refactor(core): embed llama.cpp's server binary directly for LLM infe…

…rence (#2113)

* chore: add llama-cpp-server sub crate

* chore: add llama-cpp-server to embed llama-server directly

* update

* cleanup

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

Cargo.toml

@@ -13,7 +13,7 @@ members = [
     "ee/tabby-db",
     "ee/tabby-db-macros",
     "ee/tabby-search",
-    "ee/tabby-schema",
+    "ee/tabby-schema", "crates/llama-cpp-server",
 ]
 
 [workspace.package]

crates/llama-cpp-server/Cargo.toml

@@ -0,0 +1,26 @@
+[package]
+name = "llama-cpp-server"
+version.workspace = true
+edition.workspace = true
+authors.workspace = true
+homepage.workspace = true
+
+[features]
+cuda = []
+rocm = []
+vulkan = []
+
+[dependencies]
+futures.workspace = true
+http-api-bindings = { path = "../http-api-bindings" }
+reqwest.workspace = true
+serde_json.workspace = true
+tabby-inference = { path = "../tabby-inference" }
+tracing.workspace = true
+async-trait.workspace = true
+tokio = { workspace = true, features = ["process"] }
+anyhow.workspace = true
+
+[build-dependencies]
+cmake = "0.1"
+omnicopy_to_output = "0.1.1"

crates/llama-cpp-server/build.rs

@@ -0,0 +1,79 @@
+use std::{env, path::Path};
+
+use cmake::Config;
+use omnicopy_to_output::copy_to_output;
+
+fn main() {
+    let mut config = Config::new("../llama-cpp-bindings/llama.cpp");
+    config.profile("Release");
+    config.define("LLAMA_NATIVE", "OFF");
+    config.define("INS_ENB", "ON");
+
+    if cfg!(target_os = "macos") {
+        config.define("LLAMA_METAL", "ON");
+        config.define("LLAMA_METAL_EMBED_LIBRARY", "ON");
+        println!("cargo:rustc-link-lib=framework=Foundation");
+        println!("cargo:rustc-link-lib=framework=Accelerate");
+        println!("cargo:rustc-link-lib=framework=Metal");
+        println!("cargo:rustc-link-lib=framework=MetalKit");
+    }
+    if cfg!(feature = "cuda") {
+        config.define("LLAMA_CUBLAS", "ON");
+        config.define("CMAKE_POSITION_INDEPENDENT_CODE", "ON");
+    }
+    if cfg!(feature = "rocm") {
+        let amd_gpu_targets: Vec<&str> = vec![
+            "gfx803",
+            "gfx900",
+            "gfx906:xnack-",
+            "gfx908:xnack-",
+            "gfx90a:xnack+",
+            "gfx90a:xnack-",
+            "gfx940",
+            "gfx941",
+            "gfx942",
+            "gfx1010",
+            "gfx1012",
+            "gfx1030",
+            "gfx1031",
+            "gfx1100",
+            "gfx1101",
+            "gfx1102",
+            "gfx1103",
+        ];
+
+        let rocm_root = env::var("ROCM_ROOT").unwrap_or("/opt/rocm".to_string());
+        config.define("LLAMA_HIPBLAS", "ON");
+        config.define("CMAKE_C_COMPILER", format!("{}/llvm/bin/clang", rocm_root));
+        config.define(
+            "CMAKE_CXX_COMPILER",
+            format!("{}/llvm/bin/clang++", rocm_root),
+        );
+        config.define("AMDGPU_TARGETS", amd_gpu_targets.join(";"));
+    }
+    if cfg!(feature = "vulkan") {
+        config.define("LLAMA_VULKAN", "ON");
+    }
+
+    let out = config.build();
+    let server_binary = make_output_binary(&out, "server");
+    let renamed_server_binary = if cfg!(target_os = "macos") {
+        make_output_binary(&out, "llama-server-metal")
+    } else if cfg!(feature = "cuda") {
+        make_output_binary(&out, "llama-server-cuda")
+    } else if cfg!(feature = "rocm") {
+        make_output_binary(&out, "llama-server-rocm")
+    } else if cfg!(feature = "vulkan") {
+        make_output_binary(&out, "llama-server-vulkan")
+    } else {
+        make_output_binary(&out, "llama-server")
+    };
+
+    std::fs::rename(server_binary, &renamed_server_binary).expect("Failed to rename server binary");
+    copy_to_output(&renamed_server_binary)
+        .expect("Failed to copy server binary to output directory");
+}
+
+fn make_output_binary(out: &Path, name: &str) -> String {
+    out.join("bin").join(name).display().to_string() + env::consts::EXE_SUFFIX
+}

crates/llama-cpp-server/src/lib.rs

@@ -0,0 +1,180 @@
+use std::{net::TcpListener, process::Stdio, sync::Arc};
+
+use anyhow::Result;
+use async_trait::async_trait;
+use futures::stream::BoxStream;
+use serde_json::json;
+use tabby_inference::{CompletionOptions, CompletionStream, Embedding};
+use tokio::task::JoinHandle;
+use tracing::warn;
+
+pub struct LlamaCppServer {
+    port: u16,
+    handle: JoinHandle<()>,
+    completion: Arc<dyn CompletionStream>,
+    embedding: Arc<dyn Embedding>,
+}
+
+#[async_trait]
+impl CompletionStream for LlamaCppServer {
+    async fn generate(&self, prompt: &str, options: CompletionOptions) -> BoxStream<String> {
+        self.completion.generate(prompt, options).await
+    }
+}
+
+#[async_trait]
+impl Embedding for LlamaCppServer {
+    async fn embed(&self, prompt: &str) -> Result<Vec<f32>> {
+        self.embedding.embed(prompt).await
+    }
+}
+
+impl LlamaCppServer {
+    pub fn new(device: &str, model_path: &str, parallelism: u8) -> Self {
+        let use_gpu = device != "cpu";
+        let Some(binary_name) = find_binary_name(Some(device)) else {
+            panic!("Failed to find llama-server binary for device {device}, please make sure you have corresponding llama-server binary locates in the same directory as the current executable.");
+        };
+
+        let model_path = model_path.to_owned();
+        let port = get_available_port();
+        let handle = tokio::spawn(async move {
+            loop {
+                let server_binary = std::env::current_exe()
+                    .expect("Failed to get current executable path")
+                    .parent()
+                    .expect("Failed to get parent directory")
+                    .join(&binary_name)
+                    .display()
+                    .to_string()
+                    + std::env::consts::EXE_SUFFIX;
+                let mut command = tokio::process::Command::new(server_binary);
+
+                command
+                    .arg("-m")
+                    .arg(&model_path)
+                    .arg("--port")
+                    .arg(port.to_string())
+                    .arg("-np")
+                    .arg(parallelism.to_string())
+                    .arg("--log-disable")
+                    .kill_on_drop(true)
+                    .stderr(Stdio::null())
+                    .stdout(Stdio::null());
+
+                if let Ok(n_threads) = std::env::var("LLAMA_CPP_N_THREADS") {
+                    command.arg("-t").arg(n_threads);
+                }
+
+                if use_gpu {
+                    let num_gpu_layers =
+                        std::env::var("LLAMA_CPP_N_GPU_LAYERS").unwrap_or("9999".into());
+                    command.arg("-ngl").arg(&num_gpu_layers);
+                }
+
+                let mut process = command.spawn().unwrap_or_else(|e| {
+                    panic!(
+                        "Failed to start llama-server with command {:?}: {}",
+                        command, e
+                    )
+                });
+
+                let status_code = process
+                    .wait()
+                    .await
+                    .ok()
+                    .and_then(|s| s.code())
+                    .unwrap_or(-1);
+
+                if status_code != 0 {
+                    warn!(
+                        "llama-server exited with status code {}, restarting...",
+                        status_code
+                    );
+                }
+            }
+        });
+
+        Self {
+            handle,
+            port,
+            completion: make_completion(port),
+            embedding: make_embedding(port),
+        }
+    }
+
+    pub async fn start(&self) {
+        let client = reqwest::Client::new();
+        loop {
+            let Ok(resp) = client.get(api_endpoint(self.port) + "/health").send().await else {
+                continue;
+            };
+
+            if resp.status().is_success() {
+                return;
+            }
+        }
+    }
+}
+
+fn find_binary_name(suffix: Option<&str>) -> Option<String> {
+    let current_exe = std::env::current_exe().expect("Failed to get current executable path");
+    let binary_dir = current_exe
+        .parent()
+        .expect("Failed to get parent directory");
+    let binary_name = if let Some(suffix) = suffix {
+        format!("llama-server-{}", suffix)
+    } else {
+        "llama-server".to_owned()
+    };
+    std::fs::read_dir(binary_dir)
+        .expect("Failed to read directory")
+        .filter_map(|entry| entry.ok())
+        .filter(|entry| {
+            entry
+                .file_name()
+                .to_string_lossy()
+                .starts_with(&binary_name)
+        })
+        .map(|entry| entry.path().display().to_string())
+        .next()
+}
+
+fn make_completion(port: u16) -> Arc<dyn CompletionStream> {
+    let model_spec: String = serde_json::to_string(&json!({
+        "kind": "llama",
+        "api_endpoint": api_endpoint(port),
+    }))
+    .expect("Failed to serialize model spec");
+    let (engine, _, _) = http_api_bindings::create(&model_spec);
+    engine
+}
+
+pub fn make_embedding(port: u16) -> Arc<dyn Embedding> {
+    let model_spec: String = serde_json::to_string(&json!({
+        "kind": "llama",
+        "api_endpoint": api_endpoint(port),
+    }))
+    .expect("Failed to serialize model spec");
+    http_api_bindings::create_embedding(&model_spec)
+}
+
+fn get_available_port() -> u16 {
+    (30888..40000)
+        .find(|port| port_is_available(*port))
+        .expect("Failed to find available port")
+}
+
+fn port_is_available(port: u16) -> bool {
+    TcpListener::bind(("127.0.0.1", port)).is_ok()
+}
+
+impl Drop for LlamaCppServer {
+    fn drop(&mut self) {
+        self.handle.abort();
+    }
+}
+
+fn api_endpoint(port: u16) -> String {
+    format!("http://localhost:{port}")
+}