`bridge.models.glm_vl.glm_45v_bridge`#

Module Contents#

Classes#

Megatron Bridge for GLM 4.5 Vision-Language (VL) Models.

API#

class bridge.models.glm_vl.glm_45v_bridge.GLM45VBridge#

Bases: megatron.bridge.models.conversion.model_bridge.MegatronModelBridge

Megatron Bridge for GLM 4.5 Vision-Language (VL) Models.

provider_bridge( hf_pretrained: megatron.bridge.models.hf_pretrained.vlm.PreTrainedVLM, ) → megatron.bridge.models.glm_vl.glm_45v_provider.GLM45VModelProvider#

build_conversion_tasks(hf_pretrained, megatron_model)#: Override to store config before mapping_registry is called.

classmethod get_hf_tokenizer_kwargs() → dict#

Return HuggingFace tokenizer kwargs specific to GLM 4.5V models.

GLM 4.5V requires use_fast=True to properly load the tokenizer.

mapping_registry() → megatron.bridge.models.conversion.mapping_registry.MegatronMappingRegistry#

_uses_fused_experts() → bool#

_hf_expert_suffix(base_name: str) → str#

maybe_modify_converted_hf_weight( task, converted_weights_dict: Dict[str, torch.Tensor], hf_state_dict, ) → Dict[str, torch.Tensor]#