Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Keep the surrounding {{ }} when stripping bos_token from a chat template (#10682)
Vineeth Sai Varikuntla6c54ae2cac8c
diff --git a/tests/python/test_gemma4_base_bos_token.py b/tests/python/test_gemma4_base_bos_token.pyindex eef48df1162..c61458d23eb 100644--- a/tests/python/test_gemma4_base_bos_token.py+++ b/tests/python/test_gemma4_base_bos_token.py@@ -259,6 +259,52 @@ def test_real_tokenizer_chat_bos_survives_save_reload(tmp_path, prefix):assert tok("Hello")["input_ids"] == [2, 3]+# Llama 2 and its many derivatives put bos_token inside a larger expression rather than in a+# template action of its own, so stripping it has to leave the surrounding `{{ ... }}` intact.+LLAMA2_TEMPLATE = (+ "{% for message in messages %}"+ "{% if message['role'] == 'user' %}"+ "{{ bos_token + '[INST] ' + message['content'].strip() + ' [/INST]' }}"+ "{% else %}"+ "{{ ' ' + message['content'].strip() + ' ' + eos_token }}"+ "{% endif %}"+ "{% endfor %}"+)+++def _render(template, **kwargs):+ jinja2 = pytest.importorskip("jinja2")+ return jinja2.Template(template).render(+ messages = [{"role": "user", "content": "hi"}],+ bos_token = "<s>",+ eos_token = "</s>",+ **kwargs,+ )+++def test_stripping_bos_from_an_expression_keeps_the_template_valid():+ stripped = tu._strip_bos_from_chat_template_text(LLAMA2_TEMPLATE)+ assert "bos_token" not in stripped+ # The rest of the expression has to stay an expression: dropping the opening `{{` too would+ # leave a dangling `}}` and render the Jinja source as literal text.+ assert stripped.count("{{") == stripped.count("}}")+ assert _render(stripped) == "[INST] hi [/INST]"+++def test_stripping_a_standalone_bos_action_is_unchanged():+ template = "{{ bos_token }}{% for m in messages %}<t>{{ m['content'] }}</t>{% endfor %}"+ stripped = tu._strip_bos_from_chat_template_text(template)+ assert stripped == "{% for m in messages %}<t>{{ m['content'] }}</t>{% endfor %}"+ assert _render(stripped) == "<t>hi</t>"+++def test_dedupe_leaves_a_renderable_template_for_an_expression_bos():+ tok = _gemma4_base(add_bos_token = True, chat_template = LLAMA2_TEMPLATE)+ tu._dedupe_bos_chat_template(tok)+ assert "bos_token" not in tok.chat_template+ assert _render(tok.chat_template) == "[INST] hi [/INST]"++def test_export_helper_strips_dict_chat_template_without_crash():tok = _gemma4_base(add_bos_token = True,diff --git a/unsloth/tokenizer_utils.py b/unsloth/tokenizer_utils.pyindex fd69109fe84..f06c06363e7 100644--- a/unsloth/tokenizer_utils.py+++ b/unsloth/tokenizer_utils.py@@ -149,7 +149,8 @@ def _strip_bos_from_chat_template_text(chat_template):if not isinstance(chat_template, str) or not chat_template:return chat_templatestripped = re.sub(r"\{[\s\-]*\{[\s\-]*bos\_token[\s\-]*\}[\s\-]*\}", "", chat_template, count = 1)- return re.sub(r"\{[\s\-]*\{[\s\-]*bos\_token[\s\-]*\+[\s\-]*", "", stripped, count = 1)+ # Keep the opening `{{` for Llama 2 expressions so removing bos_token does not leave a dangling `}}`.+ return re.sub(r"(\{[\s\-]*\{[\s\-]*)bos\_token[\s\-]*\+[\s\-]*", r"\1", stripped, count = 1)def _dedupe_bos_chat_template(tokenizer):