fix: guard the qwen4exp PLE prefetch on the model architecture (SIGSEGV at ubatch >= 4096)

src/llama-context.cpp

@@ -1743,7 +1743,9 @@ int llama_context::decode(const llama_batch & batch_inp) {
     {   // warm the page cache for this batch's per-layer-embedding rows while the first chunk is on the GPU;
         // posix_fadvise only, so a wrong prediction costs readahead and nothing else
         extern void qwen4exp_ple_prefetch(const llama_model & model, const llama_token * tokens, int32_t n_tokens);
-        if (batch_inp.token && batch_inp.n_tokens >= 4096) { qwen4exp_ple_prefetch(model, batch_inp.token, batch_inp.n_tokens); }
+        if (model.arch == LLM_ARCH_QWEN4EXP && batch_inp.token && batch_inp.n_tokens >= 4096) {
+            qwen4exp_ple_prefetch(model, batch_inp.token, batch_inp.n_tokens);
+        }
     }
     const uint32_t n_outputs_all = balloc->get_n_outputs();
 

src/models/qwen4exp.cpp

@@ -1749,6 +1749,12 @@ void qwen4exp_ple_prefetch(const llama_model & model_base, const llama_token * t
     if (!tokens || n_tokens < 4096) {
         return;
     }
+    // llama_context::decode calls this for every architecture, so the downcast below is only valid
+    // once the arch is known: on any other model it reads ple_disk out of an unrelated object and
+    // dereferences whatever that happens to hold.
+    if (model_base.arch != LLM_ARCH_QWEN4EXP) {
+        return;
+    }
     const auto & pmodel = static_cast<const llama_model_qwen4exp &>(model_base);
     if (!pmodel.ple_disk || !pmodel.ple_disk->page_cached()) {
         return;