bridge.models.glm.glm47_flash_bridge#
Megatron Bridge for GLM-4.7-Flash (Glm4MoeLiteForCausalLM).
This module registers only the glm4_moe_lite / Flash variant. The full
GLM-4.7 model uses the existing GLM45Bridge registration for
Glm4MoeForCausalLM / glm4_moe.
GLM-4.7-Flash combines Multi-Latent Attention (MLA, inherited from DeepSeek V3)
with GLM-style MoE routing. The safetensors checkpoint uses per-expert weight
naming (experts.{i}.gate_proj), not the fused gate_up_proj tensor used
by the runtime model, so the DeepSeek common mapping list works directly.
Module Contents#
Classes#
Megatron Bridge for GLM-4.7-Flash. |
API#
- class bridge.models.glm.glm47_flash_bridge.GLM47FlashBridge#
Bases:
megatron.bridge.models.conversion.model_bridge.MegatronModelBridgeMegatron Bridge for GLM-4.7-Flash.
GLM-4.7-Flash uses Multi-Latent Attention (MLA) with MoE routing, combining DeepSeek V3-style compressed attention with GLM MoE architecture.
.. rubric:: Example
from megatron.bridge import AutoBridge bridge = AutoBridge.from_hf_pretrained(“zai-org/GLM-4.7-Flash”) provider = bridge.to_megatron_provider()
- provider_bridge(
- hf_pretrained: megatron.bridge.models.hf_pretrained.causal_lm.PreTrainedCausalLM,
Convert HuggingFace config to MLAModelProvider.
- mapping_registry() megatron.bridge.models.conversion.mapping_registry.MegatronMappingRegistry#