Pi8ddlmZddlZddlmZidddddd d d d d dddddddddddddddddddd d!d"Zd!Z d/d&eeejfd'e d(e d)e d*e d+e d,eeejffd-Z d/d&eeejfd'e d(e d)e d*e d+e f d.Z dS)0)DictN)get_mapped_keyzmodel.embed_tokens.weightztok_embeddings.weightz'model.layers.{}.self_attn.q_proj.weightzlayers.{}.attn.q_proj.weightz%model.layers.{}.self_attn.q_proj.biaszlayers.{}.attn.q_proj.biasz'model.layers.{}.self_attn.k_proj.weightzlayers.{}.attn.k_proj.weightz%model.layers.{}.self_attn.k_proj.biaszlayers.{}.attn.k_proj.biasz'model.layers.{}.self_attn.v_proj.weightzlayers.{}.attn.v_proj.weightz%model.layers.{}.self_attn.v_proj.biaszlayers.{}.attn.v_proj.biasz'model.layers.{}.self_attn.o_proj.weightz!layers.{}.attn.output_proj.weightz-model.layers.{}.self_attn.rotary_emb.inv_freqz$model.layers.{}.mlp.gate_proj.weightzlayers.{}.mlp.w1.weightz"model.layers.{}.mlp.up_proj.weightzlayers.{}.mlp.w3.weightz$model.layers.{}.mlp.down_proj.weightzlayers.{}.mlp.w2.weightz&model.layers.{}.input_layernorm.weightzlayers.{}.sa_norm.scalez/model.layers.{}.post_attention_layernorm.weightzlayers.{}.mlp_norm.scalezmodel.norm.weightz norm.scalezlm_head.weightz output.weight F state_dict num_heads num_kv_headsdimhead_dimtie_word_embeddingsreturnci}|||z}|D]0\}}|r t|vrd|vrt|t} ||| <1|S)a Convert a state dict from HF's format to TorchTune's format, which contains the weights of a Qwen2 model. State dicts from multiple checkpoint files should be consolidated into a single state dict before calling this function. The logic is identical to :func:`~torchtune.models.convert_weights.hf_to_tune`, but may not load output projection weights. 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. tie_word_embeddings (bool): Whether the model's input and output word embeddings should be tied. Returns: Dict[str, torch.Tensor]: State dict in torchtune's format. Nzrotary_emb.inv_freq)itemsQWEN2_TIED_KEYr_FROM_HF) rrr r r r converted_state_dictkeyvaluenew_keys {/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/torchtune/models/qwen2/_convert_weights.pyqwen2_hf_to_tuner%s8)# &&(( . . U  $2c$9$9  C ' '  h//(-W%% ci}dtD}|||z}|D]\}} t||} | || <|S)a Convert a state dict from torchtune's format to HF's format. This function doesn't handle any sharding or splitting of state dicts. It follows the state_dict IN -> state_dict OUT pattern. Args: state_dict (Dict[str, torch.Tensor]): State dict in torchtune's format. num_heads (int): Number of heads in the model. num_kv_heads (int): Number of heads in the key/value projection layers. dim (int): Dimension of the model. head_dim (int): Dimension of the head. If not provided, it will be calculated as dim // num_heads. tie_word_embeddings (bool): Whether the model's input and output word embeddings should be tied. Returns: Dict[str, torch.Tensor]: State dict in HF's format. ci|]\}}|| Sr).0kvs r z$qwen2_tune_to_hf..ls???daQ???r)rrr) rrr r r r rinverted_mapping_dictrrrs rqwen2_tune_to_hfr!Rs{2??hnn.>.>???)# &&((.. U &;<<(-W%% r)rrrNF) typingrtorch torchtune.models.convert_weightsrrrstrTensorintboolrr!rrrr)s ;;;;;; !8 -/M ,-I ./M   ,-I   ./M  ,-I ./R 4T +,E )*C +,E -.G 67Q   o! ("  % * * S%,&'* * *  *  *  *  #u| * * * * ^ % # # S%,&'# # #  #  #  # # # # # # r