bridge.models.hy_v3.hy_v3_bridge#

Module Contents#

Classes#

HYV3Bridge

Megatron Bridge for Hy V3 MoE Causal LM.

Data#

__all__

API#

bridge.models.hy_v3.hy_v3_bridge.__all__#

[‘HYV3Bridge’]

class bridge.models.hy_v3.hy_v3_bridge.HYV3Bridge#

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

Megatron Bridge for Hy V3 MoE Causal LM.

This bridge handles the conversion between HuggingFace HYV3ForCausalLM and Megatron-Core GPTModel formats. Hy V3 is a fine-grained MoE decoder with GQA + QK layernorm, dense-first layer freq, sigmoid router with per-expert bias, shared experts, and grouped-GEMM routed experts.

.. rubric:: Example

from megatron.bridge import AutoBridge bridge = AutoBridge.from_hf_pretrained(“tencent/Hy3-preview-Base”) model_config = bridge.get_model_config()

MODEL_CONFIG_CLASS#

None

USE_MODEL_CONFIG_FOR_CONVERSION#

True

hf_config_to_model_config_kwargs(
hf_config: transformers.PretrainedConfig,
) → dict[str, Any]#

Convert a Hugging Face Hy V3 config to builder config kwargs.

The builder’s default MoE layer spec already uses the GPT decoder block spec, which honors the dense-first moe_layer_freq.

hf_config_to_provider_kwargs(
hf_config: transformers.PretrainedConfig,
) → dict[str, Any]#

Adapt the canonical builder mapping to the deprecated provider path.

classmethod megatron_to_hf_config(
provider: megatron.bridge.models.gpt_provider.GPTModelProvider,
) → dict#

Convert Megatron provider config to HuggingFace HYV3Config dict.

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