[VLM] Qwen2.5-VL

This commit is contained in:
Roger Wang
2025-02-05 13:31:38 -08:00
committed by GitHub
parent 9a5b1554b4
commit bf3b79efb8
14 changed files with 1315 additions and 52 deletions
File diff suppressed because it is too large Load Diff
+8 -8
View File
@@ -650,8 +650,8 @@ class Qwen2VisionTransformer(nn.Module):
return loaded_params
class Qwen2EmbeddingItems(ModalityDataItems[dict[str, torch.Tensor],
dict[str, torch.Tensor]]):
class Qwen2VLEmbeddingItems(ModalityDataItems[dict[str, torch.Tensor],
dict[str, torch.Tensor]]):
def __init__(self, data: dict, modality: str) -> None:
super().__init__(data, modality)
@@ -683,26 +683,26 @@ class Qwen2EmbeddingItems(ModalityDataItems[dict[str, torch.Tensor],
return self.data
class Qwen2ImageEmbeddingItems(Qwen2EmbeddingItems):
class Qwen2VLImageEmbeddingItems(Qwen2VLEmbeddingItems):
def __init__(self, data: dict) -> None:
super().__init__(data, "image")
class Qwen2VideoEmbeddingItems(Qwen2EmbeddingItems):
class Qwen2VLVideoEmbeddingItems(Qwen2VLEmbeddingItems):
def __init__(self, data: dict) -> None:
super().__init__(data, "video")
class Qwen2MultiModalDataParser(MultiModalDataParser):
class Qwen2VLMultiModalDataParser(MultiModalDataParser):
def _parse_image_data(
self,
data: Union[dict[str, torch.Tensor], ModalityData[ImageItem]],
) -> ModalityDataItems[Any, Any]:
if isinstance(data, dict):
return Qwen2EmbeddingItems(data, modality="image")
return Qwen2VLEmbeddingItems(data, modality="image")
return super()._parse_image_data(data)
@@ -711,7 +711,7 @@ class Qwen2MultiModalDataParser(MultiModalDataParser):
data: Union[dict[str, torch.Tensor], ModalityData[VideoItem]],
) -> ModalityDataItems[Any, Any]:
if isinstance(data, dict):
return Qwen2EmbeddingItems(data, modality="video")
return Qwen2VLEmbeddingItems(data, modality="video")
return super()._parse_video_data(data)
@@ -948,7 +948,7 @@ class Qwen2VLMultiModalProcessor(BaseMultiModalProcessor[Qwen2VLProcessingInfo]
):
def _get_data_parser(self) -> MultiModalDataParser:
return Qwen2MultiModalDataParser()
return Qwen2VLMultiModalDataParser()
def _get_prompt_replacements(
self,
+1
View File
@@ -172,6 +172,7 @@ _MULTIMODAL_MODELS = {
"PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"), # noqa: E501
"QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
"Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"), # noqa: E501
"Qwen2_5_VLForConditionalGeneration": ("qwen2_5_vl", "Qwen2_5_VLForConditionalGeneration"), # noqa: E501
"Qwen2AudioForConditionalGeneration": ("qwen2_audio", "Qwen2AudioForConditionalGeneration"), # noqa: E501
"UltravoxModel": ("ultravox", "UltravoxModel"),
# [Encoder-decoder]