llama : add SKIP_KQ_KQV option

ggml.c

@@ -16602,6 +16602,10 @@ static void ggml_compute_forward_cross_entropy_loss_back(
 static void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
     GGML_ASSERT(params);
 
+    if (tensor->op == GGML_OP_NONE) {
+        return;
+    }
+
 #ifdef GGML_USE_CUBLAS
     bool skip_cpu = ggml_cuda_compute_forward(params, tensor);
     if (skip_cpu) {

llama.cpp

@@ -5815,6 +5815,33 @@ static struct ggml_cgraph * llama_build_graph(
             GGML_ASSERT(false);
     }
 
+#if 1
+    for (int i = 0; i < result->n_nodes; ++i) {
+        struct ggml_tensor * node = result->nodes[i];
+        if (getenv("SKIP_KQ_ALL")) {
+            if (
+                    strcmp(node->name, "KQ")  == 0 ||
+                    strcmp(node->name, "KQ_scaled") == 0 ||
+                    strcmp(node->name, "KQ_masked") == 0 ||
+                    strcmp(node->name, "KQ_soft_max") == 0 ||
+                    strcmp(node->name, "KQV") == 0 ||
+                    false) {
+                //printf("skipping %s\n", dst->name);
+                node->op  = GGML_OP_NONE;
+            }
+        }
+        if (getenv("SKIP_KQ_KQV")) {
+            if (
+                    strcmp(node->name, "KQ")  == 0 ||
+                    strcmp(node->name, "KQV") == 0 ||
+                    false) {
+                //printf("skipping %s\n", dst->name);
+                node->op  = GGML_OP_NONE;
+            }
+        }
+    }
+#endif
+
     return result;
 }