diff --git a/ggml/src/ggml-cuda/allreduce.cu b/ggml/src/ggml-cuda/allreduce.cu
index 434689abd..d56129a22 100644
--- a/ggml/src/ggml-cuda/allreduce.cu
+++ b/ggml/src/ggml-cuda/allreduce.cu
@@ -184,13 +184,15 @@ static __global__ void ggml_cuda_ar_kernel(
             #pragma unroll
             for (int k = 0; k < ELEMS_PER_VEC; ++k) {
                 const T_wire d_low = ggml_cuda_cast<T_wire>(sendbuf[off + k]);
-                recvbuf[off + k] = ggml_cuda_cast<T_dst>(d_low) + ggml_cuda_cast<T_dst>(wire[k]);
+                recvbuf[off + k] = ggml_cuda_cast<T_dst>(
+                    ggml_cuda_cast<float>(d_low) + ggml_cuda_cast<float>(wire[k]));
             }
         }
         if (bid == 0 && tid < count - tail) {
             const T_wire d_low = ggml_cuda_cast<T_wire>(sendbuf[tail + tid]);
-            recvbuf[tail + tid] =
-                ggml_cuda_cast<T_dst>(d_low) + ggml_cuda_cast<T_dst>(host_other[tail + tid]);
+            recvbuf[tail + tid] = ggml_cuda_cast<T_dst>(
+                ggml_cuda_cast<float>(d_low) +
+                ggml_cuda_cast<float>(host_other[tail + tid]));
         }
     }
 }
@@ -210,7 +212,8 @@ static __global__ void ggml_cuda_ar_add_kernel(
     const int nt  = gridDim.x * blockDim.x;
     for (int i = tid; i < count; i += nt) {
         const T_src d_low = ggml_cuda_cast<T_src>(dst[i]);
-        dst[i] = ggml_cuda_cast<T_dst>(d_low) + ggml_cuda_cast<T_dst>(src[i]);
+        dst[i] = ggml_cuda_cast<T_dst>(
+            ggml_cuda_cast<float>(d_low) + ggml_cuda_cast<float>(src[i]));
     }
 }