simplify by swapping op params

ggml-ci

Author: CISC

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2852,10 +2852,8 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                         float scale = ggml_get_op_params_f32(node, 0);
 
                         i += 2; node = cgraph->nodes[i];
-                        float softcap = ggml_get_op_params_f32(node, 0);
 
-                        ggml_set_op_params_f32(node, 0, scale);
-                        ggml_set_op_params_f32(node, 1, softcap);
+                        ggml_set_op_params_f32(node, 1, scale);
                         node->src[0] = src0;
 
                         ggml_cuda_op_softcap(*cuda_ctx, node);

ggml/src/ggml-cuda/softcap.cu

@@ -26,8 +26,8 @@ void ggml_cuda_op_softcap(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     float scale;
     float softcap;
-    memcpy(&scale,   (float *) dst->op_params + 0, sizeof(float));
-    memcpy(&softcap, (float *) dst->op_params + 1, sizeof(float));
+    memcpy(&softcap, (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&scale,   (float *) dst->op_params + 1, sizeof(float));
 
     softcap_f32_cuda(src0_d, dst_d, scale, softcap, ggml_nelements(src0), stream);
 }