Fix typos, test=document_fix (#53540)

paddle/phi/backends/device_base.cc

@@ -124,7 +124,7 @@ bool DeviceInterface::QueryEvent(size_t dev_id, const event::Event* event) {
   return true;
 }
 
-// memery manage
+// memory manage
 void DeviceInterface::MemoryCopyH2D(size_t dev_id,
                                     void* dst,
                                     const void* src,

paddle/phi/core/compat/get_kerneltype_forvar_utils.h

@@ -22,7 +22,7 @@ class KernelKey;
 class DenseTensor;
 /**
  * Note: GetKernelTypeForVarContext is currently designed for oneDNN kernel when
- * the related memeber function 'GetKernelTypeForVar' is special. It is
+ * the related member function 'GetKernelTypeForVar' is special. It is
  * possible to leverage to other vendor libraries in the future.
  */
 class GetKernelTypeForVarContext {
@@ -47,7 +47,7 @@ class GetKernelTypeForVarContext {
 
  private:
   const KernelKey* kernel_key_;  // not owned
-  // Use AttributeMap in namespace 'phi' to avoid depending 'fuild'
+  // Use AttributeMap in namespace 'phi' to avoid depending 'fluid'
   const AttributeMap* attrs_;  // not owned
   std::string* var_name_;      // not owned
   DenseTensor* tensor_;        // not owned

paddle/phi/core/compat/op_utils.h

@@ -31,7 +31,7 @@ const static std::string deprecated_kernel_name = "deprecated";  // NOLINT
 
 const std::unordered_set<std::string> standard_kernel_suffixs({
     "sr",  // SelectedRows kernel
-    "raw"  // fallback kernel of origfinal fluid op
+    "raw"  // fallback kernel of original fluid op
 });
 
 /**

paddle/phi/kernels/funcs/blas/blas.h

@@ -38,7 +38,7 @@ namespace funcs {
  * if Mat A is [BatchSize, H, W], Mat B is [BatchSize, H, W]. It will be a
  * `batch_size` times of GEMM. The batched GEMM could be faster base on the
  * implementation of the blas library. The batch size could be zero. If any
- * matrix of `matmul` has a batch size, the will be a batched GEMM, too. e.g.,
+ * matrix of `matmul` has a batch size, there will be a batched GEMM, too. e.g.,
  * Mat A is [BatchSize, H1, W2], and Mat B [H2, W2], The result matrix wil be
  * [BatchSize, H1, W2]
  *

paddle/phi/kernels/funcs/blas/blaslt_impl.cu.h

@@ -39,7 +39,7 @@ namespace funcs {
 // While kMatmul, kMatmulGrad, kMatmulGradWithoutBias share the same
 // enum value, but if all elements for MatmulPlanner->GetKey() is same,
 // no matter forward or backward, they could share the same descriptor
-// cache, in that the descritpor is for decription of matmul operation.
+// cache, in that the descriptor is for description of matmul operation.
 enum MatmulFusedType {
   kMatmul = CUBLASLT_EPILOGUE_DEFAULT,
   kMatmulGrad = CUBLASLT_EPILOGUE_DEFAULT,
@@ -216,7 +216,7 @@ struct MatmulDescriptor {
     cudaDataType_t scale_type = phi::backends::gpu::ToCudaDataType<MT>();
     cublasComputeType_t compute_type = GetCudaComputeType<T>();
 
-    // Create operation desciriptor; see cublasLtMatmulDescAttributes_t for
+    // Create operation descriptor; see cublasLtMatmulDescAttributes_t for
     // details about defaults; just need to set the transforms for A and B
     PADDLE_ENFORCE_GPU_SUCCESS(
         dynload::cublasLtMatmulDescCreate(&op_desc, compute_type, scale_type));
@@ -787,7 +787,7 @@ struct LinearGradWithCublasLt : public CublasLtBase<T> {
   }
 };
 #else
-// A void structure just for successfully complile.
+// A void structure just for successfully compile.
 struct MatmulPlanner {};
 #endif  // (PADDLE_WITH_CUDA) && CUDA_VERSION >= 11060