Speed up KV in llamafile-bench

Author: jart

llama.cpp/llama-bench/llama-bench.cpp

@@ -265,8 +265,8 @@ static const cmd_params cmd_params_defaults = {
     /* n_pg          */ {},
     /* n_batch       */ {2048},
     /* n_ubatch      */ {512},
-    /* type_k        */ {GGML_TYPE_F16},
-    /* type_v        */ {GGML_TYPE_F16},
+    /* type_k        */ {X86_HAVE(AVX512_BF16) ? GGML_TYPE_BF16 : GGML_TYPE_F16},
+    /* type_v        */ {X86_HAVE(AVX512_BF16) ? GGML_TYPE_BF16 : GGML_TYPE_F16},
     /* n_threads     */ {cpu_get_num_math()},
     /* n_gpu_layers  */ {0},
     /* split_mode    */ {LLAMA_SPLIT_MODE_LAYER},