PiddlmZddlZddlmZ ddddddd d d d d dddddZ ddeeejfdededededeeejff dZ ddeeejfdedededef dZ dS))DictN)get_mapped_keyztok_embeddings.weightzlayers.{}.attn.q_proj.weightzlayers.{}.attn.k_proj.weightzlayers.{}.attn.v_proj.weightz!layers.{}.attn.output_proj.weightzlayers.{}.mlp.w1.weightzlayers.{}.mlp.w3.weightzlayers.{}.mlp.w2.weightzlayers.{}.sa_norm.scalezlayers.{}.sa_scale.scalezlayers.{}.mlp_norm.scalezlayers.{}.mlp_scale.scaleznorm.rms_norm.scalez output.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.weightz1model.layers.{}.post_feedforward_layernorm.weightz0model.layers.{}.pre_feedforward_layernorm.weightzmodel.norm.weightzlm_head.weight  state_dict num_heads num_kv_headsdimhead_dimreturnci}|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_headsr r s |/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/torchtune/models/gemma2/_convert_weights.py_permutez#gemma2_hf_to_tune.._permuteFsB FF7Ax1}c 2 2 Yq!__ Wh(3 / / zrotary_emb.inv_freqq_projk_proj)itemsr_GEMMA2_FROM_HF) rrr r r converted_state_dictrkeyvaluenew_keys `` rgemma2_hf_to_tuner$(s4)#      !&&((22 U  + +$S/::G3  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').0kvs r z%gemma2_tune_to_hf..psFFFdaQFFFrNc||dzddd|zSrrrs rrz#gemma2_tune_to_hf.._permuteusB FF7HM1c 2 2 Yq!__ Wh(3 / / rrr)rrr) rrr r r r inverted_mapping_dictrr!r"r#s `` rgemma2_tune_to_hfr.Ys,FFo.C.C.E.EFFF)#      !&&((.. U &;<< s??HUI..EE __HUL11E(-W%% r)rrrN) typingrtorch torchtune.models.convert_weightsrrstrTensorintr$r.r'rrr5sW ;;;;;;"9/M/M/M/R59,E*C,E.G7Q9S8S.%* . . S%,&'. . .  .  .  #u|  . . . . f + + S%,&'+ + +  +  + + + + + + r