o i/2@sdZddlmZddlmZddlmZmZer0ddlm Z m Z m Z m Z m Z mZmZmZdZn+ddlmZed Z ed Z ed Z ed Z ed Z edZdededefddZdZeeZGdddeZdgZdS)zxLSTM configuration.)Optional)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_ofreturncCst||d||S)z0Rounds up x to the next multiple of multiple_of.)int)rrr a/home/ubuntu/.local/lib/python3.10/site-packages/transformers/models/xlstm/configuration_xlstm.pyr 2sr FcAs,eZdZdZdZ                       dHdededeedeed eed!ed"ed#ed$ed%ed&ed'ed(ed)e d*e d+e d,e d-ed.ed/e d0ed1e d2ed3ed4ed5ed6ed7ed8ed9ed:ed;ef@fdd?Zed@dAZedBdCZedDdEZdFdGZZS)I 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 FTư>??rrrr@rrtV@.@>@rrr@ 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!" s||_|dur |n||_|dur|n||_|dur|n||_|dur$|n||_||_||_| |_| |_| |_ ||_ | |_ | |_ ||_ ||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_||_| |_t j!d|||| d|!dS)N)rNrOrMr9r )"r1r2r3r4r5r6r7r9r:r;r8r<r=r>r?r@rArBrCrDrErFrGrHrIrJrKrLrMrNrOrPsuper__init__)"selfr1r2r3r4r5r6r7r8r9r:r;r<r=r>r?r@rArBrCrDrErFrGrHrIrJrKrLrMrNrOrPkwargs __class__r r!rRsP, zxLSTMConfig.__init__cCt|j|jddSNr+)r)r r2r<rSr r r!qk_dim zxLSTMConfig.qk_dimcCrWrX)r r2r=rYr r r!v_dimr[zxLSTMConfig.v_dimcC |j|jSN)rZr6rYr r r! qk_head_dim zxLSTMConfig.qk_head_dimcCr]r^)r\r6rYr r r! v_head_dimr`zxLSTMConfig.v_head_dimcCstrgtdid|jd|jd|jd|jd|jd|jd|jd|j d |j d |j d |j d |j d |jd|jd|jd|jd|jd|jd|jd|jd|jd|jd|jd|jS|S)Nr1r3r5r6r7r:r;r8r<r=r>r?r@rArBrCrDrErFrGrHrIrJrKr )external_xlstmrr1r2r4r6r7r:r;r8r<r=r>r?r@rArBrCrDrErFrGrHrIrJrKrYr r r!to_xlstm_block_config sf   z!xLSTMConfig.to_xlstm_block_config) r$r%Nr&Nr'FTFTr(r)r*rrrrr+Trr(rr,r+r-r.rTrrr/r0)__name__ __module__ __qualname____doc__ model_typerrboolfloatrr r rr r rRpropertyrZr\r_rarc __classcell__r r rUr!r"<s[     !#%&'()Z    r"N)rgtypingrconfiguration_utilsrutilsrrxlstm.xlstm_large.modelrrr r r r r rrbrr get_loggerrdloggerr"__all__r r r r!s,  (   s