llama : add warning about multi-sqeuence SWA contexts

Author: ggerganov

include/llama.h

@@ -366,6 +366,8 @@ extern "C" {
         bool no_perf;     // measure performance timings
         bool op_offload;  // offload host tensor operations to device
         bool swa_full;    // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+                          // NOTE: setting to false when n_seq_max > 1 can cause bad perforamnce in some cases
+                          //       ref: https://github.com/ggml-org/llama.cpp/pull/13845#issuecomment-2924800573
     };
 
     // model quantization parameters

src/llama-context.cpp

@@ -123,6 +123,11 @@ llama_context::llama_context(
                 __func__, n_ctx_per_seq, hparams.n_ctx_train);
     }
 
+    if (!params.swa_full && cparams.n_seq_max > 1) {
+        LLAMA_LOG_WARN("%s: requested n_seq_max (%u) > 1, but swa_full is not enabled -- performance may be degraded: %s\n",
+                __func__, cparams.n_seq_max, "https://github.com/ggml-org/llama.cpp/pull/13845#issuecomment-2924800573");
+    }
+
     if (!hparams.vocab_only) {
         // GPU backends
         for (auto * dev : model.devices) {