bridge.models.kimi_vl.kimi_k25_vl_bridge

Module Contents

Classes

KimiK25VLBridge

Megatron Bridge for Kimi K2.5 VL.

API

class bridge.models.kimi_vl.kimi_k25_vl_bridge.KimiK25VLBridge

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

Megatron Bridge for Kimi K2.5 VL.

Converts HuggingFace Kimi K2.5 VL models (KimiK25ForConditionalGeneration) to Megatron format (KimiK25VLModel) and vice versa.

The language backbone shares the same architecture as Kimi K2 (MoE with MLA).

provider_bridge(

hf_pretrained: megatron.bridge.models.hf_pretrained.vlm.PreTrainedVLM,

) → megatron.bridge.models.kimi_vl.kimi_k25_vl_provider.KimiK25VLModelProvider

_load_and_dequantize(

key: str,

hf_state_dict: Mapping[str, torch.Tensor],

) → torch.Tensor

Load a weight, dequantizing INT4 packed tensors when present.

maybe_modify_loaded_hf_weight(

hf_param: str | dict[str, str],

hf_state_dict: Mapping[str, torch.Tensor],

) → torch.Tensor

Load HF weights, dequantizing INT4 quantized tensors when present.

_is_quantized_expert_key(key: str) → bool

maybe_modify_converted_hf_weight(

task: megatron.bridge.models.conversion.model_bridge.WeightConversionTask,

converted_weights_dict: Dict[str, torch.Tensor],

hf_state_dict: Mapping[str, torch.Tensor],

) → Dict[str, torch.Tensor]

Re-quantize converted expert weights to INT4 format.

build_conversion_tasks(hf_pretrained, megatron_model) → List

Override to synthesize virtual weight keys from INT4 quantized triplets.

The HF checkpoint stores quantized expert weights as triplets (weight_packed, weight_scale, weight_shape) without a plain ‘weight’ key. We synthesize virtual ‘weight’ keys so the mapping registry can find them, then maybe_modify_loaded_hf_weight handles dequantization at load time.

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