Skip to content

Check prefix preservation at the token level #5559

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 8 commits into from
Apr 20, 2026
Merged

Check prefix preservation at the token level #5559

Changes from 2 commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
7258c00
Check prefix preservation at the token level
qgallouedec Apr 15, 2026
a076c4a
fix vml
qgallouedec Apr 15, 2026
84e90a0
fix vlm
qgallouedec Apr 16, 2026
22dc962
Merge branch 'main' into check-prefix-preservation-at-the-token-level
qgallouedec Apr 16, 2026
af51127
Merge branch 'main' into check-prefix-preservation-at-the-token-level
qgallouedec Apr 17, 2026
5f3f91b
fix
qgallouedec Apr 17, 2026
d3f2474
rm duplicate
qgallouedec Apr 17, 2026
723c064
Merge branch 'main' into check-prefix-preservation-at-the-token-level
qgallouedec Apr 20, 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: 12 additions & 5 deletions trl/chat_template_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -419,19 +419,26 @@ def is_chat_template_prefix_preserving(tokenizer: PreTrainedTokenizer) -> bool:
{"role": "tool", "name": "dummy", "content": "dummy"},
]

if isinstance(tokenizer, ProcessorMixin):
from PIL import Image

dummy_image = Image.new("RGB", (8, 8))
messages1 = prepare_multimodal_messages(messages1, images=[dummy_image])
messages2 = prepare_multimodal_messages(messages2, images=[dummy_image])

@qgallouedec qgallouedec Apr 15, 2026

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

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

taken from #5558 ta avoid that tokenize=False -> True breaks this function for VLMs

Comment thread
qgallouedec marked this conversation as resolved.

try:
text1 = tokenizer.apply_chat_template(messages1, tokenize=False)
text2 = tokenizer.apply_chat_template(messages2, tokenize=False, add_generation_prompt=True)
ids1 = tokenizer.apply_chat_template(messages1, tokenize=True, return_dict=False)
ids2 = tokenizer.apply_chat_template(messages2, tokenize=True, return_dict=False, add_generation_prompt=True)
Comment thread
qgallouedec marked this conversation as resolved.
Outdated
except TypeError:
# Best-effort fallback for templates that reject dict args (e.g. DeepSeek-V3). This is a chat template
# bug (see transformers#45419), and the training chat template fixes it to avoid blocking users.
dummy_tool_calls = [{"type": "function", "function": {"name": "dummy", "arguments": "{}"}}]
messages1[1]["tool_calls"] = dummy_tool_calls
messages2[1]["tool_calls"] = dummy_tool_calls
text1 = tokenizer.apply_chat_template(messages1, tokenize=False)
text2 = tokenizer.apply_chat_template(messages2, tokenize=False, add_generation_prompt=True)
ids1 = tokenizer.apply_chat_template(messages1, tokenize=True, return_dict=False)
ids2 = tokenizer.apply_chat_template(messages2, tokenize=True, return_dict=False, add_generation_prompt=True)
Comment thread
qgallouedec marked this conversation as resolved.
Outdated

return text2.startswith(text1)
return ids2[: len(ids1)] == ids1
Comment thread
cursor[bot] marked this conversation as resolved.


deepseekv3_training_chat_template = (_CHAT_TEMPLATES_DIR / "deepseekv3_training.jinja").read_text()
Expand Down
Loading