Add freeContext

Author: crasm

lib/src/ensemble_llama_base.dart

@@ -22,7 +22,7 @@ void main() async {
   print(model);
 
   final ctx = await llama.newContext(model, params);
-  // await llama.freeContext(ctx);
+  await llama.freeContext(ctx);
 
   await llama.freeModel(model);
   llama.dispose();
@@ -104,9 +104,9 @@ class Llama {
     return resp.ctx!;
   }
 
-  // Future<void> freeContext(Context ctx) async {
-  //   final ctl = FreeContextCtl(ctx);
-  //   _controlPort.send(ctl);
-  //   await _response.firstWhere((e) => e is FreeContextResp && e.id == ctl.id);
-  // }
+  Future<void> freeContext(Context ctx) async {
+    final ctl = FreeContextCtl(ctx);
+    _controlPort.send(ctl);
+    await _response.firstWhere((e) => e is FreeContextResp && e.id == ctl.id);
+  }
 }

lib/src/llama_cpp_isolate_wrapper.dart

@@ -201,8 +201,10 @@ class Model {
 }
 
 class Context {
-  final int rawPointer;
-  const Context._(this.rawPointer);
+  final int _rawPointer;
+  const Context._(this._rawPointer);
+  Pointer<llama_context> get _ffiPointer =>
+      Pointer.fromAddress(_rawPointer).cast<llama_context>();
 }
 
 class _Allocations<E> {
@@ -313,5 +315,7 @@ void _onControl(ControlMessage ctl) {
       _response.send(ctl.done(Context._(rawCtx)));
 
     case FreeContextCtl():
+      libllama.llama_free(ctl.ctx._ffiPointer);
+      _response.send(ctl.done());
   }
 }