mullama 0.3.0

Comprehensive Rust bindings for llama.cpp with memory-safe API and advanced features
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
143
144
145
146
147
148
149
150
151

use axum::{
    extract::{Json, State},
    response::{sse::Event, IntoResponse, Response},
};
use futures::stream::StreamExt as _;
use tokio_stream::wrappers::ReceiverStream;

use super::error::ApiError;
use super::helpers::{protocol_err_to_api, sse_response};
use super::types::{
    validate_n_parameter, ChatChoiceDelta, ChatCompletionChunk, ChatCompletionRequest,
    ChatCompletionResponse, DeltaContent,
};
use super::AppState;

/// POST /v1/chat/completions
pub(super) async fn chat_completions(
    State(daemon): State<AppState>,
    Json(req): Json<ChatCompletionRequest>,
) -> Result<Response, ApiError> {
    validate_n_parameter(req.n, "chat completions")?;

    let has_images = req.messages.iter().any(|m| m.content.has_images());

    if req.stream {
        if has_images {
            #[cfg(feature = "multimodal")]
            return chat_completions_vision_stream(daemon, req).await;
            #[cfg(not(feature = "multimodal"))]
            return Err(ApiError::new("Vision support requires multimodal feature"));
        }
        return chat_completions_stream(daemon, req).await;
    }

    if has_images {
        #[cfg(feature = "multimodal")]
        {
            let params = crate::daemon::protocol::ChatCompletionParams::from(req);
            match daemon.handle_vision_chat_completion(params).await {
                crate::daemon::protocol::Response::ChatCompletion(resp) => {
                    return Ok(Json(ChatCompletionResponse::from(resp)).into_response());
                }
                crate::daemon::protocol::Response::Error { code, message, .. } => {
                    return Err(ApiError::from_protocol_error(code, message));
                }
                _ => return Err(ApiError::new("Unexpected response")),
            }
        }
        #[cfg(not(feature = "multimodal"))]
        return Err(ApiError::new("Vision support requires multimodal feature"));
    }

    let params = crate::daemon::protocol::ChatCompletionParams::from(req);
    let request = crate::daemon::protocol::Request::ChatCompletion(params);

    match daemon.handle_request(request).await {
        crate::daemon::protocol::Response::ChatCompletion(resp) => {
            Ok(Json(ChatCompletionResponse::from(resp)).into_response())
        }
        crate::daemon::protocol::Response::Error { code, message, .. } => {
            Err(ApiError::from_protocol_error(code, message))
        }
        _ => Err(ApiError::new("Unexpected response")),
    }
}

/// Build an SSE response from a streaming chat completion receiver.
///
/// Shared by both text and vision chat completion streaming paths.
fn chat_stream_to_sse(
    rx: tokio::sync::mpsc::Receiver<crate::daemon::protocol::StreamChunk>,
    request_id: String,
    model_alias: String,
) -> Response {
    use futures::stream;

    let created = crate::daemon::protocol::unix_timestamp_secs();

    let stream = ReceiverStream::new(rx);
    let request_id_clone = request_id.clone();
    let model_clone = model_alias.clone();

    let event_stream = stream
        .map(move |chunk| {
            let sse_chunk = ChatCompletionChunk {
                id: request_id_clone.clone(),
                object: "chat.completion.chunk".to_string(),
                created,
                model: model_clone.clone(),
                choices: vec![ChatChoiceDelta {
                    index: chunk.index,
                    delta: DeltaContent {
                        role: if chunk.index == 0 {
                            Some("assistant".to_string())
                        } else {
                            None
                        },
                        content: Some(chunk.delta),
                    },
                    finish_reason: None,
                }],
            };
            Event::default().data(serde_json::to_string(&sse_chunk).unwrap_or_default())
        })
        .chain(stream::once(async move {
            let final_chunk = ChatCompletionChunk {
                id: request_id,
                object: "chat.completion.chunk".to_string(),
                created,
                model: model_alias,
                choices: vec![ChatChoiceDelta {
                    index: 0,
                    delta: DeltaContent {
                        role: None,
                        content: None,
                    },
                    finish_reason: Some("stop".to_string()),
                }],
            };
            Event::default().data(serde_json::to_string(&final_chunk).unwrap_or_default())
        }));

    sse_response(event_stream)
}

async fn chat_completions_stream(
    daemon: AppState,
    req: ChatCompletionRequest,
) -> Result<Response, ApiError> {
    let params = crate::daemon::protocol::ChatCompletionParams::from(req);
    let (rx, _prompt_tokens, request_id, model_alias) = daemon
        .handle_chat_completion_streaming(params)
        .await
        .map_err(protocol_err_to_api)?;

    Ok(chat_stream_to_sse(rx, request_id, model_alias))
}

#[cfg(feature = "multimodal")]
async fn chat_completions_vision_stream(
    daemon: AppState,
    req: ChatCompletionRequest,
) -> Result<Response, ApiError> {
    let params = crate::daemon::protocol::ChatCompletionParams::from(req);
    let (rx, _prompt_tokens, request_id, model_alias) = daemon
        .handle_vision_chat_completion_streaming(params)
        .await
        .map_err(protocol_err_to_api)?;

    Ok(chat_stream_to_sse(rx, request_id, model_alias))
}