A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 4, 06:15 UTC
Fix DeepSeek-OCR for Transformers v4 (#41460)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
May 5, 2026 at 12:11 UTC
b0765bee17a901f710a0bf4b9490d2c8e795b0a0
1 file changed
+8
-15
vllm/transformers_utils/configs/deepseek_vl2.py
+8
-15
@@ -101,7 +101,6 @@ else:
101
102
class DeepseekVLV2Config(PretrainedConfig):
103
model_type = "deepseek_vl_v2"
104
- architectures: list[str] | None = None
104
105
tile_tag: str = "2D"
106
global_view_pos: str = "head"
@@ -114,17 +113,11 @@ class DeepseekVLV2Config(PretrainedConfig):
113
candidate_resolutions: tuple[tuple[int, int]] = ((384, 384),),
114
**kwargs,
115
):
117
- if "architectures" not in kwargs:
118
- kwargs["architectures"] = ["DeepseekVLV2ForCausalLM"]
116
+ architectures = kwargs.setdefault("architectures", ["DeepseekVLV2ForCausalLM"])
117
120
- vision_config = kwargs.pop("vision_config", {})
121
- self.vision_config = VisionEncoderConfig(**vision_config)
122
-
123
- projector_config = kwargs.pop("projector_config", {})
124
- self.projector_config = MlpProjectorConfig(**projector_config)
125
-
126
- language_config = kwargs.pop("language_config", {})
127
- self.text_config = DeepseekVLV2TextConfig(**language_config)
118
+ self.vision_config = VisionEncoderConfig(**kwargs.pop("vision_config", {}))
119
+ self.projector_config = MlpProjectorConfig(**kwargs.pop("projector_config", {}))
120
+ self.text_config = DeepseekVLV2TextConfig(**kwargs.pop("language_config", {}))
121
122
self.tile_tag = tile_tag
123
self.global_view_pos = global_view_pos
@@ -132,8 +125,8 @@ class DeepseekVLV2Config(PretrainedConfig):
125
self.vocab_size = self.text_config.vocab_size
126
127
# update model_type for OCR models
135
- if "DeepseekOCRForCausalLM" in kwargs["architectures"]:
136
- self.model_type = "deepseek_ocr"
137
- elif "DeepseekOCR2ForCausalLM" in kwargs["architectures"]:
138
- self.model_type = "deepseek_ocr2"
128
+ if "DeepseekOCRForCausalLM" in architectures:
129
+ kwargs["model_type"] = "deepseek_ocr"
130
+ elif "DeepseekOCR2ForCausalLM" in architectures:
131
+ kwargs["model_type"] = "deepseek_ocr2"
132
super().__init__(**kwargs)