ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Mapping proper tokenizer for VLMs

Browse source
This commit is contained in:
Manan17 2026-02-21 01:57:05 +00:00
commit f6ebeb1d42
1 changed files with 7 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

7
studio/backend/core/inference/inference.py
View file

@ -574,6 +574,8 @@ class InferenceBackend:
model_info = self.models[self.active_model_name]
is_vision = model_info.get("is_vision", False)
tokenizer = model_info.get("tokenizer") or model_info.get("processor")
# Unwrap processor → raw tokenizer for VLMs on the text path
tokenizer = getattr(tokenizer, "tokenizer", tokenizer)
top_k = self._normalize_top_k(top_k)
if is_vision and image:
@ -769,7 +771,11 @@ class InferenceBackend:
model_info = self.models[self.active_model_name]
model = model_info["model"]
# For VLMs the stored "tokenizer" is actually the processor.
# Unwrap to get the real tokenizer so TextIteratorStreamer's
# skip_prompt / skip_special_tokens work correctly.
tokenizer = model_info["tokenizer"]
tokenizer = getattr(tokenizer, "tokenizer", tokenizer)
try:
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
@ -879,6 +885,7 @@ class InferenceBackend:
chat_template_info = self.models[self.active_model_name].get("chat_template_info", {})
tokenizer = self.models[self.active_model_name]["tokenizer"]
tokenizer = getattr(tokenizer, "tokenizer", tokenizer)
chat_messages = []