PiKB(ddlmZmZddlmZddlmZddlmZm Z m Z m Z m Z m Z mZddlmZddlmZmZmZ d*d ed ed ededededededeedede fdZd+dedededefdZ d,d dd ddddddeededed ed ed ededededeedededed ed!ed"ed#edede f&d$Zdddddd%d&eeded ededededed ed!ed"eded#ede fd'Zdddd(deded ed!ed"eded#edefd)Zd S)-)ListOptional)nn)scale_hidden_dim_for_mlp) FeedForwardFrozenNF4LinearMultiHeadAttentionRMSNormRotaryPositionalEmbeddingsTransformerDecoderTransformerSelfAttentionLayer)(_register_reparametrize_state_dict_hooks) DoRALinearLORA_ATTN_MODULES LoRALinear Nh㈵> vocab_size num_layers num_heads num_kv_heads embed_dim max_seq_len attn_dropout rope_baseintermediate_dimnorm_epsreturnc ||z} |r|n|}|r|nt|} t| ||} tj} t |D]}t |||| tj||| zdtj||| zdtj||| zdtj||d| || }t|| }t||t|| t|| }| |tj ||}tj||d}t|| ||| t|| | S) a Build the decoder associated with the Llama3 model. This includes: - Token embeddings - num_layers number of TransformerSelfAttentionLayer blocks - RMS Norm layer applied to the output of the transformer - Final projection into token space Args: 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 :func:`~torchtune.modules.KVCache` attn_dropout (float): dropout value passed onto scaled_dot_product_attention. Default: 0.0 rope_base (int): base for the rotary positional embeddings. Default: 500_000 intermediate_dim (Optional[int]): intermediate dimension for MLP. If not specified, this is computed using :func:`~torchtune.modules.scale_hidden_dim_for_mlp` norm_eps (float): epsilon in RMS norms. Returns: TransformerDecoder: Instantiation of Llama3 model. dimrbaseFbias rrrhead_dimq_projk_projv_proj output_projpos_embeddingsrr)r" hidden_dimr"epsattnmlpsa_normmlp_normr/tok_embeddingslayersrrr'normoutput) rr r ModuleListranger Linear llama3_mlpr r append Embeddingr )rrrrrrrrrrr'r-roper8_ self_attnr2layerr7r+s /home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/torchtune/models/llama3/_component_builders.pyllama3rF+sPI%H#/><>>- x888999      e\*i88N)Iz>>>K % YH - - -   Fr"r- quantize_basec(|stj||dnt||d}|stj||dnt||d}|stj||dnt||d}t|||S)z> Build the MLP layer associated with the Llama model. Fr$ gate_proj down_projup_proj)rr=rr)r"r-rHrKrLrMs rEr>r>s  : #z.... S*5 9 9 9 : *c.... Z5 9 9 9 : #z.... S*5 9 9 9 i Q Q QQrG)rrrr lora_dropoutuse_dorarHlora_attn_modulesapply_lora_to_mlpapply_lora_to_output lora_rank lora_alpharNrOc| r| nt|}tj}t|D]}t |||||| | | |||| }|rt ||| ||||}nt |||}t||t|| t|| }| |tj ||}|rtnt}|r|||| ||ntj ||d}t||||||zt|| | }|rt||S) a Return a version of Llama3 (an instance of :func:`~torchtune.modules.TransformerDecoder`) with LoRA applied based on the passed in configuration. 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 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 :func:`~torchtune.modules.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 :func:`~torchtune.modules.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: TransformerDecoder: Instantiation of Llama3 model with LoRA applied to a subset of the attention projections in each layer. ) lora_modulesrrrrrrrSrTrNrHrO)r"r-rSrTrHrNrO)r"r-rHr.r0)rankalphadropoutFr$r5r6)rrr;r<lora_llama3_self_attentionlora_llama3_mlpr>r r r?r@rrr=r r)rPrQrRrrrrrrrrrrrSrTrNrOrHr-r8rBrCr2rDr7 adapter_clsr+models rE lora_llama3r^sD-U2J92U2U]__F :  %%.*%#%!%'      !%#%+)!CC*MC. x888999      e\*i88N!)8**jK  :       Yy*5 9 9 9 %y( YH - - -   E8 1777 LrG)rrrNrHrOrVc |stdtd||z} |r|n|}| rtnt} d|vr| ||| z||| | n0| st j||| zdnt ||| zd}d|vr| ||| z||| | n0| st j||| zdnt ||| zd}d|vr| ||| z||| | n0| st j||| zdnt ||| zd}d |vr| ||||| | n*| st j||dnt ||d}t| || }t|||| ||||||| }|S) a Return an instance of :func:`~torchtune.modules.MultiHeadAttention` with LoRA applied to a subset of its linear layers Args: lora_modules (List[LORA_ATTN_MODULES]): list of which linear layers LoRA should be applied to. Options are ``{"q_proj", "k_proj", "v_proj", "output_proj"}``. embed_dim (int): embedding dimension for self-attention 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. max_seq_len (int): maximum sequence length the model will be run with, as used by :func:`~torchtune.modules.KVCache` attn_dropout (float): dropout value passed onto scaled_dot_product_attention. Default: 0.0 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 parameters for linear layers LoRA is being applied to. Default is ``False``. 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). Returns: MultiHeadAttention: instantiation of self-attention module with LoRA applied to a subset of Q, K, V, output projections. Raises: ValueError: If lora_modules arg is an empty list zMust pass one or more of z as lora_modulesr()rWrXrYrHFr$r)r*r+r!r&) ValueErrorrrrrr=rr r )rVrrrrrrrSrTrNrHrOr'r\r(r)r*r+rArCs rErZrZ(sd   K(9 K K K   I%H#/><r^rZr[rGrErns("!!!!!!!IIIIIITSSSSSLLLLLLLLLL .&*SSSSS S  S  SSSsmSSSSSSlRRCRSRR+RRRR6$!&K'+-KKK-.KKK  KKKKKKsmKKKK"#K$%K&'K()K,-K./KKKKlJJJ()J J  J  JJJJJJJJJ !JJJJf''' '' '  '  '''''''''rG