no repacking for avx2 for kcpp because it breaks 4_0_4_4 quants

Author: LostRuins

ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp

@@ -5642,11 +5642,11 @@ static const ggml::cpu::tensor_traits * ggml_aarch64_get_optimal_repack_type(con
             }
         }
     } else if (cur->type == GGML_TYPE_Q4_K) {
-        if (ggml_cpu_has_avx2()) {
-            if (cur->ne[1] % 8 == 0) {
-                return &ggml::cpu::aarch64::q4_K_8x8_q8_K;
-            }
-        }
+        // if (ggml_cpu_has_avx2()) { //we shall just use the regular avx2 handling, no repacking otherwise massive slowdown with gpu
+        //     if (cur->ne[1] % 8 == 0) {
+        //         return &ggml::cpu::aarch64::q4_K_8x8_q8_K;
+        //     }
+        // }
     } else if (cur->type == GGML_TYPE_IQ4_NL) {
         if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
             if (cur->ne[1] % 4 == 0) {

src/llama-model.cpp

@@ -286,7 +286,7 @@ static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & de
 
     // add extra buffer types, only if no GPU device is present
     // ref: https://github.com/ggml-org/llama.cpp/issues/12481#issuecomment-2743136094
-    if (!has_gpu_device) {
+    if (true) { //kcpp needs this to be true, otherwise 4_0_4_4 quants will break. avx2 repacking dont affect us cause we disabled it
         auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
         auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
         auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)