gpt-oss : set special tokens back to user defined

Author: aldehir

src/llama-vocab.cpp

@@ -2339,6 +2339,13 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             }
         }
 
+        // @ngxson : quick hack for gpt-oss, always render these tokens
+        for (const auto & t : token_to_id) {
+            if (t.first == "<|channel|>" || t.first == "<|message|>" || t.first == "<|start|>" || t.first == "<|constrain|>") {
+                id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
+            }
+        }
+
         // sanity checks
         if (special_eos_id != LLAMA_TOKEN_NULL && special_eog_ids.count(special_eos_id) == 0) {
             special_eog_ids.insert(special_eos_id);
@@ -2381,6 +2388,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
             if (has_return && has_call && has_end) {
                 special_eog_ids.erase(end_id);
+                id_to_token[end_id].attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
                 LLAMA_LOG_WARN("%s: special_eog_ids contains both '<|return|>' and '<|call|>' tokens, removing '<|end|>' token from EOG list\n", __func__);
             }
         }