feat: add support vertex-ai http bindings (#419)

* feat: add support vertex-ai http bindings

* support prefix / suffix

crates/http-api-bindings/README.md

@@ -1,4 +1,4 @@
-## Usage
+## Examples
 
 ```bash
 export MODEL_ID="code-gecko"
@@ -8,3 +8,14 @@ export AUTHORIZATION="Bearer $(gcloud auth print-access-token)"
 
 cargo run --example simple
 ```
+
+## Usage
+
+```bash
+export MODEL_ID="code-gecko"
+export PROJECT_ID="$(gcloud config get project)"
+export API_ENDPOINT="https://us-central1-aiplatform.googleapis.com/v1/projects/${PROJECT_ID}/locations/us-central1/publishers/google/models/${MODEL_ID}:predict"
+export AUTHORIZATION="Bearer $(gcloud auth print-access-token)"
+
+cargo run serve --device experimental-http --model "{\"kind\": \"vertex-ai\", \"api_endpoint\": \"$API_ENDPOINT\", \"authorization\": \"$AUTHORIZATION\"}"
+```

crates/http-api-bindings/src/vertex_ai.rs

@@ -56,23 +56,34 @@ impl VertexAIEngine {
             client,
         }
     }
+
+    pub fn prompt_template() -> String {
+        "{prefix}<MID>{suffix}".to_owned()
+    }
 }
 
 #[async_trait]
 impl TextGeneration for VertexAIEngine {
     async fn generate(&self, prompt: &str, options: TextGenerationOptions) -> String {
-        let stop_sequences: Vec<String> =
-            options.stop_words.iter().map(|x| x.to_string()).collect();
+        let stop_sequences: Vec<String> = options
+            .stop_words
+            .iter()
+            .map(|x| x.to_string())
+            // vertex supports at most 5 stop sequence.
+            .take(5)
+            .collect();
 
+        let tokens: Vec<&str> = prompt.split("<MID>").collect();
         let request = Request {
             instances: vec![Instance {
-                prefix: prompt.to_owned(),
-                suffix: None,
+                prefix: tokens[0].to_owned(),
+                suffix: Some(tokens[1].to_owned()),
             }],
             // options.max_input_length is ignored.
             parameters: Parameters {
                 temperature: options.sampling_temperature,
-                max_output_tokens: options.max_decoding_length,
+                // vertex supports at most 64 output tokens.
+                max_output_tokens: std::cmp::min(options.max_decoding_length, 64),
                 stop_sequences,
             },
         };

crates/tabby/Cargo.toml

@@ -35,6 +35,7 @@ tantivy = { workspace = true }
 anyhow = { workspace = true }
 sysinfo = "0.29.8"
 nvml-wrapper = "0.9.0"
+http-api-bindings = { path = "../http-api-bindings" }
 
 [target.'cfg(all(target_os="macos", target_arch="aarch64"))'.dependencies]
 llama-cpp-bindings = { path = "../llama-cpp-bindings" }

crates/tabby/src/serve/completions.rs

@@ -5,8 +5,10 @@ use std::{path::Path, sync::Arc};
 
 use axum::{extract::State, Json};
 use ctranslate2_bindings::{CTranslate2Engine, CTranslate2EngineOptionsBuilder};
+use http_api_bindings::vertex_ai::VertexAIEngine;
 use hyper::StatusCode;
 use serde::{Deserialize, Serialize};
+use serde_json::Value;
 use tabby_common::{config::Config, events, path::ModelDir};
 use tabby_inference::{TextGeneration, TextGenerationOptionsBuilder};
 use tracing::{debug, instrument};
@@ -128,22 +130,55 @@ pub struct CompletionState {
 
 impl CompletionState {
     pub fn new(args: &crate::serve::ServeArgs, config: &Config) -> Self {
-        let model_dir = get_model_dir(&args.model);
-        let metadata = read_metadata(&model_dir);
-        let engine = create_engine(args, &model_dir, &metadata);
+        let (engine, prompt_template) = create_engine(args);
 
         Self {
             engine,
             prompt_builder: prompt::PromptBuilder::new(
-                metadata.prompt_template,
+                prompt_template,
                 config.experimental.enable_prompt_rewrite,
             ),
         }
     }
 }
 
+fn create_engine(args: &crate::serve::ServeArgs) -> (Box<dyn TextGeneration>, Option<String>) {
+    if args.device != super::Device::ExperimentalHttp {
+        let model_dir = get_model_dir(&args.model);
+        let metadata = read_metadata(&model_dir);
+        let engine = create_local_engine(args, &model_dir, &metadata);
+        (engine, metadata.prompt_template)
+    } else {
+        let params: Value =
+            serdeconv::from_json_str(&args.model).expect("Failed to parse model string");
+
+        let kind = params
+            .get("kind")
+            .expect("Missing kind field")
+            .as_str()
+            .expect("Type unmatched");
+
+        if kind != "vertex-ai" {
+            fatal!("Only vertex_ai is supported for http backend");
+        }
+
+        let api_endpoint = params
+            .get("api_endpoint")
+            .expect("Missing api_endpoint field")
+            .as_str()
+            .expect("Type unmatched");
+        let authorization = params
+            .get("authorization")
+            .expect("Missing authorization field")
+            .as_str()
+            .expect("Type unmatched");
+        let engine = Box::new(VertexAIEngine::create(api_endpoint, authorization));
+        (engine, Some(VertexAIEngine::prompt_template()))
+    }
+}
+
 #[cfg(not(all(target_os = "macos", target_arch = "aarch64")))]
-fn create_engine(
+fn create_local_engine(
     args: &crate::serve::ServeArgs,
     model_dir: &ModelDir,
     metadata: &Metadata,
@@ -152,7 +187,7 @@ fn create_engine(
 }
 
 #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
-fn create_engine(
+fn create_local_engine(
     args: &crate::serve::ServeArgs,
     model_dir: &ModelDir,
     metadata: &Metadata,

crates/tabby/src/serve/mod.rs

@@ -14,7 +14,7 @@ use clap::Args;
 use tabby_common::{config::Config, usage};
 use tokio::time::sleep;
 use tower_http::cors::CorsLayer;
-use tracing::info;
+use tracing::{info, warn};
 use utoipa::OpenApi;
 use utoipa_swagger_ui::SwaggerUi;
 
@@ -58,6 +58,9 @@ pub enum Device {
     #[cfg(all(target_os = "macos", target_arch = "aarch64"))]
     #[strum(serialize = "metal")]
     Metal,
+
+    #[strum(serialize = "experimental_http")]
+    ExperimentalHttp,
 }
 
 #[derive(clap::ValueEnum, strum::Display, PartialEq, Clone)]
@@ -129,22 +132,28 @@ fn should_download_ggml_files(device: &Device) -> bool {
 pub async fn main(config: &Config, args: &ServeArgs) {
     valid_args(args);
 
-    // Ensure model exists.
-    tabby_download::download_model(
-        &args.model,
-        /* download_ctranslate2_files= */
-        !should_download_ggml_files(&args.device),
-        /* download_ggml_files= */ should_download_ggml_files(&args.device),
-        /* prefer_local_file= */ true,
-    )
-    .await
-    .unwrap_or_else(|err| {
-        fatal!(
-            "Failed to fetch model due to '{}', is '{}' a valid model id?",
-            err,
-            args.model
+    if args.device != Device::ExperimentalHttp {
+        let download_ctranslate2_files = !should_download_ggml_files(&args.device);
+        let download_ggml_files = should_download_ggml_files(&args.device);
+
+        // Ensure model exists.
+        tabby_download::download_model(
+            &args.model,
+            download_ctranslate2_files,
+            download_ggml_files,
+            /* prefer_local_file= */ true,
         )
-    });
+        .await
+        .unwrap_or_else(|err| {
+            fatal!(
+                "Failed to fetch model due to '{}', is '{}' a valid model id?",
+                err,
+                args.model
+            )
+        });
+    } else {
+        warn!("HTTP device is unstable and does not comply with semver expectations.")
+    }
 
     info!("Starting server, this might takes a few minutes...");
     let app = Router::new()