From 8ef020c1164dcc0d6a8e676b1e0a23c6be4477ed Mon Sep 17 00:00:00 2001
From: niuliling123 <51102941+niuliling123@users.noreply.github.com>
Date: Wed, 22 Mar 2023 14:28:52 +0800
Subject: [PATCH] Fix type error in adagrad_kernel (#51790)

---
 paddle/phi/kernels/gpu/adagrad_kernel.cu | 4 ++--
 python/paddle/fluid/optimizer.py         | 6 +++++-
 python/paddle/optimizer/adagrad.py       | 6 +++++-
 3 files changed, 12 insertions(+), 4 deletions(-)

diff --git a/paddle/phi/kernels/gpu/adagrad_kernel.cu b/paddle/phi/kernels/gpu/adagrad_kernel.cu
index 14046ed4491..1ce78b7e0f6 100644
--- a/paddle/phi/kernels/gpu/adagrad_kernel.cu
+++ b/paddle/phi/kernels/gpu/adagrad_kernel.cu
@@ -37,7 +37,7 @@ __global__ void AdagradGPUKernel(const T* param,
                                  MT* master_param_out,
                                  int num) {
   auto idx = blockDim.x * blockIdx.x + threadIdx.x;
-  MT lr_data = static_cast<T>(lr[0]);
+  MT lr_data = static_cast<MT>(lr[0]);
 
   for (int i = idx; i < num; i += blockDim.x * gridDim.x) {
     MT grad_data = static_cast<MT>(grad[i]);
@@ -47,7 +47,7 @@ __global__ void AdagradGPUKernel(const T* param,
     MT param_out_data =
         in - (lr_data * grad_data) / (sqrt(moment_out_data) + epsilon);
 
-    param_out[i] = static_cast<MT>(param_out_data);
+    param_out[i] = static_cast<T>(param_out_data);
 
     if (master_param_out) {
       master_param_out[i] = param_out_data;
diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py
index d835a3fbfcf..ded8883ffdb 100755
--- a/python/paddle/fluid/optimizer.py
+++ b/python/paddle/fluid/optimizer.py
@@ -2081,7 +2081,11 @@ class AdagradOptimizer(Optimizer):
         for p in parameters:
             if self._multi_precision and self._is_dtype_fp16_or_bf16(p.dtype):
                 master_p = self._create_master_weight(p)
-                self._add_accumulator(self._moment_acc_str, master_p)
+                self._add_accumulator(
+                    self._moment_acc_str,
+                    master_p,
+                    fill_value=self.initial_accumulator_value,
+                )
                 continue
             if (
                 self._is_dtype_fp16_or_bf16(p.dtype)
diff --git a/python/paddle/optimizer/adagrad.py b/python/paddle/optimizer/adagrad.py
index 3d2935c7407..c19b3116de3 100644
--- a/python/paddle/optimizer/adagrad.py
+++ b/python/paddle/optimizer/adagrad.py
@@ -146,7 +146,11 @@ class Adagrad(Optimizer):
                 continue
             if self._multi_precision and self._is_dtype_fp16_or_bf16(p.dtype):
                 master_p = self._create_master_weight(p)
-                self._add_accumulator(self._moment_acc_str, master_p)
+                self._add_accumulator(
+                    self._moment_acc_str,
+                    master_p,
+                    fill_value=self.initial_accumulator_value,
+                )
                 self._already_create_accumulater.add(p.name)
                 continue
             if (
-- 
GitLab