o ̳iS7@s"ddlmZmZddlmZddlmZmZmZm Z ddl m Z ddl m Z ddlmZddlmZddlmZ d e fd d Zd3dedeedeed e fddZ       d4deedededededededed e fdd Zeed!d"Zd#e_d e fd$d%Z       d4deedededededededed e fd&d'Zeed!d"Zd(e_d e fd)d*Z       d4deedededededededed e fd+d,Z ee d!d"Z!d-e!_d e fd.d/Z"       d4deedededededededed e fd0d1Z#eed!d"Z$d2e$_d S)5)ListOptional)partial)llama2 lora_llama2llama2_classifierlora_llama2_classifier)TransformerDecoder)Llama2Tokenizer)LORA_ATTN_MODULES) _TemplateType)_get_prompt_templatereturnc CstdddddddddS)z Builder for creating a Llama2 model initialized w/ the default 7B parameter values from https://arxiv.org/abs/2307.09288 Returns: TransformerDecoder: Instantiation of Llama2 7B model } h㈵>) vocab_size num_layers num_heads num_kv_heads embed_dim max_seq_len attn_dropoutnorm_epsrrr[/home/ubuntu/.local/lib/python3.10/site-packages/torchtune/models/llama2/_model_builders.py llama2_7bsrN*torchtune.models.llama2.Llama2ChatTemplatepathrprompt_templatecCs"t|||dur t|dSddS)a Tokenizer for Llama2. Args: path (str): path to the tokenizer max_seq_len (Optional[int]): maximum sequence length for tokenizing a single list of messages, after which the input will be truncated. Default is None. prompt_template (Optional[_TemplateType]): optional specified prompt template. If a string, it is assumed to be the dotpath of a :class:`~torchtune.data.PromptTemplateInterface` class. If a dictionary, it is assumed to be a custom prompt template mapping role to the prepend/append tags. Default is :class:`~torchtune.models.llama2.Llama2ChatTemplate`. Returns: Llama2Tokenizer: Instantiation of the Llama2 tokenizer Nr!rr")r r r#rrrllama2_tokenizer-s"r$Frlora_attn_modulesapply_lora_to_mlpapply_lora_to_output lora_rank lora_alpha lora_dropoutuse_dora quantize_basecCsjtdid|d|d|ddddddd dd d d d d dddd|d|d|d|d|S)aJ Builder for creating a Llama2 7B model with LoRA enabled. The Llama2 defaults are the same as in :func:`~torchtune.models.llama2.llama2_7b`, while LoRA default params are based on https://github.com/tloen/alpaca-lora/blob/8bb8579e403dc78e37fe81ffbb253c413007323f/finetune.py#L41-L43. Args: 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 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 Returns: TransformerDecoder: Instantiation of Llama2 7B model with LoRA applied r'r(r)rrrrrrrrrrrrrr*r+r,r-r.Nrrr'r(r)r*r+r,r-r.rrrlora_llama2_7b@sB#     r1T)r.z Builder for creating a Llama2 7B model with QLoRA enabled. Base model weights in linear layers that LoRA is applied to are quantized per the QLoRA paper: https://arxiv.org/abs/2305.14314. Please see `lora_llama2_7b` for full API arguments. c Cstdddddddddd S) z Builder for creating a Llama2 model initialized w/ the default 13B parameter values from https://arxiv.org/abs/2307.09288 Returns: TransformerDecoder: Instantiation of Llama2 13B model r(6rrr rrrrrintermediate_dimrrrrrrrr llama2_13br7cCsptdid|d|d|ddddddd dd d d d ddddddd|d|d|d|d|S)a Builder for creating a Llama2 13B model with LoRA enabled. The Llama2 defaults are the same as in :func:`~torchtune.models.llama2.llama2_13b`, while LoRA default params are based on https://github.com/tloen/alpaca-lora/blob/8bb8579e403dc78e37fe81ffbb253c413007323f/finetune.py#L41-L43. Args: 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 lora_rank (int): rank of each low-rank approximation lora_alpha (float): scaling factor for the low-rank approximation 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 Returns: TransformerDecoder: Instantiation of Llama2 13B model with LoRA applied r'r(r)rrrr2rrrr3r6r4rrrrrrr*r+r,r-r.Nrr/r0rrrlora_llama2_13bF#     r9z Builder for creating a Llama2 13B model with QLoRA enabled. Base model weights in linear layers that LoRA is applied to are quantized per the QLoRA paper: https://arxiv.org/abs/2305.14314. Please see `lora_llama2_13b` for full API arguments. c Cstddddddddd d S) z Builder for creating a Llama2 model initialized w/ the default 70B parameter values from https://arxiv.org/abs/2307.09288 Returns: TransformerDecoder: Instantiation of Llama2 70B model rP@r% prrrr5rrrrr llama2_70br8r?cCsptdid|d|d|dddddd d d d d ddddddddd|d|d|d|d|S)aE Builder for creating a Llama2 70B model with LoRA enabled. The Llama2 defaults are the same as in :func:`~torchtune.models.llama2.llama2_70b`, while LoRA default params are based on https://github.com/tloen/alpaca-lora/blob/8bb8579e403dc78e37fe81ffbb253c413007323f/finetune.py#L41-L43. Args: 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 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 Returns: TransformerDecoder: Instantiation of Llama2 70B model with LoRA applied r'r(r)rrrr;rr<rr%rr=rrr6r>rrrrr*r+r,r-r.Nrr/r0rrrlora_llama2_70br:r@z Builder for creating a Llama2 70B model with QLoRA enabled. Base model weights in linear layers that LoRA is applied to are quantized per the QLoRA paper: https://arxiv.org/abs/2305.14314. Please see `lora_llama2_70b` for full API arguments. c Cstdddddddddd S)a: Builder for creating a Llama2 model initialized w/ the default 7B parameter values from https://arxiv.org/abs/2307.09288, where the output layer is a classification layer projecting to a single class for reward modelling. Returns: TransformerDecoder: Instantiation of Llama2 7B model rrrrr) num_classesrrrrrrrr)rrrrrllama2_reward_7b*s rCcCsptdid|d|d|dddddd d d d d d d dd ddddd|d|d|d|d|S)a Builder for creating a Llama2 7B reward model with LoRA enabled. The Llama2 classifier defaults are the same as in :func:`~torchtune.models.llama2.llama2_reward_7b`, while LoRA default params are based on https://github.com/tloen/alpaca-lora/blob/8bb8579e403dc78e37fe81ffbb253c413007323f/finetune.py#L41-L43. Args: 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 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 quantize_base (bool): Whether to quantize base model weights Returns: TransformerDecoder: Instantiation of Llama2 7B model with LoRA applied r'r(r)rBrArrrrrrrrrrrrrr*r+r,r-r.Nr)rr0rrrlora_llama2_reward_7b@sF!     rDz Builder for creating a Llama2 reward 7b model with QLoRA enabled. Base model weights in linear layers that LoRA is applied to are quantized per the QLoRA paper: https://arxiv.org/abs/2305.14314. Please see `lora_llama2_reward_7b` for full API arguments. )Nr )FFr%r&rFF)%typingrr functoolsr+torchtune.models.llama2._component_buildersrrrrtorchtune.modulesr "torchtune.models.llama2._tokenizerr torchtune.modules.peftr torchtune.data._prompt_templatesr r rstrintr$boolfloatr1qlora_llama2_7b__doc__r7r9qlora_llama2_13br?r@qlora_llama2_70brCrDqlora_llama2_reward_7brrrrs      $ 7 8 8 6