nemo_automodel.components.models.deepseek_v32.state_dict_adapter

State dict adapter for DeepSeek V3.2.

Extends DeepSeekV3StateDictAdapter with mappings for the new Indexer weights.

Module Contents

Classes

DeepSeekV32StateDictAdapter

State dict adapter for DeepSeek V3.2.

API

class nemo_automodel.components.models.deepseek_v32.state_dict_adapter.DeepSeekV32StateDictAdapter(

config: transformers.DeepseekV3Config,

moe_config: nemo_automodel.components.moe.config.MoEConfig,

backend: nemo_automodel.components.models.common.BackendConfig,

dtype: torch.dtype = torch.float32,

)

Bases: nemo_automodel.components.models.deepseek_v3.state_dict_adapter.DeepSeekV3StateDictAdapter

State dict adapter for DeepSeek V3.2.

Initialization

_base_non_quantized_keys

[‘input_layernorm.weight’, ‘post_attention_layernorm.weight’, ‘norm.weight’, ‘lm_head.weight’, ‘embe…

_indexer_non_quantized_keys

[‘indexer.k_norm.weight’, ‘indexer.k_norm.bias’, ‘indexer.weights_proj.weight’]

property _non_quantized_keys: list[str]

Get the full list of non-quantized keys including indexer keys.

_add_quantization_scale_inv_tensors(

state_dict: dict[str, Any],

) → dict[str, Any]

Add quantization scale tensors, handling indexer-specific keys.

convert_single_tensor_to_hf(

fqn: str,

tensor: Any,

**kwargs,

) → list[tuple[str, Any]]

Convert a single tensor from native format to HuggingFace format.

Handles both standard V3 tensors and V3.2 indexer tensors, ensuring indexer LayerNorm weights are not quantized.