Skip to content

Narrow prefix-preserving check to the actual requirement #5458

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
qgallouedec merged 14 commits into from
Apr 7, 2026
+90 −74
Merged

Narrow prefix-preserving check to the actual requirement #5458

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
4b3aa51
Narrow prefix-preserving check to the actual requirement
qgallouedec Apr 5, 2026
0894910
Merge branch 'main' into narrow-prefix-preserving-check
qgallouedec Apr 5, 2026
730070b
Update chat template examples to use multiplication function calls
qgallouedec Apr 5, 2026
4622d77
style
qgallouedec Apr 5, 2026
8a00354
tools in dummy
qgallouedec Apr 5, 2026
cd8cbfc
Enhance chat template handling to include tool call formatting in mes…
qgallouedec Apr 5, 2026
8b35321
align grpo and async
qgallouedec Apr 5, 2026
160d6a0
Merge branch 'main' into narrow-prefix-preserving-check
qgallouedec Apr 6, 2026
8dd341b
revert no content
qgallouedec Apr 6, 2026
7e3ddd7
docstyle ignore
qgallouedec Apr 6, 2026
71bf73f
Merge branch 'main' into narrow-prefix-preserving-check
qgallouedec Apr 6, 2026
103d3c9
Merge branch 'main' into narrow-prefix-preserving-check
qgallouedec Apr 6, 2026
fffcb67
Merge branch 'main' into narrow-prefix-preserving-check
qgallouedec Apr 7, 2026
87131e5
handle content=""
qgallouedec Apr 7, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 7 additions & 10 deletions tests/test_chat_template_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -153,6 +153,8 @@ def test_prefix_preserving_template(self):
{{- '<|im_start|>user\n' + message.content + '<|im_end|>\n' }}
{%- elif message.role == 'assistant' %}
{{- '<|im_start|>assistant\n' + message.content + '<|im_end|>\n' }}
{%- elif message.role == 'tool' %}
{{- '<|im_start|>tool\n' + message.content + '<|im_end|>\n' }}
{%- endif %}

{%- endfor %}
Expand All @@ -164,8 +166,9 @@ def test_prefix_preserving_template(self):

def test_non_prefix_preserving_template(self):
tokenizer = AutoTokenizer.from_pretrained("trl-internal-testing/tiny-Qwen3MoeForSequenceClassification")
# The following template is quite typical of models like Qwen3 and GPT-OSS, where the thinking part is
# only present for last assistant message, which makes it non-prefix-preserving.
# The following template is quite typical of models like Qwen3 and GPT-OSS, where the thinking part (even
# empty) is only present for last assistant message, which makes it non-prefix-preserving: appending a tool
# message changes the earlier output.
# docstyle-ignore
tokenizer.chat_template = textwrap.dedent(r"""
{%- if messages[0].role == 'system' %}
Expand Down Expand Up @@ -203,6 +206,8 @@ def test_non_prefix_preserving_template(self):
{{- '<|im_start|>' + message.role + '\n' + content }}
{%- endif %}
{{- '<|im_end|>\n' }}
{%- elif message.role == "tool" %}
{{- '<|im_start|>tool\n' + content + '<|im_end|>\n' }}
{%- endif %}
{%- endfor %}
{%- if add_generation_prompt %}
Expand All @@ -218,14 +223,6 @@ def test_non_prefix_preserving_template(self):
"tokenizer_name",
[
pytest.param("trl-internal-testing/tiny-Qwen3MoeForSequenceClassification", id="qwen3"),
pytest.param(
"trl-internal-testing/tiny-Qwen3_5ForConditionalGeneration",
id="qwen35",
marks=pytest.mark.skipif(
Version(transformers.__version__) < Version("5.0.0"),
reason="Qwen3.5 tokenizer requires transformers>=5.0.0",
),
),
],
)
class TestGetTrainingChatTemplate:
Expand Down
84 changes: 32 additions & 52 deletions trl/chat_template_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -630,6 +630,10 @@ def is_chat_template_prefix_preserving(tokenizer: PreTrainedTokenizer) -> bool:
"""
Check whether the chat template preserves prefixes when applied.

A prefix-preserving chat template renders earlier messages identically regardless of what messages follow. This
property is required by `_get_tool_suffix_ids`, which extracts tool response formatting tokens by comparing
tokenizations with and without tool messages appended.

Args:
tokenizer (`PreTrainedTokenizer`):
Tokenizer instance to check.
Expand All @@ -638,24 +642,22 @@ def is_chat_template_prefix_preserving(tokenizer: PreTrainedTokenizer) -> bool:
`bool`:
`True` if the chat template preserves prefixes, `False` otherwise.
"""
# Use the same dummy messages as _get_tool_suffix_ids to test the exact property it relies on.
dummy_tool_calls = [{"type": "function", "function": {"name": "dummy", "arguments": {}}}]
messages1 = [
{"role": "user", "content": "What color is the sky?"},
{"role": "user", "content": "dummy"},
{"role": "assistant", "tool_calls": dummy_tool_calls},
]
messages2 = [
{"role": "user", "content": "What color is the sky?"},
{"role": "assistant", "content": "It is blue."},
]
messages3 = [
{"role": "user", "content": "What color is the sky?"},
{"role": "assistant", "content": "It is blue."},
{"role": "user", "content": "And at night?"},
{"role": "user", "content": "dummy"},
{"role": "assistant", "tool_calls": dummy_tool_calls},
{"role": "tool", "name": "dummy", "content": "dummy"},
]

text1 = tokenizer.apply_chat_template(messages1, tokenize=False, add_generation_prompt=True)
text2 = tokenizer.apply_chat_template(messages2, tokenize=False)
text3 = tokenizer.apply_chat_template(messages3, tokenize=False)
text1 = tokenizer.apply_chat_template(messages1, tokenize=False)
text2 = tokenizer.apply_chat_template(messages2, tokenize=False, add_generation_prompt=True)

return text2.startswith(text1) and text3.startswith(text2)
return text2.startswith(text1)


# Modifications:
Expand Down Expand Up @@ -749,33 +751,12 @@ def is_chat_template_prefix_preserving(tokenizer: PreTrainedTokenizer) -> bool:
{%- endif %}"""


# Modifications:
# - {%- if '</think>' in content %}
# + {%- if '<think>' in content and '</think>' in content %}
# Always check for both tags to avoid edge cases where the model generates only one tag, which would otherwise be parsed incorrectly
# - {{- '<|im_start|>' + message.role + '\n' + content }}
# + {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
# Always include thinking block during training. It's important to have a prefix-preserving template.
def _patch_qwen3_5_training_template(template: str) -> str:
return template.replace(
"{%- if '</think>' in content %}",
"{%- if '<think>' in content and '</think>' in content %}",
).replace(
"{{- '<|im_start|>' + message.role + '\\n' + content }}",
"{{- '<|im_start|>' + message.role + '\\n<think>\\n' + reasoning_content + '\\n</think>\\n\\n' + content }}",
)


qwen3_5_training_chat_template_2b_and_below = _patch_qwen3_5_training_template(qwen3_5_chat_template_2b_and_below)
qwen3_5_training_chat_template_4b_and_above = _patch_qwen3_5_training_template(qwen3_5_chat_template_4b_and_above)
Copy link
Copy Markdown
Member

@albertvillanova albertvillanova Apr 7, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You mention Qwen3.5 doesn't need patching anymore. Does this depend on the transformers version?

Copy link
Copy Markdown
Member Author

@qgallouedec qgallouedec Apr 7, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

no, it's really about the chat template:

from transformers import AutoTokenizer, AutoProcessor

dummy_tool_calls = [{"type": "function", "function": {"name": "dummy", "arguments": {}}}]

messages1 = [
    {"role": "user", "content": "dummy"},
    {"role": "assistant", "content": "", "tool_calls": dummy_tool_calls},
]
messages2 = messages1 + [
    {"role": "tool", "name": "dummy", "content": "dummy"},
]


model_id = "Qwen/Qwen3-0.6B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
text1 = tokenizer.apply_chat_template(messages1, tokenize=False)
text2 = tokenizer.apply_chat_template(messages2, tokenize=False, add_generation_prompt=True)
print(f"\n{'='*60}")
print(f"{model_id}")
print(f"Prefix-preserving: {text2.startswith(text1)}")
print(repr(text1))
print(repr(text2))

model_id = "Qwen/Qwen3.5-0.8B"
tokenizer = AutoProcessor.from_pretrained(model_id)
text1 = tokenizer.apply_chat_template(messages1, tokenize=False)
text2 = tokenizer.apply_chat_template(messages2, tokenize=False, add_generation_prompt=True)
print(f"\n{'='*60}")
print(f"{model_id}")
print(f"Prefix-preserving: {text2.startswith(text1)}")
print(repr(text1))
print(repr(text2))
============================================================
Qwen/Qwen3-0.6B
Prefix-preserving: False
'<|im_start|>user\ndummy<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n<tool_call>\n{"name": "dummy", "arguments": {}}\n</tool_call><|im_end|>\n'
'<|im_start|>user\ndummy<|im_end|>\n<|im_start|>assistant\n<tool_call>\n{"name": "dummy", "arguments": {}}\n</tool_call><|im_end|>\n<|im_start|>user\n<tool_response>\ndummy\n</tool_response><|im_end|>\n<|im_start|>assistant\n'

============================================================
Qwen/Qwen3.5-0.8B
Prefix-preserving: True
'<|im_start|>user\ndummy<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n<tool_call>\n<function=dummy>\n</function>\n</tool_call><|im_end|>\n'
'<|im_start|>user\ndummy<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n<tool_call>\n<function=dummy>\n</function>\n</tool_call><|im_end|>\n<|im_start|>user\n<tool_response>\ndummy\n</tool_response><|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n'



def get_training_chat_template(tokenizer: PreTrainedTokenizer) -> str | None:
r"""
Get a prefix-preserving chat template for training, if needed.

If the tokenizer's template isn't prefix-preserving, returns a training-compatible template (currently Qwen3 and
Qwen3.5 supported). Otherwise, returns `None`.
If the tokenizer's template isn't prefix-preserving, returns a training-compatible template (currently Qwen3
supported). Otherwise, returns `None`.

Args:
tokenizer (`PreTrainedTokenizer`):
Expand All @@ -793,27 +774,30 @@ def get_training_chat_template(tokenizer: PreTrainedTokenizer) -> str | None:

>>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
>>> messages1 = [
... {"role": "user", "content": "What color is the sky?"},
... {"role": "assistant", "content": "It is blue."},
... {"role": "user", "content": "What is 2 * 3?"},
... {
... "role": "assistant",
... "tool_calls": [{"type": "function", "function": {"name": "multiply", "arguments": {"a": 2, "b": 3}}}],
... },
... ]
>>> messages2 = [
... {"role": "user", "content": "What color is the sky?"},
... {"role": "assistant", "content": "It is blue."},
... {"role": "user", "content": "And at night?"},
>>> messages2 = messages1 + [
... {"role": "tool", "name": "multiply", "content": "6"},
... ]
>>> tokenizer.apply_chat_template(messages1, tokenize=False)
'<|im_start|>user\nWhat color is the sky?<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\nIt is blue.<|im_end|>\n'
'<|im_start|>user\nWhat is 2 * 3?<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n<tool_call>\n{"name": "multiply", "arguments": {"a": 2, "b": 3}}\n</tool_call><|im_end|>\n'

>>> tokenizer.apply_chat_template(messages2, tokenize=False)
'<|im_start|>user\nWhat color is the sky?<|im_end|>\n<|im_start|>assistant\nIt is blue.<|im_end|>\n<|im_start|>user\nAnd at night?<|im_end|>\n'
>>> tokenizer.apply_chat_template(messages2, tokenize=False, add_generation_prompt=True)
'<|im_start|>user\nWhat is 2 * 3?<|im_end|>\n<|im_start|>assistant\n<tool_call>\n{"name": "multiply", "arguments": {"a": 2, "b": 3}}\n</tool_call><|im_end|>\n<|im_start|>user\n<tool_response>\n6\n</tool_response><|im_end|>\n<|im_start|>assistant\n'

>>> # ^ think tags missing
>>> # ^ think tags missing
>>> chat_template = get_training_chat_template(tokenizer)
>>> tokenizer.apply_chat_template(messages1, tokenize=False, chat_template=chat_template)
'<|im_start|>user\nWhat color is the sky?<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\nIt is blue.<|im_end|>\n'
'<|im_start|>user\nWhat is 2 * 3?<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n<tool_call>\n{"name": "multiply", "arguments": {"a": 2, "b": 3}}\n</tool_call><|im_end|>\n'

>>> tokenizer.apply_chat_template(messages2, tokenize=False, chat_template=chat_template)
'<|im_start|>user\nWhat color is the sky?<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\nIt is blue.<|im_end|>\n<|im_start|>user\nAnd at night?<|im_end|>\n'
>>> tokenizer.apply_chat_template(
... messages2, tokenize=False, add_generation_prompt=True, chat_template=chat_template
... )
'<|im_start|>user\nWhat is 2 * 3?<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n<tool_call>\n{"name": "multiply", "arguments": {"a": 2, "b": 3}}\n</tool_call><|im_end|>\n<|im_start|>user\n<tool_response>\n6\n</tool_response><|im_end|>\n<|im_start|>assistant\n'
```
"""
# First check if patching is needed
Expand All @@ -822,10 +806,6 @@ def get_training_chat_template(tokenizer: PreTrainedTokenizer) -> str | None:

if tokenizer.chat_template == qwen3_chat_template:
return qwen3_training_chat_template
if tokenizer.chat_template == qwen3_5_chat_template_2b_and_below:
return qwen3_5_training_chat_template_2b_and_below
if tokenizer.chat_template == qwen3_5_chat_template_4b_and_above:
return qwen3_5_training_chat_template_4b_and_above
else:
raise ValueError(
"The tokenizer's chat template is not prefix-preserving and patching is not supported for this template. "
Expand Down
8 changes: 7 additions & 1 deletion trl/trainer/grpo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1406,7 +1406,13 @@ def _generate_single_turn(self, prompt_ids, images, multimodal_fields):

def _get_tool_suffix_ids(self, tool_messages):
"""Get token IDs for tool result formatting by using a minimal dummy conversation."""
dummy_messages = [{"role": "user", "content": "dummy"}, {"role": "assistant", "content": "dummy"}]
dummy_messages = [
{"role": "user", "content": "dummy"},
{
"role": "assistant",
"tool_calls": [{"type": "function", "function": {"name": "dummy", "arguments": {}}}],
},
]
if self._is_vlm:
dummy_messages = prepare_multimodal_messages(dummy_messages)
prefix_ids = self.processing_class.apply_chat_template(
Expand Down
Loading