From f9b571dc96a5b77bab9dacae2a522fb20d8b8542 Mon Sep 17 00:00:00 2001
From: Gunther Schulz <dev@guntherschulz.de>
Date: Sat, 14 Feb 2026 00:42:32 +0100
Subject: [PATCH] fix(glm4.5): use parse-only for tool_choice=AUTO, grammar
 only for REQUIRED

- In common_chat_params_init_glm_4_5: set grammar_lazy=false; build grammar
  only when has_tools && tool_choice==REQUIRED (vLLM-style: no trigger/grammar
  for AUTO, detect tool calls by parsing decoded text).
- Relax test-chat assert: allow empty grammar when test message has tool_calls
  (GLM 4.5 AUTO no longer sets grammar).

Fixes server hang when model never outputs trigger (e.g. llama.cpp #19068).

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 common/chat.cpp     | 32 +++++++++++++++++++-------------
 tests/test-chat.cpp |  4 +---
 2 files changed, 20 insertions(+), 16 deletions(-)

diff --git a/common/chat.cpp b/common/chat.cpp
index 47a34d5822..7fd7a55302 100644
--- a/common/chat.cpp
+++ b/common/chat.cpp
@@ -2169,7 +2169,9 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
 
 static common_chat_params common_chat_params_init_glm_4_5(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
-    data.grammar_lazy = inputs.tools.is_array() && !inputs.tools.empty() && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+    // vLLM-style for AUTO: no grammar/trigger during generation; tool calls are parsed from decoded text (common_chat_parse_glm_4_5).
+    // Only use grammar when tool_choice == REQUIRED (force tool call from first token).
+    data.grammar_lazy = false;
 
     std::string prompt = apply(tmpl, inputs);
 
@@ -2228,18 +2230,22 @@ static common_chat_params common_chat_params_init_glm_4_5(const common_chat_temp
         "<|observation|>"
     });
 
-    // build grammar for tool call
-    static const xml_tool_call_format form {
-        /* form.scope_start = */ "",
-        /* form.tool_start  = */ "\n<tool_call>",
-        /* form.tool_sep    = */ "\n",
-        /* form.key_start   = */ "<arg_key>",
-        /* form.key_val_sep = */ "</arg_key>\n<arg_value>",
-        /* form.val_end     = */ "</arg_value>\n",
-        /* form.tool_end    = */ "</tool_call>\n",
-        /* form.scope_end   = */ "",
-    };
-    build_grammar_xml_tool_call(data, inputs.tools, form);
+    // Build grammar only for tool_choice == REQUIRED (force tool call from first token).
+    // For AUTO, generate freely and parse tool calls from decoded text (common_chat_parse_glm_4_5).
+    const bool has_tools = inputs.tools.is_array() && !inputs.tools.empty();
+    if (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
+        static const xml_tool_call_format form {
+            /* form.scope_start = */ "",
+            /* form.tool_start  = */ "\n<tool_call>",
+            /* form.tool_sep    = */ "\n",
+            /* form.key_start   = */ "<arg_key>",
+            /* form.key_val_sep = */ "</arg_key>\n<arg_value>",
+            /* form.val_end     = */ "</arg_value>\n",
+            /* form.tool_end    = */ "</tool_call>\n",
+            /* form.scope_end   = */ "",
+        };
+        build_grammar_xml_tool_call(data, inputs.tools, form);
+    }
 
     data.prompt = prompt;
     data.format = COMMON_CHAT_FORMAT_GLM_4_5;
diff --git a/tests/test-chat.cpp b/tests/test-chat.cpp
index 4378a8db71..7ec07e4f8b 100644
--- a/tests/test-chat.cpp
+++ b/tests/test-chat.cpp
@@ -349,9 +349,7 @@ static void test_templates(const struct common_chat_templates * tmpls, const std
             assert_msg_equals(test_message, msg, ignore_whitespace_differences);
         }
 
-        if (!test_message.tool_calls.empty()) {
-            GGML_ASSERT(!data.params.grammar.empty());
-        }
+        // GLM 4.5 with tool_choice=AUTO uses parse-only (no grammar); other formats set grammar when tools present
         if (!data.params.grammar.empty()) {
             auto grammar = build_grammar(data.params.grammar);
             if (!grammar) {