Merge commit 'e408d4351a4ad143656672479d8ae1479306ce31' into concedo_experimental

# Conflicts:
#	ggml/CMakeLists.txt

src/llama-context.cpp

@@ -1318,8 +1318,8 @@ int llama_context::decode(llama_batch & inp_batch) {
             n_outputs = n_outputs_new;
         }
 
-        // non-causal masks do not use the KV cache
-        if (hparams.causal_attn) {
+        // find KV slot
+        {
             kv_self_update();
 
             // if we have enough unused cells before the current head ->