Fix: move SparseK op_params indices to 7..10 to avoid overlap with precision params

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

Author: Gitty Burstein

ggml/include/ggml.h

@@ -219,7 +219,7 @@
 #define GGML_MAX_PARAMS         2048
 #define GGML_MAX_SRC            10
 #define GGML_MAX_N_THREADS      512
-#define GGML_MAX_OP_PARAMS      128
+#define GGML_MAX_OP_PARAMS      64
 
 #ifndef GGML_MAX_NAME
 #   define GGML_MAX_NAME        64

ggml/src/ggml-cpu/ops.cpp

@@ -8004,10 +8004,10 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
     // -------- SparseK op_params --------
-    const bool    use_sparsek   = ggml_get_op_params_i32(dst, 3) != 0;
-    const int32_t k_top         = ggml_get_op_params_i32(dst, 4);
-    const int32_t win_local     = ggml_get_op_params_i32(dst, 5);
-    const int32_t stride_glb    = ggml_get_op_params_i32(dst, 6);
+    const bool    use_sparsek   = ggml_get_op_params_i32(dst, 7) != 0;
+    const int32_t k_top         = ggml_get_op_params_i32(dst, 8);
+    const int32_t win_local     = ggml_get_op_params_i32(dst, 9);
+    const int32_t stride_glb    = ggml_get_op_params_i32(dst, 10);
     // ----------------------------------------------------------------------------
 
     ggml_type         const k_vec_dot_type = ggml_get_type_traits_cpu(k->type)->vec_dot_type;

ggml/src/ggml.c

@@ -5214,10 +5214,10 @@ struct ggml_tensor * ggml_flash_attn_back(
     return result;
 }
 
-#define GGML_FA_EXT_PARAM_SPARSEK_FLAG   3
-#define GGML_FA_EXT_PARAM_SPARSEK_KTOP   4
-#define GGML_FA_EXT_PARAM_SPARSEK_WIN    5
-#define GGML_FA_EXT_PARAM_SPARSEK_STRIDE 6
+#define GGML_FA_EXT_PARAM_SPARSEK_FLAG   7
+#define GGML_FA_EXT_PARAM_SPARSEK_KTOP   8
+#define GGML_FA_EXT_PARAM_SPARSEK_WIN    9
+#define GGML_FA_EXT_PARAM_SPARSEK_STRIDE 10
 
 void ggml_flash_attn_ext_set_sparsek(struct ggml_tensor * a,
                                      bool use_sparsek,