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 15, 23:30 UTC
[Bugfix] Fix qwen2.5-vl image processor (#13286)
Isotr0py committed
Feb 15, 2025 at 19:00 UTC
7fdaaf48ef2a10c26be7c87103a64a227f245cea
2 files changed
+17
-6
vllm/model_executor/models/qwen2_5_vl.py
+8
-5
@@ -33,10 +33,11 @@ import torch.nn as nn
33
import torch.nn.functional as F
34
from einops import rearrange
35
from transformers import BatchFeature
36
-from transformers.models.qwen2_5_vl import (Qwen2_5_VLImageProcessor,
37
- Qwen2_5_VLProcessor)
36
+from transformers.models.qwen2_5_vl import Qwen2_5_VLProcessor
37
from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import (
38
Qwen2_5_VLConfig, Qwen2_5_VLVisionConfig)
39
+from transformers.models.qwen2_vl import (Qwen2VLImageProcessor,
40
+ Qwen2VLImageProcessorFast)
41
42
from vllm.attention import AttentionMetadata
43
from vllm.config import VllmConfig
@@ -693,7 +694,8 @@ class Qwen2_5_VLProcessingInfo(Qwen2VLProcessingInfo):
694
) -> Qwen2_5_VLProcessor:
695
hf_processor = self.ctx.get_hf_processor(Qwen2_5_VLProcessor)
696
image_processor = hf_processor.image_processor # type: ignore
696
- assert isinstance(image_processor, Qwen2_5_VLImageProcessor)
697
+ assert isinstance(image_processor,
698
+ (Qwen2VLImageProcessor, Qwen2VLImageProcessorFast))
699
700
if min_pixels:
701
image_processor.min_pixels = min_pixels
@@ -713,14 +715,15 @@ class Qwen2_5_VLProcessingInfo(Qwen2VLProcessingInfo):
715
min_pixels: Optional[int] = None,
716
max_pixels: Optional[int] = None,
717
fps: Optional[float] = 2.0,
716
- ) -> Qwen2_5_VLImageProcessor:
718
+ ) -> Union[Qwen2VLImageProcessor, Qwen2VLImageProcessorFast]:
719
hf_processor = self.get_hf_processor(
720
min_pixels=min_pixels,
721
max_pixels=max_pixels,
722
fps=fps,
723
)
724
image_processor = hf_processor.image_processor # type: ignore
723
- assert isinstance(image_processor, Qwen2_5_VLImageProcessor)
725
+ assert isinstance(image_processor,
726
+ (Qwen2VLImageProcessor, Qwen2VLImageProcessorFast))
727
return image_processor
728
729
vllm/model_executor/models/qwen2_vl.py
+9
-1
@@ -31,7 +31,9 @@ import torch
31
import torch.nn as nn
32
import torch.nn.functional as F
33
from einops import rearrange, repeat
34
+from packaging.version import Version
35
from transformers import BatchFeature
36
+from transformers import __version__ as TRANSFORMERS_VERSION
37
from transformers.models.qwen2_vl import (Qwen2VLImageProcessor,
38
Qwen2VLProcessor)
39
from transformers.models.qwen2_vl.configuration_qwen2_vl import (
@@ -746,7 +748,13 @@ class Qwen2VLProcessingInfo(BaseProcessingInfo):
748
hf_processor = self.get_hf_processor(min_pixels=min_pixels,
749
max_pixels=max_pixels)
750
image_processor = hf_processor.image_processor # type: ignore
749
- assert isinstance(image_processor, Qwen2VLImageProcessor)
751
+ if Version(TRANSFORMERS_VERSION) >= Version("4.49"):
752
+ from transformers.models.qwen2_vl import Qwen2VLImageProcessorFast
753
+ assert isinstance(
754
+ image_processor,
755
+ (Qwen2VLImageProcessor, Qwen2VLImageProcessorFast))
756
+ else:
757
+ assert isinstance(image_processor, Qwen2VLImageProcessor)
758
return image_processor
759
760
def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]: