[MLU] add masterparam support for mlu adamw. (#46804)

paddle/fluid/operators/optimizers/adam_op_mlu.cc

@@ -291,11 +291,38 @@ class AdamWMLUKernel : public AdamMLUKernel<T> {
       skip_update = skip_update_vec[0];
     }
     bool with_decay = ctx.Attr<bool>("with_decay");
+    const bool multi_precision = ctx.Attr<bool>("multi_precision");
+    auto* param_out = ctx.Output<LoDTensor>("ParamOut");
+    auto* master_param_out = ctx.Output<LoDTensor>("MasterParamOut");
+    const auto* master_param = ctx.Input<LoDTensor>("MasterParam");
+
     VLOG(3) << "Skip update: " << skip_update << ", With decay: " << with_decay;
     if (!skip_update && with_decay) {
-      if (ctx.HasInput("MasterParam")) {
-        PADDLE_THROW(platform::errors::Unimplemented(
-            "Master Param is not supported on MLU"));
+      auto* param = ctx.Input<LoDTensor>("Param");
+      MLUCnnlTensorDesc param_desc(*param);
+      if (multi_precision) {
+        VLOG(3) << "[adamw] multi_precision, cast masterparam to param.";
+        bool has_master =
+            ctx.HasInput("MasterParam") && ctx.HasOutput("MasterParamOut");
+        PADDLE_ENFORCE_EQ(
+            has_master,
+            true,
+            platform::errors::InvalidArgument(
+                "The Input(MasterParam) and Output(MasterParamOut) "
+                "should not be null when "
+                "the attr `multi_precision` is true"));
+        // cast masterparam (fp32) to param (fp16), then paramout (fp16) to
+        // masterparamout (fp32)
+        MLUCnnlTensorDesc master_param_desc(*master_param);
+        cnnlCastDataType_t cast_type = GetCastDataType(
+            framework::TransToProtoVarType(master_param->dtype()),
+            framework::TransToProtoVarType(param->dtype()));
+        MLUCnnl::Cast(ctx,
+                      cast_type,
+                      master_param_desc.get(),
+                      GetBasePtr(master_param),
+                      param_desc.get(),
+                      const_cast<void*>(GetBasePtr(param)));
       } else {
         const auto* param_var = ctx.InputVar("Param");
         PADDLE_ENFORCE_EQ(param_var->IsType<phi::DenseTensor>(),
@@ -305,13 +332,12 @@ class AdamWMLUKernel : public AdamMLUKernel<T> {
                               "but the received is %s",
                               ctx.InputNames("Param").front(),
                               framework::ToTypeName(param_var->Type())));
-        auto* param = ctx.Input<LoDTensor>("Param");
+
         auto* lr = ctx.Input<LoDTensor>("LearningRate");
         float coeff = ctx.Attr<float>("coeff");
 
         // update param with decay coeff: mul(-1 * lr, coeff * param) + param
         MLUCnnlTensorDesc lr_desc(*lr);
-        MLUCnnlTensorDesc param_desc(*param);
         MLUCnnlOpTensorDesc mul_op_desc(
             CNNL_OP_TENSOR_MUL, ToCnnlDataType<T>(), CNNL_NOT_PROPAGATE_NAN);
 
@@ -330,6 +356,23 @@ class AdamWMLUKernel : public AdamMLUKernel<T> {
       }
     }
     AdamMLUKernel<T>::Compute(ctx);
+    if (multi_precision) {
+      VLOG(3) << "[adamw] multi_precision, cast paramout to masterparamout.";
+      // cast paramout to masterparamout
+      master_param_out->mutable_data<float>(ctx.GetPlace());
+      cnnlCastDataType_t cast_type = GetCastDataType(
+          framework::TransToProtoVarType(param_out->dtype()),
+          framework::TransToProtoVarType(master_param_out->dtype()));
+      MLUCnnlTensorDesc param_out_desc(*param_out);
+      MLUCnnlTensorDesc master_param_out_desc(*master_param_out);
+
+      MLUCnnl::Cast(ctx,
+                    cast_type,
+                    param_out_desc.get(),
+                    GetBasePtr(param_out),
+                    master_param_out_desc.get(),
+                    GetBasePtr(master_param_out));
+    }
   }
 };