Shortcuts

lora_llama2

torchtune.models.llama2.lora_llama2(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, *, vocab_size: int, num_layers: int, num_heads: int, num_kv_heads: int, embed_dim: int, max_seq_len: int, intermediate_dim: Optional[int] = None, attn_dropout: float = 0.0, norm_eps: float = 1e-05, lora_rank: int, lora_alpha: float, lora_dropout: float = 0.0, use_dora: bool = False, quantize_base: bool = False) TransformerDecoder[source]

Return a version of Llama2 (an instance of TransformerDecoder()) with LoRA applied based on the passed in configuration.

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

  • apply_lora_to_output (bool) – whether to apply LoRA to the model’s final output projection. Default: False

  • vocab_size (int) – number of tokens in vocabulary.

  • num_layers (int) – number of layers in the transformer decoder.

  • num_heads (int) – number of query heads. For MHA this is also the number of heads for key and value

  • num_kv_heads (int) – number of key and value heads. User should ensure num_heads % num_kv_heads == 0. For standard MHA set num_kv_heads == num_heads, for GQA num_kv_heads < num_heads, and for MQA set num_kv_heads == 1.

  • embed_dim (int) – embedding dimension for self-attention

  • max_seq_len (int) – maximum sequence length the model will be run with, as used by KVCache()

  • attn_dropout (float) – dropout value passed onto scaled_dot_product_attention. Default: 0.0

  • intermediate_dim (Optional[int]) – intermediate dimension for MLP. If not specified, this is computed using scale_hidden_dim_for_mlp()

  • norm_eps (float) – epsilon in RMS norms.

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

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

  • lora_dropout (float) – LoRA dropout probability. Default: 0.0

  • use_dora (bool) – Decompose the LoRA weight into magnitude and direction, as introduced in “DoRA: Weight-Decomposed Low-Rank Adaptation” (https://arxiv.org/abs/2402.09353).

  • quantize_base – (bool): Whether to quantize base model weights or not. Only applied to base weights within linear layers LoRA is applied to. The final output linear projection is not supported for quantization currently.

Returns:

Instantiation of Llama2 model with LoRA applied to a subset of the attention projections in each layer.

Return type:

TransformerDecoder

Docs

Access comprehensive developer documentation for PyTorch

View Docs

Tutorials

Get in-depth tutorials for beginners and advanced developers

View Tutorials

Resources

Find development resources and get your questions answered

View Resources