SparseK: fix op_params indices + add runtime guards

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

Author: 3 people

ggml/src/ggml-cpu/ops.cpp

@@ -8003,10 +8003,10 @@ 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 (לא משנה שום דבר חוץ מקריאת הפרמטרים) --------
-    const bool    use_sparsek   = ggml_get_op_params_i32(dst, 30) != 0;
-    const int32_t k_top         = ggml_get_op_params_i32(dst, 31);
-    const int32_t win_local     = ggml_get_op_params_i32(dst, 32);
-    const int32_t stride_glb    = ggml_get_op_params_i32(dst, 33);
+    const bool    use_sparsek   = ggml_get_op_params_i32(dst, 28) != 0;
+    const int32_t k_top         = ggml_get_op_params_i32(dst, 29);
+    const int32_t win_local     = ggml_get_op_params_i32(dst, 30);
+    const int32_t stride_glb    = ggml_get_op_params_i32(dst, 31);
     // ----------------------------------------------------------------------------
 
     ggml_type         const k_vec_dot_type = ggml_get_type_traits_cpu(k->type)->vec_dot_type;

ggml/src/ggml.c

@@ -5215,10 +5215,10 @@ struct ggml_tensor * ggml_flash_attn_back(
     return result;
 }
 
-#define GGML_FA_EXT_PARAM_SPARSEK_FLAG   30
-#define GGML_FA_EXT_PARAM_SPARSEK_KTOP   31
-#define GGML_FA_EXT_PARAM_SPARSEK_WIN    32
-#define GGML_FA_EXT_PARAM_SPARSEK_STRIDE 33
+#define GGML_FA_EXT_PARAM_SPARSEK_FLAG   28
+#define GGML_FA_EXT_PARAM_SPARSEK_KTOP   29
+#define GGML_FA_EXT_PARAM_SPARSEK_WIN    30
+#define GGML_FA_EXT_PARAM_SPARSEK_STRIDE 31
 
 void ggml_flash_attn_ext_set_sparsek(struct ggml_tensor * a,
                                      bool use_sparsek,