Avoid unnecessary CPU offload and expose full GPU profile mode

test_inference.py

@@ -46,6 +46,11 @@ class WorkerTests(unittest.TestCase):
   with self.patches(),patch('inference.probe',return_value=self.gpus),patch('inference.cgroup_headroom',return_value=32*1024**3),patch.object(self.worker,'_fit_command',return_value='CUDA0 50000 1000 1000\nCUDA1 2000 1000 1000\nHost 512 0 256\n') as fit,patch('inference.subprocess.Popen') as launch:
    with self.assertRaises(InferenceError):self.worker.ensure(self.profile)
    launch.assert_not_called();self.assertLessEqual(fit.call_count,11)
+ def test_full_gpu_refuses_cpu_fallback(self):
+  self.profile['parameters']['gpu_offload']='full'
+  with self.patches(),patch('inference.probe',return_value=self.gpus),patch('inference.cgroup_headroom',return_value=32*1024**3),patch.object(self.worker,'_fit_command',return_value='CUDA0 50000 1000 1000\nCUDA1 2000 1000 1000\nHost 512 0 256\n') as fit,patch('inference.subprocess.Popen') as launch:
+   with self.assertRaisesRegex(InferenceError,'Vollständige GPU'):self.worker.ensure(self.profile)
+   launch.assert_not_called();self.assertEqual(fit.call_count,1)
  def test_build_symlink_cannot_escape_deck_state(self):
   self.runtime.status=lambda:dict(active='outside',builds=[dict(id='outside',backend='CUDA',fit_tool=True)])
   (self.root/'runtime'/'outside').symlink_to('/usr')