Update ggml/src/ggml-cuda/cpy.cu

Co-authored-by: Johannes Gäßler <[email protected]>

Author: gcp

ggml/src/ggml-cuda/cpy.cu

@@ -98,6 +98,7 @@ template<dequantize_kernel_t dequant, int qk>
 static __device__ void cpy_blck_q_f32(const char * cxi, char * cdsti) {
     float * cdstf = (float *)(cdsti);
 
+#pragma unroll
     for (int j = 0; j < qk/2; j++) {
         float2 dq;
         dequant(cxi, 0, j, dq);