fix: resolve result_future on inference error instead of tearing down the model

src/server/worker_registry.py

@@ -97,13 +97,13 @@ def _commit_completed_packet(
 ) -> bool:
     """Complete the request future. Return True if the worker should exit."""
     if completed.error is not None:
-        logger.error(
-            f"[{model_name}] Inference failed, triggering model unload..."
-        )
+        # Resolve the caller and keep the worker/model alive; unloading here runs the
+        # pipeline destructor, which can touch a corrupted device context after certain
+        # errors and SIGABRT the whole process, taking every other loaded model with it.
+        logger.error(f"[{model_name}] Inference failed: {completed.error}")
         if packet.result_future is not None and not packet.result_future.done():
             packet.result_future.set_exception(completed.error)
-        asyncio.create_task(registry.register_unload(model_name))
-        return True
+        return False
     if packet.result_future is not None and not packet.result_future.done():
         packet.result_future.set_result(completed)
     return False

tests/unit/test_worker_registry_unit.py

@@ -310,7 +310,7 @@ async def transcribe(self, gen_config):
     assert packet.segments is None
 
 
-def test_commit_treats_error_field_not_error_prefix() -> None:
+def test_commit_resolves_error_without_unloading_the_model() -> None:
     class DummyRegistry:
         def __init__(self):
             self.unloaded = []
@@ -342,11 +342,14 @@ async def _run():
             result_future=err_future,
             error=RuntimeError("gpu oom"),
         )
-        assert worker_module._commit_completed_packet(err, err, "m", registry) is True
+        # On error the worker resolves the caller's future and keeps the model loaded;
+        # unloading here runs the pipeline destructor, which can SIGABRT the process if
+        # the device context is corrupted (see queue_worker_llm docstring).
+        assert worker_module._commit_completed_packet(err, err, "m", registry) is False
         with pytest.raises(RuntimeError, match="gpu oom"):
             err_future.result()
         await asyncio.sleep(0)
-        assert registry.unloaded == ["m"]
+        assert registry.unloaded == []
 
     asyncio.run(_run())