add cpu_read_ram_bw, metal_read_vram_bw, cuda_read_vram_bw

2024-11-29 19:03:01 +04:00
parent 0a6ffe68e0
commit 68ecabc8c3
5 changed files with 139 additions and 48 deletions
@@ -320,6 +320,7 @@ extern "C" {
        uint32_t    n_world;           // world size
        uint32_t    rank;              // my rank
        uint32_t    n_layer_window[32];// number of layers to process in each compute
+        uint32_t    n_gpu_layers;      // number of layers to process on GPU
        bool        unload;            // whether to unload layer weights after use
        char *      master_ip;         // ip address of the master node
        char *      next_node_ip;      // ip address of the next node