`bridge.models.conversion.gtp`#

Translate between stored GTP shards and the TP-local tensors used by mappings.

Module Contents#

Functions#

`_is_gtp_param`
`_get_mapping_shape`	Return the logical TP-local shape, excluding GTP padding and sharding.
`_gtp_weight_load_context`	Let TE update native FP8 storage while preserving the GTP parameter subclass.
`_slice_gtp_weight`	Take the stored GTP row shard after the mapping has applied TP splitting.
`_gather_gtp_weight`	Gather only the GTP axis, leaving TP/EP/PP handling to the mapping.

API#

bridge.models.conversion.gtp._is_gtp_param(param: torch.Tensor | None) → bool#

bridge.models.conversion.gtp._get_mapping_shape(param: torch.Tensor) → torch.Size#: Return the logical TP-local shape, excluding GTP padding and sharding.

bridge.models.conversion.gtp._gtp_weight_load_context( param: torch.Tensor, module: torch.nn.Module, ) → collections.abc.Iterator[None]#: Let TE update native FP8 storage while preserving the GTP parameter subclass.

bridge.models.conversion.gtp._slice_gtp_weight( weight: torch.Tensor, param: torch.Tensor, ) → torch.Tensor#: Take the stored GTP row shard after the mapping has applied TP splitting.

bridge.models.conversion.gtp._gather_gtp_weight(param: torch.Tensor | None) → torch.Tensor | None#: Gather only the GTP axis, leaving TP/EP/PP handling to the mapping.