mistral.rs server core

About

This crate powers mistral.rs server. It exposes the underlying functionality allowing others to implement and extend the server implementation.

Features

1. Incorporate mistral.rs server into another axum.rs project.
2. Hook into the mistral.rs server lifecycle.

Example

use std::sync::Arc;

use axum::{
    extract::State,
    routing::{get, post},
    Json, Router,
};
use utoipa::OpenApi;
use utoipa_swagger_ui::SwaggerUi;

use mistralrs_core::{
    initialize_logging, AutoDeviceMapParams, ChatCompletionChunkResponse, ModelDType, ModelSelected,
};
use mistralrs_server_core::{
    chat_completion::{
        create_streamer, handle_error, parse_request, process_non_streaming_response,
        ChatCompletionOnChunkCallback, ChatCompletionOnDoneCallback, ChatCompletionResponder,
    },
    handler_core::{create_response_channel, send_request},
    mistralrs_for_server_builder::MistralRsForServerBuilder,
    mistralrs_server_router_builder::MistralRsServerRouterBuilder,
    openai::ChatCompletionRequest,
    openapi_doc::get_openapi_doc,
    types::SharedMistralRsState,
};

#[derive(OpenApi)]
#[openapi(
    paths(root, custom_chat),
    tags(
        (name = "hello", description = "Hello world endpoints")
    ),
    info(
        title = "Hello World API",
        version = "1.0.0",
        description = "A simple API that responds with a greeting"
    )
)]
struct ApiDoc;

#[derive(Clone)]
pub struct AppState {
    pub mistralrs_state: SharedMistralRsState,
    pub db_create: fn(),
}

#[tokio::main]
async fn main() {
    initialize_logging();

    let plain_model_id = String::from("meta-llama/Llama-3.2-1B-Instruct");
    let tokenizer_json = None;
    let arch = None;
    let organization = None;
    let write_uqff = None;
    let from_uqff = None;
    let imatrix = None;
    let calibration_file = None;
    let hf_cache_path = None;

    let dtype = ModelDType::Auto;
    let topology = None;
    let max_seq_len = AutoDeviceMapParams::DEFAULT_MAX_SEQ_LEN;
    let max_batch_size = AutoDeviceMapParams::DEFAULT_MAX_BATCH_SIZE;
    let matformer_config_path = None;
    let matformer_slice_name = None;

    let model = ModelSelected::Plain {
        model_id: plain_model_id,
        tokenizer_json,
        arch,
        dtype,
        topology,
        organization,
        write_uqff,
        from_uqff,
        imatrix,
        calibration_file,
        max_seq_len,
        max_batch_size,
        hf_cache_path,
        matformer_config_path,
        matformer_slice_name,
    };

    let shared_mistralrs = MistralRsForServerBuilder::new()
        .with_model(model)
        .with_in_situ_quant("8".to_string())
        .set_paged_attn(Some(true))
        .build()
        .await
        .unwrap();

    let mistralrs_base_path = "/api/mistral";

    let mistralrs_routes = MistralRsServerRouterBuilder::new()
        .with_mistralrs(shared_mistralrs.clone())
        .with_include_swagger_routes(false)
        .with_base_path(mistralrs_base_path)
        .build()
        .await
        .unwrap();

    let mistralrs_doc = get_openapi_doc(Some(mistralrs_base_path));
    let mut api_docs = ApiDoc::openapi();
    api_docs.merge(mistralrs_doc);

    let app_state = Arc::new(AppState {
        mistralrs_state: shared_mistralrs,
        db_create: mock_db_call,
    });

    let app = Router::new()
        .route("/", get(root))
        .route("/chat", post(custom_chat))
        .with_state(app_state.clone())
        .nest(mistralrs_base_path, mistralrs_routes)
        .merge(SwaggerUi::new("/api-docs").url("/api-docs/openapi.json", api_docs));

    let listener = tokio::net::TcpListener::bind("0.0.0.0:3000").await.unwrap();
    axum::serve(listener, app).await.unwrap();

    println!("Listening on 0.0.0.0:3000");
}

#[utoipa::path(
    get,
    path = "/",
    tag = "hello",
    responses(
        (status = 200, description = "Successful response with greeting message", body = String)
    )
)]
async fn root() -> &'static str {
    "Hello, World!"
}

#[utoipa::path(
    post,
    tag = "Custom",
    path = "/chat",
    request_body = ChatCompletionRequest,
    responses((status = 200, description = "Chat completions"))
)]
pub async fn custom_chat(
    State(state): State<Arc<AppState>>,
    Json(oai_request): Json<ChatCompletionRequest>,
) -> ChatCompletionResponder {
    let mistralrs_state = state.mistralrs_state.clone();
    let (tx, mut rx) = create_response_channel(None);

    let (request, is_streaming) =
        match parse_request(oai_request, mistralrs_state.clone(), tx).await {
            Ok(x) => x,
            Err(e) => return handle_error(mistralrs_state, e.into()),
        };

    dbg!(request.clone());

    if let Err(e) = send_request(&mistralrs_state, request).await {
        return handle_error(mistralrs_state, e.into());
    }

    if is_streaming {
        let db_fn = state.db_create;

        let on_chunk: ChatCompletionOnChunkCallback =
            Box::new(move |mut chunk: ChatCompletionChunkResponse| {
                dbg!(&chunk);

                if let Some(original_content) = &chunk.choices[0].delta.content {
                    chunk.choices[0].delta.content = Some(format!("CHANGED! {}", original_content));
                }

                chunk.clone()
            });

        let on_done: ChatCompletionOnDoneCallback =
            Box::new(move |chunks: &[ChatCompletionChunkResponse]| {
                dbg!(chunks);
                (db_fn)();
            });

        let streamer = create_streamer(rx, mistralrs_state.clone(), Some(on_chunk), Some(on_done));

        ChatCompletionResponder::Sse(streamer)
    } else {
        let response = process_non_streaming_response(&mut rx, mistralrs_state.clone()).await;

        match &response {
            ChatCompletionResponder::Json(json_response) => {
                dbg!(json_response);
                (state.db_create)();
            }
            _ => {
                //
            }
        }

        response
    }
}

pub fn mock_db_call() {
    println!("Saving to DB");
}

mistralrs-server-core

Core crate that powers mistral.rs server.

Documentation: https://docs.rs/mistralrs-server-core/