model, mtmd: fix gemma4 vision handling (#28335)

* model, mtmd: fix gemma4 vision handling

* nits

src/models/deepseek4.cpp

@@ -68,7 +68,7 @@ void llama_model_deepseek4::load_arch_hparams(llama_model_loader & ml) {
     hparams.set_swa_pattern(0);
     // tokens of an image span attend bidirectionally to the whole span, the window only applies to older tokens
     // ref: get_window_topk_idxs_visible in the reference impl
-    hparams.swa_full_non_causal = true;
+    hparams.non_causal_type = LLAMA_NON_CAUSAL_TYPE_SWA_FULL;
     for (uint32_t il = hparams.n_layer(); il < hparams.n_layer_all; ++il) {
         hparams.is_swa_impl[il] = true;
     }

src/models/gemma4.cpp

@@ -19,6 +19,11 @@ void llama_model_gemma4::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_SWA,  hparams.n_embd_head_v_swa);
     ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING,     hparams.f_final_logit_softcapping, false);
 
+    // when non_causal is set, the model will use bidirectional attention on SWA layers only, while dense layers will remain causal
+    // ref: use_bidirectional_attention == "vision" in HF config
+    // note: E2B/E4B are always causal, bypassing this logic
+    hparams.non_causal_type = LLAMA_NON_CAUSAL_TYPE_SWA_ONLY;
+
     switch (hparams.n_layer()) {
         case 30: type = LLM_TYPE_26B_A4B; break;
         case 35: type = LLM_TYPE_E2B; break;