[BUGFIX] fix ovis2_5 multimodal tokens (#52623)

Signed-off-by: Milosz Grunwald <milosz.grunwald@intel.com>
This commit is contained in:
Miłosz Grunwald
2026-09-25 07:18:54 +08:00
committed by GitHub
parent 6a86bf626e
commit bd4cb3fe4e
2 changed files with 19 additions and 3 deletions
+12 -1
View File
@@ -175,6 +175,18 @@ class VisualTokenizer(torch.nn.Module):
class Ovis2_5ProcessingInfo(BaseProcessingInfo):
def __init__(self, ctx) -> None:
super().__init__(ctx)
# extra_special_tokens adds tokens to the tokenizer lazily, on first
# access, but the renderer deep-copies this tokenizer into a thread
# pool right after this processor is constructed; the copies never
# see mutations made afterwards. Force it now, before that happens,
# on the tokenizer we actually use (get_hf_processor()'s cache can
# otherwise hand back one bound to a self-loaded tokenizer instead).
hf_processor = self.get_hf_processor()
hf_processor.tokenizer = self.get_tokenizer()
hf_processor.extra_special_tokens # noqa: B018
def get_hf_config(self):
return self.ctx.get_hf_config()
@@ -390,7 +402,6 @@ class Ovis2_5MultiModalProcessor(BaseMultiModalProcessor[Ovis2_5ProcessingInfo])
out_mm_kwargs: MultiModalKwargsItems,
) -> list[PromptReplacement]:
hf_processor = self.info.get_hf_processor()
placeholder = {
"image": hf_processor.get_token_value("image_token"),
"video": hf_processor.get_token_value("video_token"),