diff --git a/GPTQ-for-LLaMa/quant_cuda_kernel.cu b/GPTQ-for-LLaMa/quant_cuda_kernel.cu
index de0c0d6..ed9627a 100644
--- a/GPTQ-for-LLaMa/quant_cuda_kernel.cu
+++ b/GPTQ-for-LLaMa/quant_cuda_kernel.cu
@@ -631,7 +631,7 @@ __global__ void VecQuant4TransposeMatMulHalfKernel(
 #if __CUDA_ARCH__ < 700 && __CUDA_ARCH__ > 600
   atomicAddHalf(&mul2[n_cols * height * 8 + n_rows], res);
 #else
-  atomicAddHalf(&mul2[n_cols * height * 8 + n_rows], res);
+  atomicAdd(&mul2[n_cols * height * 8 + n_rows], res);
 #endif
 #endif
 }