fix perplexity for batch size > 32

Author: JohannesGaessler

ggml-cuda.cu

@@ -4040,7 +4040,9 @@ static __global__ void cpy_f32_q8_0(
         src = &zero;
     } else {
         src = x;
-        memcpy(&dst[1 + iqs/8].qs[sizeof(float) * (iqs % 8)], src, sizeof(float));
+        if (i0 / QK8_0 == (i_blck_0 + ne00) / QK8_0) {
+            memcpy(&dst[1 + iqs/8].qs[sizeof(float) * (iqs % 8)], src, sizeof(float));
+        }
     }
 
     float val;