Revert embedding normal path and fix amp list

oneflow/core/job_rewriter/auto_mixed_precision_lists.cpp

@@ -90,6 +90,7 @@ const AMPList& AutoMixedPrecisionLists::ClearList() {
                                "identity",
                                "flatten",
                                "squeeze",
+                               "embedding",
                                "expand_dims",
                                "cast_to_static_shape",
                                "parallel_cast",

oneflow/user/kernels/gather_kernel_util.cpp

@@ -85,7 +85,7 @@ void GatherKernelUtilImpl<DeviceType::kCPU, T, K>::Forward(ep::Stream* stream, c
         const T* from = in + outer_idx * gather_dim_size * inner_dim_size + idx * inner_dim_size;
         std::copy(from, from + inner_dim_size, to);
       } else {
-        std::memset(reinterpret_cast<void*>(to), 0, inner_dim_size * sizeof(K));
+        std::memset(reinterpret_cast<void*>(to), 0, inner_dim_size * sizeof(T));
       }
     }
   }

python/oneflow/nn/modules/sparse.py

@@ -160,9 +160,12 @@ def forward(self, indices):
                 flow._C.embedding_renorm_(
                     self.weight, indices, self.max_norm, self.norm_type
                 )
-        return flow._C.embedding(
-            self.weight, indices, self.padding_idx, self.scale_grad_by_freq
-        )
+        if self.padding_idx is None and not self.scale_grad_by_freq:
+            return flow._C.gather(self.weight, indices, axis=0)
+        else:
+            return flow._C.embedding(
+                self.weight, indices, self.padding_idx, self.scale_grad_by_freq
+            )
 
 
 def embedding(
@@ -232,7 +235,10 @@ def embedding(
         with flow.no_grad():
             weight = flow._C.embedding_renorm_(weight, input, max_norm, norm_type)
 
-    return flow._C.embedding(weight, input, padding_idx, scale_grad_by_freq)
+    if padding_idx is None and not scale_grad_by_freq:
+        return flow._C.gather(weight, input, axis=0)
+    else:
+        return flow._C.embedding(weight, input, padding_idx, scale_grad_by_freq)
 
 
 if __name__ == "__main__":