support phi4

Author: Animaxx

src/llama.cpp

@@ -12842,7 +12842,13 @@ struct llm_build_context {
         struct ggml_tensor * inp_pos = build_inp_pos();
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa();
+        struct ggml_tensor * KQ_mask = nullptr;
+        if (hparams.n_swa == 0) {
+            // Phi-4 doesn't use sliding window attention
+            KQ_mask = build_inp_KQ_mask();
+        } else {
+            KQ_mask = build_inp_KQ_mask_swa();
+        }
 
         for (int il = 0; il < n_layer; ++il) {
             auto residual = inpL;
@@ -12900,7 +12906,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, lctx, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask_swa, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
             }
 
             if (il == n_layer - 1) {