bridge.models.qwen3_asr.qwen3_asr_bridge#

Module Contents#

Classes#

Qwen3ASRBridge

Megatron Bridge for Qwen3-ASR Conditional Generation.

Functions#

_get_thinker_config

Normalize the legacy checkpoint layout without replacing native HF registrations.

API#

bridge.models.qwen3_asr.qwen3_asr_bridge._get_thinker_config(
hf_config: transformers.PretrainedConfig,
) → megatron.bridge.models.qwen3_asr.hf_qwen3_asr.configuration_qwen3_asr.Qwen3ASRThinkerConfig#

Normalize the legacy checkpoint layout without replacing native HF registrations.

class bridge.models.qwen3_asr.qwen3_asr_bridge.Qwen3ASRBridge#

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

Megatron Bridge for Qwen3-ASR Conditional Generation.

Handles conversion between HuggingFace Qwen3ASRForConditionalGeneration and Megatron-Core Qwen3ASRModel formats.

Key differences from Qwen25OmniBridge:

  • Qwen3-based LLM (not Qwen2): no QKV bias, has QK layernorm

  • Audio-only: no vision/video mappings

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

  • QK layernorm weight mappings (q_layernorm, k_layernorm)

provider_bridge(
hf_pretrained: megatron.bridge.models.hf_pretrained.causal_lm.PreTrainedCausalLM,
) → megatron.bridge.models.qwen3_asr.qwen3_asr_provider.Qwen3ASRModelProvider#

Create a Qwen3ASRModelProvider from a HuggingFace pretrained model.

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

Return MegatronMappingRegistry containing parameter mappings for Qwen3-ASR models.