Skip to content

Fix DeepSeek-OCR for Transformers v4 #41460

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
Isotr0py merged 4 commits into from
May 5, 2026
+8 −15
Merged

Fix DeepSeek-OCR for Transformers v4 #41460

Changes from 2 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
97588ea
Fix DeepSeek-OCR
hmellor May 1, 2026
f983f21
Keep more original order
hmellor May 1, 2026
417ef8b
Merge branch 'main' into fix-deepseek-ocr
Isotr0py May 4, 2026
a8ce95e
Merge branch 'main' into fix-deepseek-ocr
hmellor May 5, 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
23 changes: 8 additions & 15 deletions vllm/transformers_utils/configs/deepseek_vl2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -101,7 +101,6 @@ class DeepseekVLV2TextConfig(DeepseekV2Config):

class DeepseekVLV2Config(PretrainedConfig):
model_type = "deepseek_vl_v2"
architectures: list[str] | None = None

tile_tag: str = "2D"
global_view_pos: str = "head"
Expand All @@ -114,26 +113,20 @@ def __init__(
candidate_resolutions: tuple[tuple[int, int]] = ((384, 384),),
**kwargs,
):
if "architectures" not in kwargs:
kwargs["architectures"] = ["DeepseekVLV2ForCausalLM"]
architectures = kwargs.setdefault("architectures", ["DeepseekVLV2ForCausalLM"])

vision_config = kwargs.pop("vision_config", {})
self.vision_config = VisionEncoderConfig(**vision_config)

projector_config = kwargs.pop("projector_config", {})
self.projector_config = MlpProjectorConfig(**projector_config)

language_config = kwargs.pop("language_config", {})
self.text_config = DeepseekVLV2TextConfig(**language_config)
self.vision_config = VisionEncoderConfig(**kwargs.pop("vision_config", {}))
self.projector_config = MlpProjectorConfig(**kwargs.pop("projector_config", {}))
self.text_config = DeepseekVLV2TextConfig(**kwargs.pop("language_config", {}))

Comment thread
hmellor marked this conversation as resolved.
self.tile_tag = tile_tag
self.global_view_pos = global_view_pos
self.candidate_resolutions = candidate_resolutions
self.vocab_size = self.text_config.vocab_size
Comment thread
hmellor marked this conversation as resolved.

# update model_type for OCR models
if "DeepseekOCRForCausalLM" in kwargs["architectures"]:
self.model_type = "deepseek_ocr"
elif "DeepseekOCR2ForCausalLM" in kwargs["architectures"]:
self.model_type = "deepseek_ocr2"
if "DeepseekOCRForCausalLM" in architectures:
kwargs["model_type"] = "deepseek_ocr"
elif "DeepseekOCR2ForCausalLM" in architectures:
kwargs["model_type"] = "deepseek_ocr2"
super().__init__(**kwargs)
Loading