Fix KleidiAI compilation with improved fallback and error handling

- Revert to GitHub tarball download for stability
- Add debug logging for KleidiAI kernel fallback scenarios
- Improve error messages when no suitable kernels available
- Keep ARM64 architecture requirement for KleidiAI enabling
- Ensure graceful fallback to standard CPU implementation

Author: baonudesifeizhai

ggml/src/ggml-cpu/CMakeLists.txt

@@ -495,17 +495,17 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         # Fetch KleidiAI sources:
         include(FetchContent)
         set(KLEIDIAI_COMMIT_TAG "v1.9.0")
-        set(KLEIDIAI_DOWNLOAD_URL "https://git.gitlab.arm.com/kleidi/kleidiai/-/archive/${KLEIDIAI_COMMIT_TAG}/kleidiai-${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "e4c9fcb5de397ba3532d593672d56e95")
+        set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
+        set(KLEIDIAI_ARCHIVE_MD5  "2a8e1bb55d201557553545536489a017")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)
         endif()
 
         FetchContent_Declare(KleidiAI_Download
-            GIT_REPOSITORY https://git.gitlab.arm.com/kleidi/kleidiai.git
-            GIT_TAG ${KLEIDIAI_COMMIT_TAG}
-            GIT_SHALLOW TRUE)
+            URL ${KLEIDIAI_DOWNLOAD_URL}
+            DOWNLOAD_EXTRACT_TIMESTAMP NEW
+            URL_HASH MD5=${KLEIDIAI_ARCHIVE_MD5})
 
         FetchContent_MakeAvailable(KleidiAI_Download)
         FetchContent_GetProperties(KleidiAI_Download

ggml/src/ggml-cpu/kleidiai/kleidiai.cpp

@@ -104,7 +104,9 @@ class tensor_traits : public ggml::cpu::tensor_traits {
     bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
         ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, op);
         if (!kernels) {
-            return false;  // No suitable kernel available
+            // No suitable KleidiAI kernel available, fallback to standard CPU implementation
+            GGML_LOG_DEBUG("%s: No suitable KleidiAI kernel available for operation, falling back to standard CPU implementation\n", __func__);
+            return false;  // Let the system fallback to standard CPU implementation
         }
         kernel_info * kernel = op->src[1]->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
 
@@ -151,7 +153,9 @@ class tensor_traits : public ggml::cpu::tensor_traits {
 
         ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
         if (!kernels) {
-            return false;  // No suitable kernel available
+            // No suitable KleidiAI kernel available, fallback to standard CPU implementation
+            GGML_LOG_DEBUG("%s: No suitable KleidiAI kernel available for KV cache operation, falling back to standard CPU implementation\n", __func__);
+            return false;  // Let the system fallback to standard CPU implementation
         }
 
         kernel_info * kernel = src1->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
@@ -281,7 +285,9 @@ class tensor_traits : public ggml::cpu::tensor_traits {
 
         ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
         if (!kernels) {
-            return false;  // No suitable kernel available
+            // No suitable KleidiAI kernel available, fallback to standard CPU implementation
+            GGML_LOG_DEBUG("%s: No suitable KleidiAI kernel available for Q4_0 operation, falling back to standard CPU implementation\n", __func__);
+            return false;  // Let the system fallback to standard CPU implementation
         }
 
         kernel_info * kernel = src1->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;