comp files

kernels_req_5/local_attention.cuh

@@ -1,9 +1,6 @@
 #ifndef __ATTENTION_CUH__
 #define __ATTENTION_CUH__
-
-// WINDOW = 129 because the CPU reference uses an inclusive range [q_t - 128, q_t],
-// which covers 129 positions (128 look-back steps + the current token itself).
-#define WINDOW 129
+#define WINDOW 128
 
 #include <cuda_runtime.h>
 #include "../utils/cuda_utils.cuh"

kernels_req_5/local_softmax.cuh

@@ -1,5 +1,6 @@
 #ifndef __SOFTMAX_KERNEL_CUH__
 #define __SOFTMAX_KERNEL_CUH__
+#define WINDOW 129
 
 #include <cuda_runtime.h>
 #include <math.h>