lora_qwen2_5_3b

torchtune.models.qwen2_5.lora_qwen2_5_3b(lora_attn_modules: List[Literal['q_proj', 'k_proj', 'v_proj', 'output_proj']], apply_lora_to_mlp: bool = False, apply_lora_to_output: bool = False, lora_rank: int = 8, lora_alpha: float = 16, lora_dropout: float = 0.0, use_dora: bool = False, quantize_base: bool = False) → TransformerDecoder

Builder for creating a Qwen2.5 3B model (base or instruct) with LoRA enabled.

The Qwen2.5 defaults are the same as in qwen2_5_3b(), while LoRA default params are based on https://github.com/tloen/alpaca-lora/blob/8bb8579e403dc78e37fe81ffbb253c413007323f/finetune.py#L41-L43.

Parameters:

• lora_attn_modules (List[LORA_ATTN_MODULES]) – list of which linear layers LoRA should be applied to in each self-attention block. Options are {"q_proj", "k_proj", "v_proj", "output_proj"}.

• apply_lora_to_mlp (bool) – whether to apply LoRA to the MLP in each transformer layer. Default: False

• lora_rank (int) – rank of each low-rank approximation

• lora_alpha (float) – scaling factor for the low-rank approximation

• lora_dropout (float) – dropout probability for the low-rank approximation. Default: 0.0

• quantize_base (bool) – Whether to quantize base model weights

Returns:

Instantiation of Qwen2.5 3B model with LoRA applied

Return type:

TransformerDecoder

Note

Qwen2.5 0.5B-3B model builders will enable tie_word_embeddings by default (see qwen2())