o ۷iPH@sdZddlmZddlmZeeZGdddeZGdddeZ Gdd d eZ Gd d d eZ Gd d d eZ gdZ dS)zBlt model configuration)PretrainedConfig)loggingcFeZdZdZdZ          dfdd ZZS)BltLocalEncoderConfigzB Configuration class for the Blt Local Encoder component. blt_local_encoderFNh㈵>`Asilu {Gz?c ||_||_||_||_||_||_|p||_|||_|p%td|d|_ ||_ | |_ | |_ | |_ | |_| |_||_||_|ddtjdi|ddidSNrtie_word_embeddingsF vocab_sizecross_attn_all_layers cross_attn_khidden_size_global hidden_sizenum_attention_headsnum_key_value_headshead_dimintintermediate_sizenum_hidden_layers rms_norm_epsdropoutmax_position_embeddings rope_theta rope_scaling hidden_actinitializer_rangepopsuper__init__selfrrrrrrr r$r%r&r'r(r)r*r#r+kwargs __class__r_/home/ubuntu/vllm_env/lib/python3.10/site-packages/transformers/models/blt/configuration_blt.pyr.&   zBltLocalEncoderConfig.__init__)rFrr r r Nr r rrrNrrr__name__ __module__ __qualname____doc__ model_typer. __classcell__rrr2r4r(rcr)BltLocalDecoderConfigzB Configuration class for the Blt Local Decoder component. blt_local_decoderrTrr r r N r rrrrrrc rrrr/r2rr4r.Qr5zBltLocalDecoderConfig.__init__)rTrr r r Nr@r rrrNrrrr6rrr2r4r>Jr=r>cs>eZdZdZdZ        dfdd ZZS)BltGlobalTransformerConfigzG Configuration class for the Blt Global Transformer component. blt_global_transformerr r Nr rrrrc  s||_||_|p ||_|||_| ptd|d|_||_||_||_||_ ||_ | |_ | |_ | |_ | ddtjdi| ddidSr)rrr r!r"r#r$r%r&r'r(r)r*r+r,r-r.)r0rrr r$r%r&r'r(r)r*r#r+r1r2rr4r.s   z#BltGlobalTransformerConfig.__init__) r r NrCr rrDrNrrErr6rrr2r4rA|s rAcs>eZdZdZdZ        dfdd ZZS)BltPatcherConfiga Configuration class for the Blt Patcher/Entropy model component. Args: vocab_size (`int`, *optional*, defaults to 260): Vocabulary size of the Blt patcher model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling the patcher model. hidden_size (`int`, *optional*, defaults to 768): Dimension of the hidden representations. num_hidden_layers (`int`, *optional*, defaults to 14): Number of hidden layers in the Transformer decoder. num_attention_heads (`int`, *optional*, defaults to 12): Number of attention heads for each attention layer in the Transformer decoder. num_key_value_heads (`int`, *optional*): This is the number of key_value heads that should be used to implement Grouped Query Attention. If `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group. For more details, check out [this paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `num_attention_heads`. max_position_embeddings (`int`, *optional*, defaults to 8192): The maximum sequence length that this model might ever be used with. rms_norm_eps (`float`, *optional*, defaults to 1e-05): The epsilon used by the rms normalization layers. dropout (`float`, *optional*, defaults to 0.0): The dropout ratio for the attention probabilities. rope_theta (`float`, *optional*, defaults to 10000.0): The base period of the RoPE embeddings. intermediate_size (`int`, *optional*, defaults to 2048): Dimension of the MLP representations. rope_scaling (`dict`, *optional*): Dictionary containing the RoPE scaling configuration. initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. blt_patcherr N r r@r rc  s||_||_||_||_|||_|dur|n||_||_||_||_| |_ d|_ | p3t d|jd|_ | |_ | |_| ddtjdi| ddidS)NrrrrFr)rrr$rr!r r'r%r&r(r*r"r#r)r+r,r-r.)r0rrr$rr r'r%r&r(r#r)r+r1r2rr4r.s   zBltPatcherConfig.__init__) rrHrIrJNrKr rrLr Nrr6rrr2r4rFs %rFcsbeZdZdZdZdgZeeee dZ         dfdd Z Z S) BltConfigas This is the configuration class to store the configuration of a [`BltModel`]. It is used to instantiate a Blt model according to the specified arguments, defining the model architecture. Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PretrainedConfig`] for more information. Args: vocab_size (`int`, *optional*, defaults to 260): Vocabulary size of the Blt model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`BltModel`]. max_position_embeddings (`int`, *optional*, defaults to 4096): The maximum sequence length that this model might ever be used with. patch_in_forward (`bool`, *optional*, defaults to `True`): Whether to perform patching during the forward pass. patch_size (`int`, *optional*, defaults to 4): Size of the patches used in the patching mechanism. patching_mode (`str`, *optional*, defaults to `"entropy"`): The mode used for patching, such as entropy-based patching. patching_threshold (`float`, *optional*, defaults to 1.34): Threshold value used for determining when to apply patches. patching_batch_size (`int`, *optional*, defaults to 1): Batch size used during the patching process. max_patch_length (`int`, *optional*): Maximum length of patches that can be generated. cross_attn_k (`int`, *optional*, defaults to 2): Number of cross-attention heads used in the model. encoder_hash_byte_group_size (`list`, *optional*): List of byte group sizes used in the encoder hash function. encoder_hash_byte_group_vocab (`int`, *optional*, defaults to 500002): Vocabulary size for the encoder hash byte groups. encoder_hash_byte_group_nb_functions (`int`, *optional*, defaults to 1): Number of hash functions used in the encoder byte grouping. patcher_config (`BltPatcherConfig`, *optional*): Configuration for the patcher component of the model. encoder_config (`BltLocalEncoderConfig`, *optional*): Configuration for the local encoder component of the model. decoder_config (`BltLocalDecoderConfig`, *optional*): Configuration for the local decoder component of the model. global_config (`BltGlobalTransformerConfig`, *optional*): Configuration for the global transformer component of the model. tie_word_embeddings (`bool`, *optional*, defaults to `False`): Whether to tie weight embeddings. initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. rope_theta (`float`, *optional*, defaults to 500000.0): The base period of the RoPE embeddings. rope_scaling (`dict`, *optional*): Dictionary containing the RoPE scaling configuration. ```python >>> from transformers import BltModel, BltConfig >>> # Initializing a Blt configuration >>> configuration = BltConfig() >>> # Initializing a model from the configuration >>> model = BltModel(configuration) >>> # Accessing the model configuration >>> configuration = model.config ``` Checkpoint: [facebook/blt](https://huggingface.co/facebook/blt) bltpast_key_values)patcher_configencoder_configdecoder_config global_configrrDTentropy]?r Nr"Frrc sF||_||_||_||_||_||_||_||_||_||_ ||_ | dd|_ | dd|_ | d|_| dd|_| |_| pDgd|_| |_| |_| dur\t|d |_td nt| trp| d |tdi| |_nt| trx| |_|durt|d |_td nt|tr|d |tdi||_nt|tr||_|durt|d |_td nt|tr|d |tdi||_nt|tr||_|durt|d |_ tdnt|tr|d |tdi||_ nt|tr||_ |jj!|j}||j j!kr |nd|j _"|#ddt$j%dd|i|dS)Npatching_devicecudarealtime_patchingTpatching_threshold_add monotonicityF)rrTr)r+z8patcher_config is None, using default Blt patcher configr+z8encoder_config is None, using default Blt encoder configz8decoder_config is None, using default Blt decoder configz6global_config is None, using default Blt global configrr)&rr'r+r(r)patch_in_forward patch_size patching_modepatching_thresholdpatching_batch_sizemax_patch_lengthgetrXrZr[r\rencoder_hash_byte_group_sizeencoder_hash_byte_group_vocab$encoder_hash_byte_group_nb_functionsrFrPloggerinfo isinstancedict setdefaultrrQr>rRrArSrencoder_cross_output_sizer,r-r.)r0rr'r`rarbrcrdrerrgrhrirPrQrRrSrr+r(r)r1ror2rr4r.>sp                      zBltConfig.__init__)rrDTrTrUrVr NrNrWr NNNNFrrN) r7r8r9r:r;keys_to_ignore_at_inferencerFrr>rA sub_configsr.r<rrr2r4rMs<B rM)rMrFrr>rAN)r:configuration_utilsrutilsr get_loggerr7rjrr>rArFrM__all__rrrr4s   22*L 0