fix(sdk): enable thinking mode on the llama.cpp VLM path

Gemma-4 ships an mmproj file so it loads as a VLM and runs through
vlm.cpp, which hardcoded enable_thinking=false and dropped special
tokens in token_to_piece. Pass enable_thinking through and render
special tokens so reasoning channel delimiters reach the caller.
Also enable allow_special_tokens for all models on the LLM path so
reasoning delimiters are never silently swallowed.

Closes qcom-ai-hub/geniex#1208

Signed-off-by: Paul Zhu <paulzhuzx@gmail.com>

Author: vinovo

sdk/plugins/llama_cpp/src/llm.cpp

@@ -48,7 +48,12 @@ int32_t LlamaLlm::create_impl(const geniex_LlmCreateInput* input) {
         bool is_gpt_oss_model =
             (model_path_lower.find("gpt") != std::string::npos) && (model_path_lower.find("oss") != std::string::npos);
 
-        this->allow_special_tokens = is_gpt_oss_model;
+        // Special tokens must be rendered as text so the FSM in the CLI can
+        // detect thinking/channel delimiters (e.g. <|channel>/<channel|> for
+        // Gemma-4, <|channel|> for GPT-OSS). EOS is already intercepted by
+        // llama_vocab_is_eog() before token_to_piece is called, so enabling
+        // this globally is safe for all models.
+        this->allow_special_tokens = true;
         if (is_gpt_oss_model) {
             tensor_overrides[0]        = {"\\.ffn_(up|down|gate)_exps\\.(weight|bias)", ggml_backend_cpu_buffer_type()};
             tensor_overrides[1]        = {nullptr, nullptr};  // Null terminator

sdk/plugins/llama_cpp/src/vlm.cpp

@@ -154,10 +154,7 @@ int32_t LlamaVlm::apply_chat_template(
         tmpl_inputs.tools = common_chat_tools_parse_oaicompat(nlohmann::ordered_json::parse(std::string(input->tools)));
     }
 
-    if (input->enable_thinking) {
-        GENIEX_LOG_WARN("thinking mode not supported for llama.cpp VLM; ignoring enable_thinking=true");
-    }
-    tmpl_inputs.enable_thinking = false;
+    tmpl_inputs.enable_thinking = input->enable_thinking;
     GENIEX_LOG_DEBUG("applying chat template with add_generation_prompt=true, use_jinja={}", tmpl_inputs.use_jinja);
 
     // Apply chat template
@@ -399,7 +396,9 @@ int32_t LlamaVlm::generate(const geniex_VlmGenerateInput* input, geniex_VlmGener
             break;
         }
 
-        int n = llama_token_to_piece(vocab, token, token_buffer, sizeof(token_buffer) - 1, 0, false);
+        // Render special tokens so reasoning/channel delimiters (e.g. Gemma-4's
+        // <|channel>/<channel|>) reach the caller; EOS is already handled above.
+        int n = llama_token_to_piece(vocab, token, token_buffer, sizeof(token_buffer) - 1, 0, true);
         if (n < 0) n = 0;
         token_buffer[n] = '\0';