bridge.models.glm.glm_moe_mappings#
GLM MoE mapping helpers for fused expert weights.
These are thin aliases around the shared FusedExpertMapping / FusedGatedExpertMapping classes in param_mapping.py. Kept for backwards compatibility with existing imports.
Module Contents#
Classes#
FusedExpertMapping for GLM down-projection expert weights. |
API#
- class bridge.models.glm.glm_moe_mappings.GLMExpertDownProjMapping(
- megatron_param: str,
- hf_param: str,
- permute_dims: tuple[int, ...] | None = None,
- transpose_on_export: bool = False,
Bases:
megatron.bridge.models.conversion.param_mapping.FusedExpertMappingFusedExpertMapping for GLM down-projection expert weights.
Grouped-export transpose is disabled for the current GLM fused-expert export paths. The per-expert tensors are already in HF layout by the time they are stacked into the fused expert weight, so applying an additional grouped- export transpose would produce the wrong orientation.
Initialization