merge: TurboQuant KV cache types onto upstream master (MTP built-in)

tools/llama-bench/llama-bench.cpp

@@ -493,6 +493,15 @@ static ggml_type ggml_type_from_name(const std::string & s) {
     if (s == "iq4_nl") {
         return GGML_TYPE_IQ4_NL;
     }
+    if (s == "turbo2") {
+        return GGML_TYPE_TURBO2_0;
+    }
+    if (s == "turbo3") {
+        return GGML_TYPE_TURBO3_0;
+    }
+    if (s == "turbo4") {
+        return GGML_TYPE_TURBO4_0;
+    }
 
     return GGML_TYPE_COUNT;
 }

tools/server/server-context.cpp

@@ -886,8 +886,8 @@ private:
 
         int n_ctx_slot = llama_n_ctx_seq(ctx_tgt);
         if (n_ctx_slot > n_ctx_train) {
-            SRV_WRN("the slot context (%d) exceeds the training context of the model (%d) - capping\n", n_ctx_slot, n_ctx_train);
-            n_ctx_slot = n_ctx_train;
+            SRV_WRN("the slot context (%d) exceeds the training context of the model (%d) - using rope scaling to extend\n", n_ctx_slot, n_ctx_train);
+            // Do not cap: caller has configured rope scaling (--rope-scale / --rope-scaling yarn) to handle extended context.
         }
 
         slots.clear();