support cpm on xpu

training/kunlunxin/cpm-pytorch/config/config_R300x1x8.py

@@ -14,7 +14,7 @@
 max_steps = 10000
 
 warmup = 0.2
-learning_rate = 0.000125
+learning_rate = 0.0005
 
 beta_1: float = 0.9
 beta_2: float = 0.99

training/kunlunxin/cpm-pytorch/extern/trainer_adapter.py

@@ -19,23 +19,6 @@ def create_optimizer(config, model):
 
     return optimizer
 
-class Wrap_Module(nn.Module):
-
-    def __init__(self, module):
-        super(Wrap_Module, self).__init__()
-        self.add_module('module', module)
-
-    def forward(self, *inputs, **kwargs):
-        return self.module(*inputs, **kwargs)
-
-    def state_dict(self, destination=None, prefix='', keep_vars=False):
-        return self.module.state_dict(destination, prefix, keep_vars)
-
-    def load_state_dict(self, state_dict, strict=True):
-        self.module.load_state_dict(state_dict, strict=strict)
-
 def model_to_fp16(config, model: nn.Module,
                   optimizer: Optimizer) -> Tuple[nn.Module, Optimizer]:
-    # we don't support fp16 now, but we should follow the FP16_Module behavior
-    model = Wrap_Module(model)
-    return model, optimizer
+    return model, optimizer