bridge.models.deepseek.deepseek_v3_bridge

Module Contents

Classes

DeepSeekV3Bridge

Megatron Bridge for DeepSeek-V3.

Data

__all__
_dequant_fp8_blockwise

API

bridge.models.deepseek.deepseek_v3_bridge.__all__

[‘DeepSeekV3Bridge’, ‘_dequant_fp8_blockwise’]

bridge.models.deepseek.deepseek_v3_bridge._dequant_fp8_blockwise

None

class bridge.models.deepseek.deepseek_v3_bridge.DeepSeekV3Bridge

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

Megatron Bridge for DeepSeek-V3.

static generate_pipeline_layout(

num_layers: int,

pp: int,

mtp_layers: int = 1,

) → list[list[str]]

Generate a pipeline-parallel layout for DeepSeek V3 conversion.

DeepSeek V3 has 61 decoder layers, so the model cannot use ordinary pipeline partitioning for the practical PP sizes needed to hold the full checkpoint. The conversion launcher calls this hook to distribute decoder layers unevenly while keeping embeddings on the first stage and MTP plus loss on the last stage.

Parameters:

• num_layers – Number of decoder layers.

• pp – Pipeline parallel size.

• mtp_layers – Number of MTP layers.

Returns:

A flexible pipeline layout with exactly pp stages.

provider_bridge(

hf_pretrained: megatron.bridge.models.hf_pretrained.causal_lm.PreTrainedCausalLM,

) → megatron.bridge.models.mla_provider.MLAModelProvider

classmethod megatron_to_hf_config(

provider: megatron.bridge.models.mla_provider.MLAModelProvider,

) → dict

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

maybe_modify_loaded_hf_weight(

hf_param: Union[str, dict[str, str]],

hf_state_dict: Mapping[str, torch.Tensor],

) → Union[torch.Tensor, dict[str, torch.Tensor]]

Load HF weights and dequantize FP8 tensors on the fly.

DeepSeek-V3 ships linear weights as float8_e4m3fn with per-block scale factors stored in <key>_scale_inv (128x128 blocks). The true bf16 weight is::

w_bf16 = fp8_weight.float() * scale_inv_block

Without this override the bridge would do a bare .to(bf16) cast in ColumnParallelMapping.hf_to_megatron (param_mapping.py:905), discarding the per-block scales — the resulting model produces random-looking logits.

static _maybe_dequantize_fp8(

weight: torch.Tensor,

param_name: str,

hf_state_dict: Mapping[str, torch.Tensor],

) → torch.Tensor

Dequantize weight if it is stored as FP8 with a matching *_scale_inv.

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]

Preserve source-required shared MTP aliases and rotary frequencies.