fix error

Author: JamesLim-sy

paddle/fluid/operators/optimizers/lars_momentum_op.cu

@@ -279,13 +279,12 @@ __global__ void MomentumLarsKernel(
                       rescale_grad, gridDim.x, &param_norm, &grad_norm);
 #else
   const MT rescale_grad_pow = rescale_grad * rescale_grad;
-  MT param_parital_norm = threadIdx.x < thresh ? p_buffer[threadIdx.x] : 0;
-  MT grad_parital_norm = threadIdx.x < thresh ? g_buffer[threadIdx.x] : 0;
+  MT param_part_norm = threadIdx.x < thresh ? p_buffer[threadIdx.x] : 0;
+  MT grad_part_norm = threadIdx.x < thresh ? g_buffer[threadIdx.x] : 0;
   __syncthreads();
-  MT param_norm =
-      Sqrt(math::blockReduceSum<MT>(param_parital_norm, FINAL_MASK));
+  MT param_norm = Sqrt(math::blockReduceSum<MT>(param_part_norm, FINAL_MASK));
   MT grad_norm = Sqrt(rescale_grad_pow *
-                      math::blockReduceSum<MT>(grad_parital_norm, FINAL_MASK));
+                      math::blockReduceSum<MT>(grad_part_norm, FINAL_MASK));
 #endif
 
   const MT lr = learning_rate[0];
@@ -499,9 +498,9 @@ class LarsMomentumOpCUDAKernel : public framework::OpKernel<T> {
       MT* master_param_out_data = nullptr;
 
       if (multi_precision) {
-        auto master_param = ctx.MultiInput<framework::Tensor>("MasterParam");
+        auto master_param = ctx.MultiInput<framework::LoDTensor>("MasterParam");
         auto master_param_out =
-            ctx.MultiOutput<framework::Tensor>("MasterParamOut");
+            ctx.MultiOutput<framework::LoDTensor>("MasterParamOut");
         master_param_data = master_param[0]->data<MT>();
         master_param_out_data =
             master_param_out[0]->mutable_data<MT>(ctx.GetPlace());

paddle/fluid/operators/optimizers/lars_momentum_op.h

@@ -43,7 +43,7 @@ class LarsMomentumOpKernel : public framework::OpKernel<T> {
 
     T mu = static_cast<T>(ctx.Attr<float>("mu"));
     T lars_coeff = ctx.Attr<float>("lars_coeff");
-    T lars_weight_decay = (ctx.Attr<std::vector<float>>("lars_weight_decay"))[0];
+    T lars_weight_decay = ctx.Attr<std::vector<float>>("lars_weight_decay")[0];
     T epsilon = ctx.Attr<float>("epsilon");
 
     auto p_out = framework::EigenVector<T>::Flatten(*(param_out[0]));

python/paddle/fluid/tests/unittests/test_fleet_lars_meta_optimizer.py

@@ -51,7 +51,7 @@ def net(self, main_prog, startup_prog):
                 strategy.lars = True
                 strategy.lars_configs = {
                     "lars_coeff": 0.001,
-                    "lars_weight_decay": [0.0005],
+                    "lars_weight_decay": 0.0005,
                     "epsilon": 0,
                     "exclude_from_weight_decay": ["batch_norm", ".b"],
                 }
@@ -134,7 +134,7 @@ def test_lars_apply_with_amp(self):
         strategy.lars = True
         strategy.lars_configs = {
             "lars_coeff": 0.001,
-            "lars_weight_decay": [0.0005],
+            "lars_weight_decay": 0.0005,
             "epsilon": 0,
             "exclude_from_weight_decay": ["batch_norm", ".b"],
         }

python/paddle/fluid/tests/unittests/test_momentum_op.py

@@ -286,9 +286,9 @@ def setUp(self):
         grads = []
         velocitys = []
         learning_rates = []
-        master_params = []
         param_outs = []
         velocity_outs = []
+        master_params = []
         master_param_outs = []
         for i in range(self.params_num):
             master_param = np.random.random((123, 321)).astype("float32")
@@ -376,8 +376,8 @@ def setUp(self):
             gnorm = np.sqrt(np.square(grad).sum())
             local_lr = learning_rate * lars_coeff * pnorm / (
                 gnorm + lars_weight_decay[i] * param)
-            velocity_out = mu * velocity + local_lr * (grad + lars_weight_decay[i]
-                                                       * param)
+            velocity_out = mu * velocity + local_lr * (
+                grad + lars_weight_decay[i] * param)
             param_out = param - velocity_out
 
             params.append(("SubParam_" + str(i), param))