Skip to content

Support tokenizer override per model for multi-model Triton + vLLM serving with OpenAI-Compatible #8321

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

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Support tokenizer override per model for multi-model Triton + vLLM serving with OpenAI-Compatible #8321

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
cde2484
Add tokenizer override logic for mapped HF model names
JunmooByun Jul 31, 2025
36d7b43
Merge branch 'main' into feat/tokenizer-override
JunmooByun Aug 4, 2025
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: 15 additions & 2 deletions python/openai/openai_frontend/engine/triton_engine.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -395,10 +395,16 @@ def _get_tokenizer(self, tokenizer_name: str):
def _get_model_metadata(self) -> Dict[str, TritonModelMetadata]:
# One tokenizer and creation time shared for all loaded models for now.
model_metadata = {}

# Mapping of custom model identifiers to their corresponding Hugging Face model names
HF_MODEL_NAME_MAP = {
"llama-3.1-8b-instruct": "meta-llama/Meta-Llama-3.1-8B-Instruct",
"mistral-nemo-instruct-2407": "mistralai/Mistral-Nemo-Instruct-2407",
}

# Read all triton models and store the necessary metadata for each
for name, _ in self.server.models().keys():
model = self.server.model(name)
model = self.server.model(name)
backend = model.config()["backend"]
# Explicitly handle ensembles to avoid any runtime validation errors
if not backend and model.config()["platform"] == "ensemble":
Expand All @@ -410,12 +416,19 @@ def _get_model_metadata(self) -> Dict[str, TritonModelMetadata]:
lora_names = _get_vllm_lora_names(
self.server.options.model_repository, name, model.version
)
# Map to Hugging Face model name if available
hf_model_name = HF_MODEL_NAME_MAP.get(name, name)
# Try to get tokenizer for the mapped model name
tokenizer_override = get_tokenizer(hf_model_name)
# Use the override tokenizer if available; otherwise fall back to default
tokenizer = tokenizer_override if tokenizer_override else self.tokenizer

metadata = TritonModelMetadata(
name=name,
backend=backend,
model=model,
tokenizer=self.tokenizer,
# tokenizer=self.tokenizer,
tokenizer=tokenizer,
lora_names=lora_names,
create_time=self.create_time,
request_converter=self._determine_request_converter(backend),
Expand Down