Separate vision projector device from LLM GPU split

test_vision.py

@@ -17,8 +17,10 @@ class CapabilitiesTests(unittest.TestCase):
   for value in (-1,32769,True,1.5):
    with self.assertRaises(ValueError):chat_parameters({'gpu_devices':['GPU-a'],'gpu_reserve_mode':'manual','gpu_reserve_mib':{'GPU-a':value}})
   with self.assertRaises(ValueError):chat_parameters({'gpu_reserve_mode':'manual'})
- def test_projector_device_requires_selected_gpu(self):
+ def test_projector_device_validates_uuid_independently(self):
   with self.assertRaises(ValueError):chat_parameters({'vision_device':'GPU-missing'})
+  uuid='GPU-4834d9d7-5b61-3004-1fb3-4ae49d482d4b'
+  self.assertEqual(chat_parameters({'vision_device':uuid})['vision_device'],uuid)
 
 class ProjectorLaunchTests(WorkerTests):
  def test_cpu_and_gpu_projector_flags_and_reserve(self):
@@ -32,4 +34,16 @@ class ProjectorLaunchTests(WorkerTests):
    if device=='cpu':self.assertIn('--no-mmproj-offload',args)
    else:self.assertEqual(args[args.index('--mmproj-device')+1],'CUDA1');self.assertEqual(self.worker.memory_plan['gpus'][1]['required_mib'],7072)
    self.worker.process=None
+ def test_projector_only_gpu_excluded_from_model_devices(self):
+  self.profile['parameters'].update(gpu_devices=['gpu-first'],split_mode='none',tensor_split=[],vision_projector='p',vision_device='gpu-second',gpu_offload='full')
+  self.worker.catalog.entry=lambda ident:dict(id='p',file='mmproj.gguf',size=1024**3,role='vision_projector') if ident=='p' else self.entry
+  process=Mock();process.poll.return_value=None;http=Mock();http.getresponse.return_value.status=200
+  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 9000 1000 1000\nHost 512 0 256\n'),patch('inference.subprocess.Popen',return_value=process) as launch,patch('inference.http.client.HTTPConnection',return_value=http),patch('inference.threading.Thread'):
+   self.worker.ensure(self.profile)
+  args=launch.call_args.args[0]
+  self.assertEqual(args[args.index('--device')+1],'CUDA0')
+  self.assertEqual(args[args.index('--mmproj-device')+1],'CUDA1')
+  self.assertEqual(self.worker.memory_plan['gpus'][1]['required_mib'],3072)
+  self.assertEqual(launch.call_args.kwargs['env']['CUDA_VISIBLE_DEVICES'],'gpu-first,gpu-second')
+  self.worker.process=None
 if __name__=='__main__':unittest.main()