Pi+XddlZddlmZmZmZddlZddddddd d d d d dd Zddddd dd dd d d ddd ZdedeeefdefdZ deeej fdeeej ffdZ deeej fdeeej ffdZ d2deeej fde de de de deeej ff dZ d2deeej fde de de de f dZd d!d"d#Zd$d%d&d'd(d)d*d+d,Zgd-Z d3d.eeefd/eefd0Z d2deeej fde de de dee f d1ZdS)4N)AnyDictOptionaltok_embeddings.weightz norm.scale output.weightzlayers.{}.attn.k_proj.weightzlayers.{}.attn.q_proj.weightzlayers.{}.attn.v_proj.weightz!layers.{}.attn.output_proj.weightzlayers.{}.sa_norm.scalezlayers.{}.mlp_norm.scalezlayers.{}.mlp.w1.weightzlayers.{}.mlp.w2.weightzlayers.{}.mlp.w3.weight) rz norm.weightrzlayers.{}.attention.wk.weightzlayers.{}.attention.wq.weightzlayers.{}.attention.wv.weightzlayers.{}.attention.wo.weightzlayers.{}.attention_norm.weightzlayers.{}.ffn_norm.weightz layers.{}.feed_forward.w1.weightz layers.{}.feed_forward.w2.weightz layers.{}.feed_forward.w3.weight) zmodel.embed_tokens.weightz'model.layers.{}.self_attn.q_proj.weightz'model.layers.{}.self_attn.k_proj.weightz'model.layers.{}.self_attn.v_proj.weightz'model.layers.{}.self_attn.o_proj.weightz-model.layers.{}.self_attn.rotary_emb.inv_freqz$model.layers.{}.mlp.gate_proj.weightz"model.layers.{}.mlp.up_proj.weightz$model.layers.{}.mlp.down_proj.weightz&model.layers.{}.input_layernorm.weightz/model.layers.{}.post_attention_layernorm.weightzmodel.norm.weightzlm_head.weightkey mapping_dictreturncv td|dDr\tjdd|}tjd|d}||}||}n||}n&#t$r}td|d|d}~wwxYw|S) Nc3>K|]}|VdSN)isdigit).0ks t/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/torchtune/models/convert_weights.py z!get_mapped_key..2s*33qqyy{{333333.z(\.\d+)z.{}z\d+rz8Error converting the state dict. Found unexpected key: "zG". Please make sure you're loading a checkpoint with the right format. ) anysplitresubsearchgroupformatKeyError Exception)rr abstract_key layer_numnew_keyes rget_mapped_keyr"/s 33CIIcNN333 3 3 (6*eS99L &#..44Q77I"<0GnnY//GG"3'G  Ss S S S    NsBB B6B11B6 state_dictczi}|D]#\}}|dvrt|t}|||<$|S)a Convert a state dict from Meta's format to torchtune's format. State dicts from multiple checkpoint files should be consolidated into a single state dict before calling this function. Eg of Meta-format state dict can be found in the ``meta-llama/Llama-2-7b`` repo in HF (https://huggingface.co/meta-llama/Llama-2-7b). Args: state_dict (Dict[str, torch.Tensor]): State dict in Meta's format. Returns: Dict[str, torch.Tensor]: State dict in torchtune's format. )z rope.freqs)itemsr" _FROM_META)r#converted_state_dictrvaluer s r meta_to_tuner)CsU &&((22 U n $ $$S*55G,1  ) rci}dtD}|D]\}}t||}|||<|S)av Convert a state dict from torchtune's format to Meta's format. This function doesn't handle any sharding or splitting of state dicts. It follows the state_dict IN -> state_dict OUT pattern. Args: state_dict (Dict[str, torch.Tensor]): State dict in torchtune's format. Returns: Dict[str, torch.Tensor]: State dict in Meta's format. ci|]\}}|| Sr,rrvs r z tune_to_meta..hsAAAdaQAAAr)r&r%r")r#r'inverted_mapping_dictrr(r s r tune_to_metar1[smAAj.>.>.@.@AAA &&((.. U &;<<(-W%% r  num_heads num_kv_headsdimhead_dimci}|zfd}|D]D\}}d|vr;t|t} d|vr |||}nd|vr |||}||| <E|S)a( Convert a state dict from HF's format to torchtune's format. State dicts from multiple checkpoint files should be consolidated into a single state dict before calling this function. Eg of HF-format state dict can be found in the ``meta-llama/Llama-2-7b-hf`` repo in HF (https://huggingface.co/meta-llama/Llama-2-7b-hf). Args: state_dict (Dict[str, torch.Tensor]): State dict in HF's format. num_heads (int): Number of heads in the model. num_kv_heads (int): Number of heads in the key/value projection layers. dim (int): Dimension of the model. head_dim (int): Dimension of the head. If not provided, it will be calculated as dim // num_heads. Returns: Dict[str, torch.Tensor]: State dict in torchtune's format. Nc||ddzdd|zSNview transposereshapetn_headsr6r7s r_permutezhf_to_tune.._permutesB FF7Ax1}c 2 2 Yq!__ Wh(3 / / rzrotary_emb.inv_freqq_projk_proj)r%r"_FROM_HF) r#r4r5r6r7r'rDrr(r s `` r hf_to_tunerHqs4)#      !&&((22 U  + +$S(33G3  22S  55,1  ) rci}dtD}|zfd}|D];\}} t||} d|vr || |} nd|vr || |} | || <<|S)an Convert a state dict from torchtune's format to HF's format. This function doesn't handle any sharding or splitting of state dicts. It follows the state_dict IN -> state_dict OUT pattern. Args: state_dict (Dict[str, torch.Tensor]): State dict in torchtune's format. num_heads (int): Number of heads in the model. num_kv_heads (int): Number of heads in the key/value projection layers. dim (int): Dimension of the model. head_dim (int): Dimension of model attention heads. Default None. Returns: Dict[str, torch.Tensor]: State dict in HF's format. ci|]\}}|| Sr,r,r-s rr/ztune_to_hf..s???daQ???rNc||dzddd|zSr:r=rAs rrDztune_to_hf.._permutesB FF7HM1c 2 2 Yq!__ Wh(3 / / rrErF)rGr%r") r#r4r5r6r7r'r0rDrr(r s `` r tune_to_hfrLs,??hnn.>.>???)#      !&&((.. U &;<< s??HUI..EE __HUL11E(-W%% rlora_Alora_Blora_magnitude_vector)lora_alora_b magnituderErFv_projo_proj gate_proj down_projup_projlm_head)rErFrS output_projw1w2w3output)target_modulesr lora_alphaadapter_configbase_model_name_or_pathcZtfdtDs,tdtddD]}|tvrtd|t t tjdd<|r|d<S)Nc<g|]}|vSr,)keys)rxras r z/tune_to_peft_adapter_config..s*OOOq^((***OOOrzPEFT adapter config requires z, found r^zUnknown target module rb)all_PEFT_CONFIG_EXPECTED_KEYS ValueErrorre_TO_PEFT_TARGET_MODULESlistmapget)rarbrs` rtune_to_peft_adapter_configros OOOO4NOOO P P  g,F g gP^PcPcPePe g g   , -;; + + +9a99:: : ,'+ # '8H)IJJ((N#$ L4K01 rchi}i}tD]\}}tD]\} } | |dkr3| dd|} | dd|} n4| dd|d} | dd|d} || | i||zfd} |D]F\}}t ||}d|vrd|vr | ||}nd|vrd|vr | ||}||d|z<G|S) NrRz.weightrc|jd}||dzd|dd|z|S)Nr;r<)shaper>r?r@)rBrCrankr7s r_permute_lora_matrixz:tune_to_peft_adapter_weights.._permute_lora_matrixsMwr{ FF7HM1d 3 3 Yq!__ Wh(4 0 0 rrErNrFzbase_model.model.) _TO_PEFT_KEYSr%rGreplaceupdater")r#r4r5r6r7r' full_mappingpeft_keypeft_valhf_keyhf_val adapter_key adapter_valrurr(r s ` rtune_to_peft_adapter_weightsrsL,1133<<(&nn.. < C07:;; r)r2r2r3Nr )rtypingrrrtorchr&rGstrr"Tensorr)r1intrHrLrvrkrirorr,rrrs &&&&&&&&&& 5$%C%C%C%H'@!;(A(A(A   ""9/M/M/M/R59,E*C,E.G7Q%%  "4S>c( T#u|"34 c5<>O9P    0 T#u|"34 c5<>O9P    0 . . S%,&'. . .  .  .  #u|  . . . . f + + S%,&'+ + +  +  + + + + `(      CBB .2cN%c]2" . . S%,&'. . .  . sm . . . . . . r