model: support vllm d3 models (#30383)

* models: support pplx-decider

* vllm d2

* update

* clarify naming rule

* stricter validation

* d3

tools/server/server-context.cpp

@@ -446,6 +446,10 @@ struct server_slot {
 
     bool can_split() const {
         GGML_ASSERT(task);
+        // a decision over a non-causal context reads the whole prompt in one batch
+        if (task->type == SERVER_TASK_TYPE_DECISION && !llama_get_causal_attn(ctx_tgt)) {
+            return false;
+        }
         // MTP supports splitting - uses task->need_embd() not need_embd()
         if (!task->need_embd()) {
             return true;