refactor: separate CPU kernels from shared operator extensions

src/core/ggml_extensions.h

@@ -6,8 +6,29 @@
 void sd_ggml_register_extensions(ggml_backend_t backend);
 
 #ifndef SD_USE_UPSTREAM_GGML
-const struct ggml_ext_op* sd_ggml_rope_op();
+#include "ggml-ext.h"
+
+ggml_ext_op_t sd_ggml_rope_op();
+bool sd_ggml_rope_describe(const ggml_tensor* node);
 ggml_tensor* sd_ggml_rope_apply(ggml_context* ctx, ggml_backend_t backend, ggml_tensor* x, ggml_tensor* pe, bool interleaved);
+
+struct SDI8Operation {
+    const ggml_tensor* weight;
+    const ggml_tensor* input;
+    const ggml_tensor* input_scale;
+    const ggml_tensor* weight_scale;
+    const ggml_tensor* bias;
+    int group_size;
+    bool packed;
+};
+
+ggml_ext_op_t sd_ggml_i8_op();
+ggml_ext_op_t sd_ggml_convrot_op();
+bool sd_ggml_i8_describe(const ggml_tensor* node, SDI8Operation* operation);
+bool sd_ggml_convrot_describe(const ggml_tensor* node, int* group_size);
+ggml_tensor* sd_ggml_mul_mat_i8(ggml_context* ctx, ggml_tensor* weight, ggml_tensor* input, ggml_tensor* input_scale, ggml_tensor* weight_scale = nullptr, ggml_tensor* bias = nullptr);
+ggml_tensor* sd_ggml_linear_i8(ggml_context* ctx, ggml_tensor* weight, ggml_tensor* input, ggml_tensor* weight_scale, ggml_tensor* bias, int group_size);
+ggml_tensor* sd_ggml_quantize_i8_convrot(ggml_context* ctx, ggml_tensor* input, int group_size);
 #endif
 
 #endif