
    }
}