[DCU] Fix NAN problem when training BERT on DUC platform (#44643)

paddle/fluid/operators/optimizers/distributed_fused_lamb_op.cu

@@ -166,7 +166,11 @@ static void MultiTensorL2Norm(const platform::CUDAPlace &place,
 
   constexpr int kNumTensor = MaxTensorNumPerLaunch;
   constexpr int kNumChunk = MaxChunkNumPerLaunch;
+#ifdef PADDLE_WITH_HIP
+  constexpr int kBlockDim = 256;
+#else
   constexpr int kBlockDim = 512;
+#endif
 
   int max_chunk_num = -1;
   int vec_size = 8;
@@ -805,7 +809,11 @@ static void MultiTensorUpdateLambParamAndBetaPows(
         platform::errors::InvalidArgument("Beta2Pow should be nullptr."));
   }
 
+#ifdef PADDLE_WITH_HIP
+  const int block_dim = 256;
+#else
   const int block_dim = 512;
+#endif
 
   int vec_size = 8;
   for (int i = 0; i < n; ++i) {

paddle/fluid/platform/device/gpu/rocm/rocm_device_function.h

@@ -134,7 +134,11 @@ __device__ T reduceSum(T val, int tid, int len) {
   // I use Warp-Level Parallelism and assume the Warp size
   // is 32 which may be different for different GPU,
   // but most card's warp size is 32.
+#ifdef PADDLE_WITH_HIP
+  const int warpSize = 64;
+#else
   const int warpSize = 32;
+#endif
   __shared__ T shm[warpSize];
   unsigned mask = 0u;
   CREATE_SHFL_MASK(mask, tid < len);