ggml : fix unaligned access in AMX code

Author: ggerganov

ggml/src/ggml-cpu/amx/mmq.cpp

@@ -1432,7 +1432,10 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q4_0, float, BLOCK_M, BLOCK_N, BLO
         auto compute = [&](auto col, auto i) {
             // load a and compute compensation
             if constexpr (col == 0) {
-                const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
+                //const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
+                int32_t a_ptr[8];
+                memcpy((char *)a_ptr, (const char *)A[0 * KB + i].qs, 32);
+
                 vcomp = _mm512_setzero_si512();
                 for (int k = 0; k < 8; ++k) {
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
@@ -1495,7 +1498,10 @@ struct tinygemm_kernel_vnni<block_q8_1, block_q4_1, float, 1, BLOCK_N, BLOCK_K>
         auto compute = [&](auto col, auto i) {
             // load a
             if constexpr (col == 0) {
-                const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
+                //const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
+                int32_t a_ptr[8];
+                memcpy((char *)a_ptr, (const char *)A[0 * KB + i].qs, 32);
+
                 for (int k = 0; k < 8; ++k) {
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
                 }