From 4b6387881577c3de8afa6dae61bfbd12ca107378 Mon Sep 17 00:00:00 2001
From: sebasbad <sebasbad@gmail.com>
Date: Wed, 29 Jul 2026 20:16:29 +0200
Subject: [PATCH] fix(engine): dispose model, context, and multimodal instance
 in worker shutdown to release C++ VRAM allocations

---
 lib/src/isolate/worker.dart | 18 +++++++++++++-----
 1 file changed, 13 insertions(+), 5 deletions(-)

diff --git a/lib/src/isolate/worker.dart b/lib/src/isolate/worker.dart
index ee736ff6..07a2d5f2 100644
--- a/lib/src/isolate/worker.dart
+++ b/lib/src/isolate/worker.dart
@@ -860,11 +860,19 @@ void _shutdown(
     } catch (_) {/* ignore */}
   }
   state.sessions.clear();
-  // We deliberately do NOT dispose the model or context here. With multiple
-  // engines in a process, freeing one worker's model/context can crash the
-  // other worker's outstanding operations because the backend is shared and
-  // some teardown paths touch process-global state. The OS reclaims memory
-  // on process exit; that is good enough for M3.
+  
+  if (state.multimodal != null) {
+    try {
+      state.multimodal!.dispose();
+    } catch (_) {}
+  }
+  try {
+    state.context.dispose();
+  } catch (_) {}
+  try {
+    state.model.dispose();
+  } catch (_) {}
+
   LlamaLibrary.dispose();
   reply.send(EngineShutdownComplete(requestId));
   commandRx.close();