Tighten YuE2 VRAM use on the 5080 and empty GPU before launch.

Co-authored-by: Cursor <cursoragent@cursor.com>

tests/test_yue2_worker.py

@@ -1,5 +1,6 @@
 """CPU-only YuE2 worker regression tests; never import torch or run inference."""
 import importlib.util
+import os
 from pathlib import Path
 import unittest
 from types import SimpleNamespace
@@ -26,13 +27,14 @@ class WorkerTests(unittest.TestCase):
             backends=SimpleNamespace(cudnn=SimpleNamespace(is_available=lambda: True)),
         )
         attention, pipeline = worker.resolve_attention_backend(torch_mod)
-        self.assertEqual(attention, 'cudnn')
-        self.assertEqual(pipeline, 'torch')
-        self.assertNotEqual(attention, 'flash')
-        torch_mod.backends.cudnn.is_available = lambda: False
-        attention, pipeline = worker.resolve_attention_backend(torch_mod)
         self.assertEqual(attention, 'sdpa')
         self.assertEqual(pipeline, 'torch-eager')
+        self.assertNotEqual(attention, 'flash')
+        self.assertTrue(os.environ.get('PYTORCH_CUDA_ALLOC_CONF', '').startswith('expandable_segments'))
+
+    def test_cuda_oom_detection(self):
+        self.assertTrue(worker.is_cuda_oom(RuntimeError('CUDA out of memory. Tried to allocate 2.49 GiB')))
+        self.assertFalse(worker.is_cuda_oom(RuntimeError('bad lyrics')))
 
 
 if __name__ == '__main__':