[Models] Consolidate Deepseek-OCR2 processor (#33909)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py
2026-02-05 18:29:20 +00:00
committed by GitHub
parent a57c8228ff
commit 87d0d17ab5
5 changed files with 52 additions and 336 deletions
+1 -1
View File
@@ -31,7 +31,7 @@ class CustomQwen2Decoder(nn.Module):
num_key_value_heads: int = 2,
intermediate_size: int = 4864,
vocab_size: int = 151936,
attn_implementation: str = "sdpa", # ⭐
attn_implementation: str = "sdpa",
rms_norm_eps: float = 1e-06,
rope_theta: float = 1000000.0,
attention_dropout: float = 0.0,
+10 -2
View File
@@ -52,7 +52,6 @@ from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekVLV2Config
from vllm.transformers_utils.processors.deepseek_ocr import (
BASE_SIZE,
CROP_MODE,
IMAGE_SIZE,
DeepseekOCRProcessor,
count_tiles,
)
@@ -66,6 +65,7 @@ from .deepencoder import DeepCLIPVisionTransformer, build_sam_vit_b
from .deepseek_vl2 import MlpProjector
# The image token id may be various
IMAGE_SIZE = 640
_IMAGE_TOKEN = "<image>"
@@ -190,7 +190,15 @@ class DeepseekOCRProcessingInfo(BaseProcessingInfo):
return self.ctx.get_hf_config(DeepseekVLV2Config)
def get_hf_processor(self, **kwargs: object):
return self.ctx.get_hf_processor(DeepseekOCRProcessor, **kwargs)
v1_processor_config = dict(
image_size=IMAGE_SIZE,
base_size=BASE_SIZE,
crop_mode=CROP_MODE,
strategy="v1",
)
return self.ctx.get_hf_processor(
DeepseekOCRProcessor, **{**kwargs, **v1_processor_config}
)
def get_supported_mm_limits(self) -> Mapping[str, int | None]:
return {"image": None}
+12 -4
View File
@@ -48,11 +48,10 @@ from vllm.multimodal.processing import (
from vllm.sequence import IntermediateTensors
from vllm.tokenizers import cached_tokenizer_from_config
from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekVLV2Config
from vllm.transformers_utils.processors.deepseek_ocr2 import (
from vllm.transformers_utils.processors.deepseek_ocr import (
BASE_SIZE,
CROP_MODE,
IMAGE_SIZE,
DeepseekOCR2Processor,
DeepseekOCRProcessor,
)
from ...transformers_utils.processors.deepseek_ocr import count_tiles
@@ -62,6 +61,7 @@ from .deepseek_ocr import DeepseekOCRImagePixelInputs
from .deepseek_vl2 import MlpProjector
# The image token id may be various
IMAGE_SIZE = 768 # different from deepseek-ocr
_IMAGE_TOKEN = "<image>"
@@ -70,7 +70,15 @@ class DeepseekOCR2ProcessingInfo(BaseProcessingInfo):
return self.ctx.get_hf_config(DeepseekVLV2Config)
def get_hf_processor(self, **kwargs: object):
return self.ctx.get_hf_processor(DeepseekOCR2Processor, **kwargs)
v2_processor_config = dict(
image_size=IMAGE_SIZE,
base_size=BASE_SIZE,
crop_mode=CROP_MODE,
strategy="v2",
)
return self.ctx.get_hf_processor(
DeepseekOCRProcessor, **{**kwargs, **v2_processor_config}
)
def get_supported_mm_limits(self) -> Mapping[str, int | None]:
return {"image": None}