o ̳i@s^ddlmZmZddlmZmZmZmZddlm Z ddlm Z ddl m Z ddl mZddlmZddlmZ d e fd d Zd)dedeedee d efddZ       d*deedededededededed e fdd Zeed!d"Zd#e_d e fd$d%Z       d*deedededededededed e fd&d'Zeed!d"Zd(e_d S)+)ListOptional)mistral lora_mistralmistral_classifierlora_mistral_classifier) _TemplateType)_get_prompt_template)TransformerDecoder)MistralTokenizer)LORA_ATTN_MODULES)partialreturnc Cstdddddddddd S) z Builder for creating a Mistral 7B model initialized w/ the default 7b parameter values from https://mistral.ai/news/announcing-mistral-7b/ Returns: TransformerDecoder: Instantiation of Mistral 7B model } 8h㈵>) vocab_size num_layers num_heads num_kv_heads embed_dimintermediate_dim max_seq_len attn_dropoutnorm_eps)rr r \/home/ubuntu/.local/lib/python3.10/site-packages/torchtune/models/mistral/_model_builders.py mistral_7bs r"N,torchtune.models.mistral.MistralChatTemplatepathrprompt_templatecCs"t|||dur t|dSddS)a Tokenizer for Mistral models. Args: path (str): path to the tokenizer max_seq_len (Optional[int]): maximum sequence length for tokenizing a single list of messages, after which the input will be truncated. Default is None. prompt_template (Optional[_TemplateType]): optional specified prompt template. If a string, it is assumed to be the dotpath of a :class:`~torchtune.data.PromptTemplateInterface` class. If a dictionary, it is assumed to be a custom prompt template mapping role to the prepend/append tags. Default is :class:`~torchtune.models.mistral.MistralChatTemplate`. Returns: MistralTokenizer: Instantiation of the Mistral tokenizer Nr$rr%)r r r&r r r!mistral_tokenizer3s"r'Frrlora_attn_modulesapply_lora_to_mlpapply_lora_to_output lora_rank lora_alpha lora_dropoutuse_dora quantize_basecCsvtdid|d|d|ddddddd d d d d dddddddddd|d|d|d|d|S)ao Builder for creating a Mistral 7B model with LoRA enabled. 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): dropout probability for the low-rank approximation. 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 Mistral 7B model with LoRA applied r)r*r+rrrrrrrrrrrrrrrrr rope_base'r,r-r.r/r0Nr )rr)r*r+r,r-r.r/r0r r r!lora_mistral_7bFsJ     r4T)r0z Builder for creating a Mistral 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_mistral_7b` for full API arguments. c Cstdddddddddd d S) a~ Builder for creating a Mistral 7B model initialized w/ the default 7b parameter values from: https://huggingface.co/Ray2333/reward-model-Mistral-7B-instruct-Unified-Feedback where the output layer is a classification layer projecting to a single class for reward modelling. Returns: TransformerDecoder: Instantiation of Mistral 7B classifier model rrrrrrrr) num_classesrrrrrrrrr)rr r r r!mistral_reward_7bs r7cCs|tdid|d|d|dddddd d d d d d dddddddddddd|d|d|d|d|S) av Builder for creating a Mistral reward 7B model with LoRA enabled. 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): dropout probability for the low-rank approximation. 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 Mistral 7B model with LoRA applied r)r*r+r6r5rrrrrrrrrrrrrrrrrr1r2r,r-r.r/r0Nr )rr3r r r!lora_mistral_reward_7bsN     r8a Builder for creating a Mistral 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_mistral_reward_7b` for full API arguments. )Nr#)FFrr(rFF)typingrr,torchtune.models.mistral._component_buildersrrrr torchtune.data._prompt_templatesrr torchtune.modulesr #torchtune.models.mistral._tokenizerr torchtune.modules.peftr functoolsr r"strintr'boolfloatr4qlora_mistral_7b__doc__r7r8qlora_mistral_reward_7br r r r!s      $ 5 6