Find out a bug. When set batch_size = -1 to use the autobatch.

reproduce:

utils/torch_utils.py

@@ -68,7 +68,7 @@ def select_device(device='', batch_size=None, newline=True):
     if cuda:
         devices = device.split(',') if device else '0'  # range(torch.cuda.device_count())  # i.e. 0,1,6,7
         n = len(devices)  # device count
-        if n > 1 and batch_size:  # check batch_size is divisible by device_count
+        if n > 1 and batch_size > 0:  # check batch_size is divisible by device_count
             assert batch_size % n == 0, f'batch-size {batch_size} not multiple of GPU count {n}'
         space = ' ' * (len(s) + 1)
         for i, d in enumerate(devices):