bridge.models.qwen_omni.qwen25_omni_bridge#

Module Contents#

Classes#

Qwen25OmniBridge

Megatron Bridge for Qwen2.5-Omni Conditional Generation.

API#

class bridge.models.qwen_omni.qwen25_omni_bridge.Qwen25OmniBridge#

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

Megatron Bridge for Qwen2.5-Omni Conditional Generation.

Handles conversion between HuggingFace Qwen2_5OmniForConditionalGeneration and Megatron-Core Qwen25OmniModel formats.

Key differences from Qwen3OmniMoeBridge:

  • Dense LLM (Qwen2), not MoE -> no router/expert mappings

  • QKV bias mappings (Qwen2 has attention bias)

  • No QK layernorm weight mappings

  • Vision: ReplicatedMapping for HF vision encoder (thinker.visual.**)

  • Audio: ReplicatedMapping for HF audio encoder (thinker.audio_model.** -> thinker.audio_tower.**)

  • LLM layer norms use mlp.linear_fc1.layer_norm_weight (not pre_mlp_layernorm)

provider_bridge(
hf_pretrained: megatron.bridge.models.hf_pretrained.vlm.PreTrainedVLM,
) → megatron.bridge.models.qwen_omni.qwen25_omni_provider.Qwen25OmniModelProvider#

Create a Qwen25OmniModelProvider from a HuggingFace pretrained model.

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

Return MegatronMappingRegistry containing parameter mappings for dense Qwen2.5 Omni models.