diff --git a/.vscode/launch.json b/.vscode/launch.json
index 2fe42f98..e0402de4 100644
--- a/.vscode/launch.json
+++ b/.vscode/launch.json
@@ -7,17 +7,18 @@
         {
             "type": "lldb",
             "request": "launch",
-            "name": "rllm-cpp orca",
+            "name": "rllm-cpp phi",
             "cwd": "rllm-cpp",
+            "preLaunchTask": "rllm-cpp: build",
             "program": "${workspaceFolder}/target/debug/rllm-cpp",
             "env": {
-                "RUST_LOG": "info,rllm=debug,aicirt=info"
+                "RUST_LOG": "info,tokenizers=error,rllm=trace,aicirt=info,llama_cpp_low=trace"
             },
             "args": [
                 "--verbose",
                 "--aicirt=${workspaceFolder}/target/release/aicirt",
-                "--model=https://huggingface.co/TheBloke/Orca-2-13B-GGUF/blob/main/orca-2-13b.Q8_0.gguf",
-                "--tokenizer=orca"
+                "--model=https://huggingface.co/TheBloke/phi-2-GGUF/blob/main/phi-2.Q8_0.gguf",
+                "--gpu-layers=100"
             ]
         }
     ]
diff --git a/.vscode/tasks.json b/.vscode/tasks.json
new file mode 100644
index 00000000..94332b78
--- /dev/null
+++ b/.vscode/tasks.json
@@ -0,0 +1,17 @@
+{
+	"version": "2.0.0",
+	"tasks": [
+		{
+			"type": "cargo",
+			"command": "build",
+			"options": {
+				"cwd": "${workspaceFolder}/rllm-cpp"
+			},
+			"problemMatcher": [
+				"$rustc"
+			],
+			"group": "build",
+			"label": "rllm-cpp: build"
+		}
+	]
+}
\ No newline at end of file
diff --git a/rllm-cpp/src/llamacpp/loader.rs b/rllm-cpp/src/llamacpp/loader.rs
index 9d6d9c86..7b367154 100644
--- a/rllm-cpp/src/llamacpp/loader.rs
+++ b/rllm-cpp/src/llamacpp/loader.rs
@@ -1,7 +1,7 @@
 use std::sync::Arc;
 
-use rllm::{config::ModelMeta, LoaderArgs, Repo, RllmEngine};
 use anyhow::{bail, Result};
+use rllm::{config::ModelMeta, LoaderArgs, Repo, RllmEngine};
 
 use llama_cpp_low as cpp;
 
@@ -44,16 +44,22 @@ fn do_load(args: &LoaderArgs, model_args: &mut CppLoaderArgs) -> Result<cpp::Mod
 
         let mut mparams = cpp::ModelParams::default();
         // TODO: make this configurable
-        mparams.set_split_mode(cpp::SplitMode::None);
+        mparams.set_split_mode(cpp::SplitMode::Layer);
         mparams.n_gpu_layers = model_args.n_gpu_layers.unwrap_or(0) as i32;
         log::info!("{} layer(s) offloaded to GPU", mparams.n_gpu_layers);
-        // don't GPU offload on Intel macs - it just fails there
-        #[cfg(all(target_os = "macos", target_arch = "x86_64"))]
+        #[cfg(target_os = "macos")]
         {
+            // don't GPU offload on Intel macs - it just fails there
+            #[cfg(target_arch = "x86_64")]
             if mparams.n_gpu_layers > 0 {
                 log::warn!("disabling GPU (Intel macOS)");
                 mparams.n_gpu_layers = 0;
             }
+            #[cfg(target_arch = "aarch64")]
+            {
+                log::info!("disabling mmap");
+                mparams.use_mmap = false;
+            }
         }
 
         let m = cpp::Model::from_file(file.to_str().unwrap(), mparams)?;