bridge.recipes.qwen_vl.qwen25_vl_dataset

Module Contents

Classes

MockQwen25VLDataset	Mock vision-language dataset for Qwen2.5-VL that yields text+image samples.
MockQwen25VLDatasetProvider	DatasetProvider for a mock Qwen2.5-VL vision-language dataset.

API

class bridge.recipes.qwen_vl.qwen25_vl_dataset.MockQwen25VLDataset(size: int, config: Any)

Bases: torch.utils.data.Dataset

Mock vision-language dataset for Qwen2.5-VL that yields text+image samples.

Each sample contains:

• tokens: torch.LongTensor [L]

• labels: torch.LongTensor [L]

• attention_mask: torch.BoolTensor [L] (all ones by default)

• loss_mask: torch.FloatTensor [L]

• position_ids: torch.LongTensor [L]

• pixel_values: torch.FloatTensor [num_images, C, H, W]

• image_grid_thw: torch.LongTensor [num_images, 3]

Initialization

__len__() → int

_generate_random_image() → PIL.Image.Image

_build_inputs() → Tuple[torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor]]

__getitem__(idx: int) → Dict[str, torch.Tensor]

class bridge.recipes.qwen_vl.qwen25_vl_dataset.MockQwen25VLDatasetProvider

Bases: megatron.bridge.training.config.DatasetProvider

DatasetProvider for a mock Qwen2.5-VL vision-language dataset.

Builds train/valid/test datasets using a HF AutoProcessor and the MockQwen25VLDataset implementation.

sequence_length: int

None

hf_model_path: str

‘Qwen/Qwen2.5-VL-3B-Instruct’

prompt: str

‘Describe this image.’

random_seed: int

0

image_size: Tuple[int, int]

(256, 256)

pad_to_max_length: bool

True

create_attention_mask: bool

True

skip_getting_attention_mask_from_dataset: bool

True

num_images: int

1

_processor: Optional[Any]

None

build_datasets(

context: megatron.bridge.training.config.DatasetBuildContext,

)

Create mock Qwen2.5-VL datasets for train/valid/test splits.

Parameters:

context – Provides sample counts and optional tokenizer.

Returns:

Tuple[Optional[Dataset], Optional[Dataset], Optional[Dataset]]