bridge.training.vlm_step#

Module Contents#

Functions#

_unwrap_forward_module

Return the innermost wrapped module used for forward signature checks.

_filter_visual_kwargs_for_model

Drop visual kwargs that the target model forward cannot consume.

_project_visual_inputs_for_pp_stage

Drop visual payload tensors from PP stages that only need visual metadata.

get_batch_from_iterator

Get a batch of data from the iterator.

get_batch

Generate a batch.

forward_step

Forward training step.

Data#

_VISUAL_PAYLOAD_FIELDS

_PACKED_SEQ_DEVICE_KEYS

_PACKED_SEQ_HOST_KEYS

_PACKED_SEQ_PARAM_KEYS

API#

bridge.training.vlm_step._VISUAL_PAYLOAD_FIELDS#

‘frozenset(…)’

bridge.training.vlm_step._PACKED_SEQ_DEVICE_KEYS#

(‘cu_seqlens_q’, ‘cu_seqlens_kv’, ‘cu_seqlens_q_padded’, ‘cu_seqlens_kv_padded’)

bridge.training.vlm_step._PACKED_SEQ_HOST_KEYS#

(‘max_seqlen_q’, ‘max_seqlen_kv’)

bridge.training.vlm_step._PACKED_SEQ_PARAM_KEYS#

()

bridge.training.vlm_step._unwrap_forward_module(model: Any) → Any#

Return the innermost wrapped module used for forward signature checks.

bridge.training.vlm_step._filter_visual_kwargs_for_model(
model: Any,
visual_kwargs: collections.abc.Mapping[str, torch.Tensor],
) → dict[str, torch.Tensor]#

Drop visual kwargs that the target model forward cannot consume.

Shared VLM processors may return model-specific fields such as mm_token_type_ids. Keep those fields for models that accept them, but avoid passing them through wrappers into models with stricter signatures.

bridge.training.vlm_step._project_visual_inputs_for_pp_stage(
visual_inputs: Any,
*,
is_first_pp_stage: bool,
) → Any#

Drop visual payload tensors from PP stages that only need visual metadata.

bridge.training.vlm_step.get_batch_from_iterator(
data_iterator: Iterable,
use_mtp: bool = False,
skip_getting_attention_mask_from_dataset: bool = True,
*,
is_first_pp_stage: bool,
is_last_pp_stage: bool,
) → dict[str, Any]#

Get a batch of data from the iterator.

Parameters:

  • data_iterator – The data iterator to get the batch from.

  • use_mtp – Whether Multi-Token Prediction layers are enabled.

  • skip_getting_attention_mask_from_dataset – If set, the dataset will pass a None attention mask.

Returns:

A dictionary containing the batch data.

Return type:

dict[str, torch.Tensor]

bridge.training.vlm_step.get_batch(
data_iterator: Iterable,
cfg: megatron.bridge.training.config.ConfigContainer,
use_mtp: bool = False,
*,
pg_collection,
) → tuple[...]#

Generate a batch.

Parameters:

  • data_iterator – Input data iterator

  • cfg – Configuration container

  • use_mtp – Whether Multi-Token Prediction layers are enabled

Returns:

tuple of tensors containing tokens, labels, loss_mask, attention_mask, position_ids, packed sequence metadata, and visual_inputs.

bridge.training.vlm_step.forward_step(
state: megatron.bridge.training.state.GlobalState,
data_iterator: Iterable,
model: megatron.core.models.gpt.GPTModel,
return_schedule_plan: bool = False,
) → tuple[torch.Tensor, functools.partial]#

Forward training step.

Parameters:

  • state – Global state for the run

  • data_iterator – Input data iterator

  • model – The GPT Model

  • return_schedule_plan (bool) – Whether to return the schedule plan instead of the output tensor

Returns:

tuple containing the output tensor and the loss function