feat(cookbook): add Gemma4 thinking chat template (#2955)

* feat(cookbook): add Gemma4 thinking chat template

* fix(cookbook): place Gemma4 thinking token in system turn

static/js/cookbook.js

@@ -188,6 +188,19 @@ export function _isMetal() {
   return ['metal', 'mps', 'apple'].includes(String(_hwfitCache?.system?.backend || '').toLowerCase());
 }
 
+const GEMMA4_THINKING_CHAT_TEMPLATE = `{% for message in messages %}{% if message['role'] == 'system' %}<|turn>system\n<|think|>{{ message['content'] }}<turn|>\n{% elif message['role'] == 'user' %}<|turn>user\n{{ message['content'] }}<turn|>\n{% elif message['role'] == 'assistant' %}<|turn>model\n{{ message['content'] }}<turn|>\n{% endif %}{% endfor %}{% if add_generation_prompt %}<|turn>model\n<|channel>thought{% endif %}`;
+
+function _isGemma4ThinkingModel(modelName) {
+  const n = (modelName || '').toLowerCase();
+  return n.includes('gemma-4') || n.includes('gemma4');
+}
+
+function _gemma4ThinkingChatTemplateArg(modelName) {
+  return _isGemma4ThinkingModel(modelName)
+    ? _shellQuote(GEMMA4_THINKING_CHAT_TEMPLATE)
+    : '';
+}
+
 /** Detect model-specific vLLM optimizations */
 function _detectModelOptimizations(modelName) {
   const n = (modelName || '').toLowerCase();
@@ -388,6 +401,8 @@ export function _buildServeCmd(f, modelName, backend) {
     const _extraEnv = (f.extra_env ?? '').toString().replace(/\s+/g, ' ').trim();
     if (_extraEnv) cmd += _extraEnv + ' ';
     cmd += `${_vllmBin} serve ${modelName} --host 0.0.0.0 --port ${f.port || '8000'}`;
+    const _gemma4ChatTemplate = _gemma4ThinkingChatTemplateArg(modelName);
+    if (_gemma4ChatTemplate) cmd += ` --chat-template ${_gemma4ChatTemplate}`;
     cmd += ` --tensor-parallel-size ${f.tp || '1'}`;
     cmd += ` --max-model-len ${f.ctx || '8192'}`;
     cmd += ` --gpu-memory-utilization ${f.gpu_mem || '0.90'}`;
@@ -418,6 +433,8 @@ export function _buildServeCmd(f, modelName, backend) {
     const _extraEnv = (f.extra_env ?? '').toString().replace(/\s+/g, ' ').trim();
     if (_extraEnv) cmd += _extraEnv + ' ';
     cmd += `${_py3Bin} -m sglang.launch_server --model-path ${modelName} --host 0.0.0.0 --port ${f.port || '30000'}`;
+    const _gemma4ChatTemplate = _gemma4ThinkingChatTemplateArg(modelName);
+    if (_gemma4ChatTemplate) cmd += ` --chat-template ${_gemma4ChatTemplate}`;
     if (f.tp && f.tp !== '1') cmd += ` --tp ${f.tp}`;
     if (f.ctx) cmd += ` --context-length ${f.ctx}`;
     if (f.gpu_mem && f.gpu_mem !== '0.90') cmd += ` --mem-fraction-static ${f.gpu_mem}`;

tests/test_cookbook_gemma4_thinking_template.py

@@ -0,0 +1,31 @@
+"""Regression coverage for issue #2929: Gemma 4 thinking chat template.
+
+Gemma 4 thinking models need the `<|think|>` control token in the system
+instruction, while the generation prompt should start the model turn with the
+thought channel. Cookbook serve commands should supply that template for
+OpenAI-compatible servers instead of relying on a generic chat template that
+cannot toggle thinking mode.
+"""
+from pathlib import Path
+
+SRC = Path(__file__).resolve().parent.parent / "static/js/cookbook.js"
+
+
+def test_gemma4_thinking_template_uses_google_documented_thinking_placement():
+    text = SRC.read_text(encoding="utf-8")
+
+    assert "GEMMA4_THINKING_CHAT_TEMPLATE" in text
+    assert "<|turn>system\\n<|think|>{{ message['content'] }}<turn|>" in text
+    assert "<|turn>user" in text
+    assert "<|turn>model" in text
+    assert "<|turn>model\\n<|channel>thought" in text
+    assert "<|turn>model\\n<|think|><|channel>thought" not in text
+
+
+def test_vllm_and_sglang_apply_gemma4_thinking_template():
+    text = SRC.read_text(encoding="utf-8")
+
+    assert "function _isGemma4ThinkingModel" in text
+    assert "const _gemma4ChatTemplate" in text
+    assert "if (_gemma4ChatTemplate) cmd += ` --chat-template ${_gemma4ChatTemplate}`;" in text
+    assert text.count("_gemma4ThinkingChatTemplateArg(modelName)") >= 2