Do not allocate KV cache for unused layers (#843)

* Do not allocate KV cache for unused layers

* Do not apply experts weight scale if it is 1

---------

Co-authored-by: Iwan Kawrakow <[email protected]>

Author: ikawrakow

src/llama-build-context.cpp

@@ -863,7 +863,7 @@ llm_expert_gating_func_type   gating_op,
 
         weights = ggml_reshape_3d(ctx, weights, 1, n_expert_used, n_tokens);
     }
-    if (scale_w) {
+    if (scale_w && std::abs(w_scale-1) > 1e-5f) {
         weights = ggml_scale(ctx, weights, w_scale);
         cb(weights, "ffn_moe_weights_scaled", il);
     }

src/llama.cpp

@@ -532,7 +532,7 @@ static bool llama_kv_cache_init(
 
     const struct llama_hparams & hparams = model.hparams;
 
-    const int64_t  n_layer = hparams.n_layer;
+    const int64_t  n_layer = hparams.n_layer - hparams.nextn_predict_layers;
 
     cache.has_shift = false;