diff --git a/.gitmodules b/.gitmodules
index e150ac479..6f96885ce 100644
--- a/.gitmodules
+++ b/.gitmodules
@@ -1,3 +1,3 @@
 [submodule "crates/llama-cpp-bindings/llama.cpp"]
 	path = crates/llama-cpp-bindings/llama.cpp
-	url = https://github.com/TabbyML/llama.cpp
+	url = https://github.com/ggerganov/llama.cpp
diff --git a/crates/llama-cpp-bindings/build.rs b/crates/llama-cpp-bindings/build.rs
index bc00a56b1..bd6a63128 100644
--- a/crates/llama-cpp-bindings/build.rs
+++ b/crates/llama-cpp-bindings/build.rs
@@ -23,9 +23,13 @@ fn build_llama_cpp() {
     let mut config = Config::new("llama.cpp");
     config.define("LLAMA_NATIVE", "OFF");
     config.define("INS_ENB", "ON");
+    if cfg!(not(debug_assertions)) {
+        config.define("CMAKE_BUILD_TYPE", "Release");
+    }
 
     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");
diff --git a/crates/llama-cpp-bindings/llama.cpp b/crates/llama-cpp-bindings/llama.cpp
index 807ee66db..c29af7e22 160000
--- a/crates/llama-cpp-bindings/llama.cpp
+++ b/crates/llama-cpp-bindings/llama.cpp
@@ -1 +1 @@
-Subproject commit 807ee66dbfb9e395362388a7cfb12684a41d1d09
+Subproject commit c29af7e2252d288f2ea58a7d437c1cb7c0abf160