3i1dZddlmZddlmZmZerddlmZmZm Z m Z m Z m Z m Z mZdZn0ddlmZedZed Zed Z ed Z ed Z ed Z dededefdZ dZej&eZGddeZdgZy)zxLSTM configuration.)PreTrainedConfig)is_xlstm_availablelogging)BackendModeTypeChunkwiseKernelType DtypeTypeSequenceKernelTypeStepKernelTypeWeightModeTyperound_up_to_next_multiple_ofxLSTMLargeConfigT)Literal)traintrain_with_padding inference)chunkwise--native_autogradzparallel--native_autograd)float32bfloat16float16native_sequence__nativenative)singlefusedx multiple_ofreturnc0t||zdz |z|zS)z0Rounds up x to the next multiple of multiple_of.)int)rrs i/mnt/e/genesis-system/.venv/lib/python3.12/site-packages/transformers/models/xlstm/configuration_xlstm.pyr r 0s Q_q([8KGHHFcABeZdZdZdZ d*dedededzdedzdedzd ed ed ed ed ededededede de de dedede dede dedededede ded ed!ed"ed#ef@fd$ Zed%Zed&Zed'Zed(Zd)ZxZS)+ xLSTMConfiga This is the configuration class to store the configuration of a [`xLSTM`]. It is used to instantiate a xLSTM model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the xLSTM-7b [NX-AI/xLSTM-7b](https://huggingface.co/NX-AI/xLSTM-7b) model. 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, *optional*, defaults to 50304): Vocabulary size of the xLSTM model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`xLSTMModel`]. Defaults to the GPT2-NeoX tokenizer size. hidden_size (int, optional, *optional*, defaults to 4096): Dimensionality of the embeddings or hidden states. embedding_dim (int, optional, *optional*, defaults to 4096): Dimensionality of the embeddings or hidden states, use hidde_size if None. num_hidden_layers (int, optional, *optional*, defaults to 32): Number of blocks of the xLSTM model. num_blocks (int, optional, *optional*, defaults to 32): Number of blocks of the xLSTM model, use num_hidden_layers if None. num_heads (int, optional, *optional*, defaults to 8): Number of heads for the xLSTM Layer/Cell. use_bias (bool, optional, *optional*, defaults to `False`): Whether to use biases in the xLSTM model. norm_reduction_force_float32 (bool, optional, *optional*, defaults to `True`): Whether to force the float32 norm reduction op to be done in fp32 precision. tie_word_embeddings (bool, optional, *optional*, defaults to `False`): Whether to tie word embeddings to the lm head weights. add_out_norm (bool, optional, *optional*, defaults to `True`): Whether to add an output norm after the blocks before the LMHead. norm_eps (float, optional, *optional*, defaults to 1e-06): Norm eps for RMSNorm and Layer Norm. qk_dim_factor (float, optional, *optional*, defaults to 0.5): Scale factor for the query and key dimension. v_dim_factor (float, optional, *optional*, defaults to 1.0): Scale factor for the value dimension. chunkwise_kernel (ChunkwiseKernelType, optional, *optional*, defaults to `"chunkwise--native_autograd"`): Kernel type for chunkwise processing mode. sequence_kernel (SequenceKernelType, optional, *optional*, defaults to `"native_sequence__native"`): Kernel type for sequence processing mode. step_kernel (StepKernelType, optional, *optional*, defaults to `"native"`): Kernel type for step processing mode. mode (BackendModeType, optional, *optional*, defaults to `"inference"`): Operation mode (inference is needed for generation). chunk_size (int, optional, *optional*, defaults to 64): Internal chunk size. return_last_states (bool, optional, *optional*, defaults to `True`): If to return the last states / cache internally. Needed as True for generation. autocast_kernel_dtype (DtypeType, optional, *optional*, defaults to `"bfloat16"`): Kernel dtype for the states. eps (float, optional, *optional*, defaults to 1e-06): Epsilon for the mLSTM cell post norm. inference_state_dtype (DtypeType, optional, *optional*, defaults to `"float32"`): Kernel dtype for states in inference. ffn_proj_factor (float, optional, *optional*, defaults to 2.667): Size factor of the post-up projection gated Feed Forward network. ffn_round_up_to_multiple_of (int, optional, *optional*, defaults to 64): Size factor round value of the post-up projection gated Feed Forward network. gate_soft_cap (float, optional, *optional*, defaults to 15.0): Gate soft cap scale. output_logit_soft_cap (float, optional, *optional*, defaults to 30.0): Output logit soft cap scale. weight_mode (`Literal`, *optional*, defaults to `"single"`): Whether parallel linear layers are separated or fused (single). use_cache (bool, optional, *optional*, defaults to `True`): Whether to use the cache (xLSTMCache). pad_token_id (int, optional, *optional*, defaults to 1): Pad token id needed for generation. bos_token_id (int, optional, *optional*, defaults to 0): BOS token id needed for generation. eos_token_id (int, optional, *optional*, defaults to 2): EOS token id needed for generation. max_inference_chunksize (int, optional, *optional*, defaults to 16384): Limit the chunk size for inference to save memory. Example: ```python >>> from transformers import xLSTMConfig, xLSTMModel >>> # Initializing a xLSTM configuration >>> configuration = xLSTMConfig() >>> # Initializing a model (with random weights) from the configuration >>> model = xLSTMModel(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```xlstmN vocab_size hidden_size embedding_dimnum_hidden_layers num_blocks num_headsuse_biasnorm_reduction_force_float32tie_word_embeddings add_out_normnorm_eps qk_dim_factor v_dim_factorchunkwise_kernelsequence_kernel step_kernelmode chunk_sizereturn_last_statesautocast_kernel_dtypeepsinference_state_dtypeffn_proj_factorffn_round_up_to_multiple_of gate_soft_capoutput_logit_soft_cap weight_mode use_cache pad_token_id bos_token_id eos_token_idmax_inference_chunksizec! <||_||n||_||n||_||n||_||n||_||_||_| |_| |_| |_ ||_ | |_ | |_ ||_ ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| |_||_||_||_| |_tA"|di|!y)N)"r&r'r(r)r*r+r,r.r/r0r-r1r2r3r4r5r6r7r8r9r:r;r<r=r>r?r@rArBrCrDrEsuper__init__)#selfr&r'r(r)r*r+r,r-r.r/r0r1r2r3r4r5r6r7r8r9r:r;r<r=r>r?r@rArBrCrDrEkwargs __class__s# r!rIzxLSTMConfig.__init__sLX%*5*A;}.;.G][6G6S!2Yc(2(>*DU"  #6 (  ,H)*( 0.& $"4%:"%:".+F(*%:"&"((('>$(((#6  "6"r"cJt|j|jzdSN@)r)r r'r1rJs r!qk_dimzxLSTMConfig.qk_dims&+   t11 1  r"cJt|j|jzdSrN)r r'r2rPs r!v_dimzxLSTMConfig.v_dims&+   t00 0  r"c4|j|jzSN)rQr+rPs r! qk_head_dimzxLSTMConfig.qk_head_dims{{dnn,,r"c4|j|jzSrU)rSr+rPs r! v_head_dimzxLSTMConfig.v_head_dimszzT^^++r"ctrAtdid|jd|jd|jd|j d|j d|jd|jd|jd |jd |jd |jd |jd |jd|jd|j d|j"d|j$d|j&d|j(d|j*d|j,d|j.d|j0d|j2S|S)Nr&r(r*r+r,r/r0r-r1r2r3r4r5r6r7r8r9r:r;r<r=r>r?r@rG)external_xlstmrr&r'r)r+r,r/r0r-r1r2r3r4r5r6r7r8r9r:r;r<r=r>r?r@rPs r!to_xlstm_block_configz!xLSTMConfig.to_xlstm_block_configs #??".. 11..    ".. .2-N-N#00".."&!6!6!% 4 4!,, YY!" ??#$$(#:#:%&'+&@&@'(HH)*'+&@&@+.!% 4 4/0-1,L,L14#0056'+&@&@78!,,9 >Kr") iiN NFTFTư>g?g?rrrrrOTrr^rgtV@rOg.@g>@rTrri@)__name__ __module__ __qualname____doc__ model_typer boolfloatrr r rr r rIpropertyrQrSrVrXr[ __classcell__)rLs@r!r$r$:sJY vJ $((*!%-1$)!"!0L.G&. +#'+5+4!&+-#'+&.',SV#V#V#Tz V# : V# $J V#V#V#'+V#"V#V#V#V#V#".#V#$,%V#&$'V#*+V#,-V#0!1V#2 )3V#45V#6 )7V#:;V#<&)=V#@AV#B %CV#F$GV#JKV#LMV#NOV#PQV#R"%SV#p    --,,!r"r$N)rcconfiguration_utilsrutilsrrxlstm.xlstm_large.modelrrr r r r r rrZtypingrr get_loggerr`loggerr$__all__rGr"r!rps 30   NHIO! %89I !:;X&N./NII#I#IN   H %m"m` /r"