feat: add llama.cpp backend (#94)

* feat: add `llama.cpp` backend

* fix(ci): install stable toolchain instead of nightly

* fix(ci): use different model

.github/workflows/release.yml

@@ -12,7 +12,6 @@ env:
   RUSTFLAGS: "-D warnings -W unreachable-pub"
   RUSTUP_MAX_RETRIES: 10
   FETCH_DEPTH: 0 # pull in the tags for the version string
-  MACOSX_DEPLOYMENT_TARGET: 10.15
   CARGO_TARGET_AARCH64_UNKNOWN_LINUX_GNU_LINKER: aarch64-linux-gnu-gcc
   CARGO_TARGET_ARM_UNKNOWN_LINUX_GNUEABIHF_LINKER: arm-linux-gnueabihf-gcc
 

.github/workflows/test.yml

@@ -40,10 +40,7 @@ jobs:
           DEBIAN_FRONTEND=noninteractive apt install -y pkg-config protobuf-compiler libssl-dev curl build-essential git-all gfortran
 
       - name: Install Rust toolchain
-        uses: actions-rust-lang/setup-rust-toolchain@v1
-        with:
-          rustflags: ''
-          toolchain: nightly
+        uses: dtolnay/rust-toolchain@stable
 
       - name: Install Python 3.10
         uses: actions/setup-python@v5

crates/custom-types/src/llm_ls.rs

@@ -67,10 +67,9 @@ pub enum Backend {
         #[serde(default = "hf_default_url", deserialize_with = "parse_url")]
         url: String,
     },
-    // TODO:
-    // LlamaCpp {
-    //   url: String,
-    // },
+    LlamaCpp {
+        url: String,
+    },
     Ollama {
         url: String,
     },

crates/llm-ls/src/backend.rs

@@ -67,6 +67,37 @@ fn parse_api_text(text: &str) -> Result<Vec<Generation>> {
     }
 }
 
+#[derive(Debug, Serialize, Deserialize)]
+struct LlamaCppGeneration {
+    content: String,
+}
+
+impl From<LlamaCppGeneration> for Generation {
+    fn from(value: LlamaCppGeneration) -> Self {
+        Generation {
+            generated_text: value.content,
+        }
+    }
+}
+
+#[derive(Debug, Deserialize)]
+#[serde(untagged)]
+enum LlamaCppAPIResponse {
+    Generation(LlamaCppGeneration),
+    Error(APIError),
+}
+
+fn build_llamacpp_headers() -> HeaderMap {
+    HeaderMap::new()
+}
+
+fn parse_llamacpp_text(text: &str) -> Result<Vec<Generation>> {
+    match serde_json::from_str(text)? {
+        LlamaCppAPIResponse::Generation(gen) => Ok(vec![gen.into()]),
+        LlamaCppAPIResponse::Error(err) => Err(Error::LlamaCpp(err)),
+    }
+}
+
 #[derive(Debug, Serialize, Deserialize)]
 struct OllamaGeneration {
     response: String,
@@ -192,6 +223,9 @@ pub(crate) fn build_body(
                 request_body.insert("parameters".to_owned(), params);
             }
         }
+        Backend::LlamaCpp { .. } => {
+            request_body.insert("prompt".to_owned(), Value::String(prompt));
+        }
         Backend::Ollama { .. } | Backend::OpenAi { .. } => {
             request_body.insert("prompt".to_owned(), Value::String(prompt));
             request_body.insert("model".to_owned(), Value::String(model));
@@ -208,6 +242,7 @@ pub(crate) fn build_headers(
 ) -> Result<HeaderMap> {
     match backend {
         Backend::HuggingFace { .. } => build_api_headers(api_token, ide),
+        Backend::LlamaCpp { .. } => Ok(build_llamacpp_headers()),
         Backend::Ollama { .. } => Ok(build_ollama_headers()),
         Backend::OpenAi { .. } => build_openai_headers(api_token, ide),
         Backend::Tgi { .. } => build_tgi_headers(api_token, ide),
@@ -217,6 +252,7 @@ pub(crate) fn build_headers(
 pub(crate) fn parse_generations(backend: &Backend, text: &str) -> Result<Vec<Generation>> {
     match backend {
         Backend::HuggingFace { .. } => parse_api_text(text),
+        Backend::LlamaCpp { .. } => parse_llamacpp_text(text),
         Backend::Ollama { .. } => parse_ollama_text(text),
         Backend::OpenAi { .. } => parse_openai_text(text),
         Backend::Tgi { .. } => parse_tgi_text(text),

crates/llm-ls/src/document.rs

@@ -168,7 +168,7 @@ impl TryFrom<Vec<tower_lsp::lsp_types::PositionEncodingKind>> for PositionEncodi
 }
 
 impl PositionEncodingKind {
-    pub fn to_lsp_type(&self) -> tower_lsp::lsp_types::PositionEncodingKind {
+    pub fn to_lsp_type(self) -> tower_lsp::lsp_types::PositionEncodingKind {
         match self {
             PositionEncodingKind::Utf8 => tower_lsp::lsp_types::PositionEncodingKind::UTF8,
             PositionEncodingKind::Utf16 => tower_lsp::lsp_types::PositionEncodingKind::UTF16,
@@ -205,9 +205,10 @@ impl Document {
     ) -> Result<()> {
         match change.range {
             Some(range) => {
-                if range.start.line < range.end.line
+                if range.start.line > range.end.line
                     || (range.start.line == range.end.line
-                        && range.start.character <= range.end.character) {
+                        && range.start.character > range.end.character)
+                {
                     return Err(Error::InvalidRange(range));
                 }
 
@@ -219,7 +220,10 @@ impl Document {
 
                 // 1. Get the line at which the change starts.
                 let change_start_line_idx = range.start.line as usize;
-                let change_start_line = self.text.get_line(change_start_line_idx).ok_or_else(|| Error::OutOfBoundLine(change_start_line_idx, self.text.len_lines()))?;
+                let change_start_line =
+                    self.text.get_line(change_start_line_idx).ok_or_else(|| {
+                        Error::OutOfBoundLine(change_start_line_idx, self.text.len_lines())
+                    })?;
 
                 // 2. Get the line at which the change ends. (Small optimization
                 // where we first check whether start and end line are the
@@ -228,7 +232,9 @@ impl Document {
                 let change_end_line_idx = range.end.line as usize;
                 let change_end_line = match same_line {
                     true => change_start_line,
-                    false => self.text.get_line(change_end_line_idx).ok_or_else(|| Error::OutOfBoundLine(change_end_line_idx, self.text.len_lines()))?,
+                    false => self.text.get_line(change_end_line_idx).ok_or_else(|| {
+                        Error::OutOfBoundLine(change_end_line_idx, self.text.len_lines())
+                    })?,
                 };
 
                 fn compute_char_idx(
@@ -330,7 +336,7 @@ impl Document {
                             self.tree = Some(new_tree);
                         }
                         None => {
-                            return Err(Error::TreeSitterParseError);
+                            return Err(Error::TreeSitterParsing);
                         }
                     }
                 }
@@ -416,7 +422,9 @@ mod test {
         let mut rope = Rope::from_str(
             "let a = '🥸 你好';\rfunction helloWorld() { return '🤲🏿'; }\nlet b = 'Hi, 😊';",
         );
-        let mut doc = Document::open(&LanguageId::JavaScript.to_string(), &rope.to_string()).await.unwrap();
+        let mut doc = Document::open(&LanguageId::JavaScript.to_string(), &rope.to_string())
+            .await
+            .unwrap();
         let mut parser = Parser::new();
 
         parser
@@ -464,7 +472,9 @@ mod test {
     #[tokio::test]
     async fn test_text_document_apply_content_change_bounds() {
         let rope = Rope::from_str("");
-        let mut doc = Document::open(&LanguageId::Unknown.to_string(), &rope.to_string()).await.unwrap();
+        let mut doc = Document::open(&LanguageId::Unknown.to_string(), &rope.to_string())
+            .await
+            .unwrap();
 
         assert!(doc
             .apply_content_change(new_change!(0, 0, 0, 1, ""), PositionEncodingKind::Utf16)
@@ -513,7 +523,9 @@ mod test {
     async fn test_document_update_tree_consistency_easy() {
         let a = "let a = '你好';\rlet b = 'Hi, 😊';";
 
-        let mut document = Document::open(&LanguageId::JavaScript.to_string(), a).await.unwrap();
+        let mut document = Document::open(&LanguageId::JavaScript.to_string(), a)
+            .await
+            .unwrap();
 
         document
             .apply_content_change(new_change!(0, 9, 0, 11, "𐐀"), PositionEncodingKind::Utf16)
@@ -541,7 +553,9 @@ mod test {
     async fn test_document_update_tree_consistency_medium() {
         let a = "let a = '🥸 你好';\rfunction helloWorld() { return '🤲🏿'; }\nlet b = 'Hi, 😊';";
 
-        let mut document = Document::open(&LanguageId::JavaScript.to_string(), a).await.unwrap();
+        let mut document = Document::open(&LanguageId::JavaScript.to_string(), a)
+            .await
+            .unwrap();
 
         document
             .apply_content_change(new_change!(0, 14, 2, 13, "，"), PositionEncodingKind::Utf16)

crates/llm-ls/src/error.rs

@@ -33,6 +33,8 @@ pub enum Error {
     InvalidRepositoryId,
     #[error("invalid tokenizer path")]
     InvalidTokenizerPath,
+    #[error("llama.cpp error: {0}")]
+    LlamaCpp(crate::backend::APIError),
     #[error("ollama error: {0}")]
     Ollama(crate::backend::APIError),
     #[error("openai error: {0}")]
@@ -50,7 +52,7 @@ pub enum Error {
     #[error("tgi error: {0}")]
     Tgi(crate::backend::APIError),
     #[error("tree-sitter parse error: timeout possibly exceeded")]
-    TreeSitterParseError,
+    TreeSitterParsing,
     #[error("tree-sitter language error: {0}")]
     TreeSitterLanguage(#[from] tree_sitter::LanguageError),
     #[error("tokenizer error: {0}")]
@@ -60,7 +62,7 @@ pub enum Error {
     #[error("unknown backend: {0}")]
     UnknownBackend(String),
     #[error("unknown encoding kind: {0}")]
-    UnknownEncodingKind(String)
+    UnknownEncodingKind(String),
 }
 
 pub(crate) type Result<T> = std::result::Result<T, Error>;

crates/llm-ls/src/main.rs

@@ -417,6 +417,17 @@ async fn get_tokenizer(
 fn build_url(backend: Backend, model: &str) -> String {
     match backend {
         Backend::HuggingFace { url } => format!("{url}/models/{model}"),
+        Backend::LlamaCpp { mut url } => {
+            if url.ends_with("/completions") {
+                url
+            } else if url.ends_with('/') {
+                url.push_str("completions");
+                url
+            } else {
+                url.push_str("/completions");
+                url
+            }
+        }
         Backend::Ollama { url } => url,
         Backend::OpenAi { url } => url,
         Backend::Tgi { url } => url,
@@ -540,7 +551,8 @@ impl LanguageServer for LlmService {
                 general_capabilities
                     .position_encodings
                     .map(TryFrom::try_from)
-            }).unwrap_or(Ok(document::PositionEncodingKind::Utf16))?;
+            })
+            .unwrap_or(Ok(document::PositionEncodingKind::Utf16))?;
 
         *self.position_encoding.write().await = position_encoding;
 

crates/testbed/repositories-ci.yaml

@@ -2,10 +2,10 @@
 context_window: 2000
 fim:
   enabled: true
-  prefix: <fim_prefix>
-  middle: <fim_middle>
-  suffix: <fim_suffix>
-model: bigcode/starcoder
+  prefix: "<PRE> "
+  middle: " <MID>"
+  suffix: " <SUF>"
+model: codellama/CodeLlama-13b-hf
 backend: huggingface
 request_body:
   max_new_tokens: 150
@@ -14,8 +14,8 @@ request_body:
   top_p: 0.95
 tls_skip_verify_insecure: false
 tokenizer_config:
-  repository: bigcode/starcoder
-tokens_to_clear: ["<|endoftext|>"]
+  repository: codellama/CodeLlama-13b-hf
+tokens_to_clear: ["<EOT>"]
 repositories:
   - source:
       type: local