Skip to content

[fix] Switch placement of image placeholder for mistral 3.1 #6435

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
merged 1 commit into from
Jul 30, 2025
Merged

[fix] Switch placement of image placeholder for mistral 3.1 #6435

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
6 changes: 4 additions & 2 deletions tensorrt_llm/inputs/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -254,10 +254,12 @@ class MultimodalPlaceholderPlacement(enum.Enum):
"mllama": MultimodalPlaceholderPlacement.BEFORE_TEXT,
"hyperclovax_vlm": MultimodalPlaceholderPlacement.AFTER_TEXT,
"gemma3": MultimodalPlaceholderPlacement.BEFORE_TEXT,
# NOTE: for mistral3 multimodal models, it does not strictly have to be after the text.
# NOTE: for mistral3 multimodal models, it does not strictly have to be before the text.
# Ref: https://github.com/mistralai/mistral-common/blob/039465db2bdc0486df36365c9bdb428188482a18/
# src/mistral_common/tokens/tokenizers/base.py#L326
"mistral3": MultimodalPlaceholderPlacement.AFTER_TEXT,
# However, accuracy tests show that the model generates higher quality output when the image
# precedes the text (the relative difference can be as much as ~30% for both vLLM and TRT-LLM).
"mistral3": MultimodalPlaceholderPlacement.BEFORE_TEXT,
"phi4mm": MultimodalPlaceholderPlacement.BEFORE_TEXT,
}
assert len(PLACEHOLDER_PLACEMENT_MAP) == len(ALL_SUPPORTED_MULTIMODAL_MODELS)
Expand Down
7 changes: 5 additions & 2 deletions tests/integration/defs/test_e2e.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2039,8 +2039,11 @@ def test_ptp_quickstart_multimodal(llm_root, llm_venv, model_name, model_path,
"mistral-small-3.1-24b-instruct": {
"image": [
["dramatic", "seascape", "ocean", "turbulent", "waves", "dark"],
["scenic", "rock", "landscape", "snow", "altitude"],
["highway", "traffic", "directions", "lanes", "Jurong"],
["scenic", "rock", "landscape", "monolith", "formation"],
[
"multi-lane", "highway", "moderate", "traffic", "flow",
"vehicles", "congestion"
],
],
},
"gemma-3-27b-it": {
Expand Down