bridge.models.qwen_vl.modelling_qwen3_vl.token_classification#

Module Contents#

Classes#

Qwen3VLForTokenClassification

Qwen3.5 VL model with replicated per-token classification postprocessing.

Functions#

_token_classification_output_processor

Project token logits and compute non-vocabulary-parallel loss when requested.

Data#

TokenClassificationOutput

API#

bridge.models.qwen_vl.modelling_qwen3_vl.token_classification.TokenClassificationOutput: TypeAlias#

None

bridge.models.qwen_vl.modelling_qwen3_vl.token_classification._token_classification_output_processor(
*,
hidden_states: torch.Tensor,
output_layer: torch.nn.Module,
output_weight: torch.Tensor | None,
labels: torch.Tensor | None,
runtime_gather_output: bool | None,
**_: Any,
) → torch.Tensor#

Project token logits and compute non-vocabulary-parallel loss when requested.

class bridge.models.qwen_vl.modelling_qwen3_vl.token_classification.Qwen3VLForTokenClassification#

Bases: megatron.bridge.models.qwen_vl.modelling_qwen3_vl.model.Qwen3VLModel

Qwen3.5 VL model with replicated per-token classification postprocessing.

forward(
input_ids: torch.Tensor | None,
position_ids: torch.Tensor | None = None,
attention_mask: torch.Tensor | None = None,
labels: torch.Tensor | None = None,
loss_mask: torch.Tensor | None = None,
inference_params: megatron.core.InferenceParams | None = None,
packed_seq_params: megatron.core.packed_seq_params.PackedSeqParams | None = None,
extra_block_kwargs: dict[str, object] | None = None,
pixel_values: torch.Tensor | None = None,
pixel_values_videos: torch.Tensor | None = None,
image_grid_thw: torch.Tensor | None = None,
video_grid_thw: torch.Tensor | None = None,
image_input_mask: torch.Tensor | None = None,
video_input_mask: torch.Tensor | None = None,
cp_img_num: list[int] | None = None,
images_padded: list[bool] | None = None,
inference_context: object | None = None,
runtime_gather_output: bool | None = None,
mm_token_type_ids: torch.Tensor | None = None,
**kwargs: object,
) → bridge.models.qwen_vl.modelling_qwen3_vl.token_classification.TokenClassificationOutput#

Run the VLM and apply token-classification logits or per-token loss.

Parameters:

  • input_ids – Input token IDs.

  • position_ids – Optional Qwen MRoPE position IDs.

  • attention_mask – Optional language-model attention mask.

  • labels – Optional token-classification labels.

  • loss_mask – Optional per-token supervision mask.

  • inference_params – Megatron inference state; currently unsupported by Qwen3VL.

  • packed_seq_params – Optional packed-sequence metadata.

  • extra_block_kwargs – Extra transformer-block keyword arguments.

  • pixel_values – Optional image patch values.

  • pixel_values_videos – Optional video patch values.

  • image_grid_thw – Image temporal/height/width grid metadata.

  • video_grid_thw – Video temporal/height/width grid metadata.

  • image_input_mask – Positions receiving image embeddings.

  • video_input_mask – Positions receiving video embeddings.

  • cp_img_num – Per-context-parallel-rank image counts.

  • images_padded – Whether individual images were padded.

  • inference_context – Compatibility placeholder for inference context.

  • runtime_gather_output – Runtime output-gather override.

  • mm_token_type_ids – Multimodal token type IDs retained for API compatibility.

  • **kwargs – Additional language-model keyword arguments.

Returns:

Token logits or per-token losses; context-parallel and non-last-stage paths may return a loss-mask tuple or a stage-output dictionary.

Raises:

ValueError – If a caller tries to override the reserved output processor.