【PIR Dist Op Reg No.18】 reg lars_momentum (#60838)

* fix

* fix

* fix

* fix

* fix

* add f

* fix

* change sequence

* add lars_momentum_

* fix

* fix

Author: enkilee

paddle/fluid/operators/optimizers/lars_momentum_op.cc

@@ -71,14 +71,14 @@ class LarsMomentumOpMaker : public framework::OpProtoAndCheckerMaker {
         .AsDispensable();
     AddAttr<float>("mu", "(float) Momentum coefficient");
     AddAttr<float>("lars_coeff", "(float, default 0.001) LARS coefficient.")
-        .SetDefault(0.001);
+        .SetDefault(0.001f);
     AddAttr<std::vector<float>>(
         "lars_weight_decay",
         "(std::vector<float>, default 0.0005) LARS weight decay params")
-        .SetDefault({0.0005});
+        .SetDefault({0.0005f});
     AddAttr<float>("epsilon",
                    "(float, default 0.0) epsilon to avoid Division by Zero.")
-        .SetDefault(0.0);
+        .SetDefault(0.0f);
     AddAttr<bool>("multi_precision",
                   "(bool, default false) "
                   "Whether to use multi-precision during weight updating.")

paddle/fluid/pir/dialect/op_generator/ops_api_gen.py

@@ -137,6 +137,7 @@
     'fused_dot_product_attention',
     'nce',
     'lars_momentum',
+    'lars_momentum_',
     'max_pool2d_v2',
     'recv_v2',
     'rnn_',

paddle/fluid/pir/dialect/operator/ir/ops.yaml

@@ -1583,16 +1583,17 @@
   backward: fused_feedforward_grad
 
 - op: lars_momentum
-  args: (Tensor param,  Tensor velocity, Tensor grad, Tensor learning_rate, Tensor master_param, float mu, float lars_coeff=0.001f, float[] lars_weight_decay={0.0005}, float epsilon=0, bool multi_precision=false, float rescale_grad=1.0f)
-  output: Tensor(param_out), Tensor(velocity_out), Tensor(master_param_out)
+  args: (Tensor[] param, Tensor[] grad, Tensor[] velocity, Tensor[] learning_rate, Tensor[] master_param, float mu, float lars_coeff=0.001f, float[] lars_weight_decay={0.0005f}, float epsilon=0.0f, bool multi_precision=false, float rescale_grad=1.0f)
+  output: Tensor[](param_out){param.size()}, Tensor[](velocity_out){param.size()}, Tensor[](master_param_out){param.size()}
   infer_meta:
-    func: SparseMomentumInferMeta
-    param: [param, learning_rate, velocity]
+    func: LarsMomentumInferMeta
+    param: [param, velocity, learning_rate, grad, master_param, lars_weight_decay, mu, lars_coeff, epsilon, multi_precision, rescale_grad]
   kernel:
     func: lars_momentum
     param: [param, velocity, learning_rate, grad, master_param, lars_weight_decay, mu, lars_coeff, epsilon, multi_precision, rescale_grad]
     data_type: param
   optional: master_param, master_param_out
+  inplace : master_param -> master_param_out
 
 - op: match_matrix_tensor
   args: (Tensor x, Tensor y, Tensor w, int dim_t=1)