feat: implement SparseK attention core logic

Co-authored-by: Yael <[email protected]>
Co-authored-by: Gitty <[email protected]>

Author: 3 people

ggml/include/ggml.h

@@ -2236,9 +2236,9 @@ extern "C" {
 
     GGML_API struct ggml_tensor * ggml_sparsek_attn(
               struct ggml_context * ctx,
-              struct ggml_tensor  * Q,   
-              struct ggml_tensor  * K,  
-              struct ggml_tensor  * V,   
+              struct ggml_tensor  * Q,
+              struct ggml_tensor  * K,
+              struct ggml_tensor  * V,
               int32_t               k_top,
               int32_t               win_local,
               int32_t               stride_global);

ggml/src/ggml-cpu/ggml-cpu.c

@@ -1950,7 +1950,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
         case GGML_OP_SPARSEK_ATTN:
             {
                 ggml_compute_forward_sparsek_attn(params, tensor);
-            } break;          
+            } break;
         case GGML_OP_FLASH_ATTN_BACK:
             {
                 int32_t t = ggml_get_op_params_i32(tensor, 0);

tests/test-backend-ops.cpp

@@ -7497,7 +7497,6 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 // Test cases for performance evaluation: should be representative of real-world use cases
 static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     std::vector<std::unique_ptr<test_case>> test_cases;
-
     // Conv2d: K=CRS=NPQ=4096 matmul performance
     uint32_t                        iwh_idx  = 0;
     uint32_t                        kwh_idx  = 1;