qi15ddlmZddlZddlmZddlmZmZddlm Z ddl m Z ddl m Z dd lmZdd lmZd d lmZmZmZmZmZmZmZmZmZd d lmZmZej>e Z!GddeZ"GddeZ#GddeZ$GddeZ%GddeZ&GddeZ'GddeZ(GddeZ)GddeZ*Gd d!eZ+gd"Z,y)#)CallableN)Cache)PreTrainedConfiglayer_type_validation)FlashAttentionKwargs)RopeParameters)ALL_ATTENTION_FUNCTIONS)Unpack)logging) LlamaAttentionLlamaDecoderLayerLlamaForCausalLMLlamaForQuestionAnsweringLlamaForSequenceClassificationLlamaForTokenClassificationLlamaPreTrainedModelapply_rotary_pos_embeager_attention_forward) Qwen2ModelQwen2RotaryEmbeddingc2eZdZdZdZdgZdZddddddddZdgd gfd d gd gfd gd gfd Z d'de d zde d zde d zde d zde d zde d zde d zde d zde d zde d zde d zde d zde d zde d zde ee e fzd zde d zde d zde d zd e d zd!e d zd"e d zd#e d zd$e d zd%e d zf0fd& ZxZS)( SmolLM3Configa! This is the configuration class to store the configuration of a [`SmolLM3Model`]. It is used to instantiate a SmolLM3 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 SmolLM3 3B. e.g. [HuggingFaceTB/SmolLM3-3B](https://huggingface.co/HuggingFaceTB/SmolLM3-3B) 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 128256): Vocabulary size of the SmolLM3 model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`SmolLM3Model`] hidden_size (`int`, *optional*, defaults to 2048): Dimension of the hidden representations. intermediate_size (`int`, *optional*, defaults to 11008): Dimension of the MLP representations. num_hidden_layers (`int`, *optional*, defaults to 36): Number of hidden layers in the Transformer encoder. num_attention_heads (`int`, *optional*, defaults to 16): Number of attention heads for each attention layer in the Transformer encoder. num_key_value_heads (`int`, *optional*, defaults to 4): 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 checkout [this paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `16`. hidden_act (`str` or `function`, *optional*, defaults to `"silu"`): The non-linear activation function (function or string) in the decoder. max_position_embeddings (`int`, *optional*, defaults to 32768): The maximum sequence length that this model might ever be used with. initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. rms_norm_eps (`float`, *optional*, defaults to 1e-06): The epsilon used by the rms normalization layers. use_cache (`bool`, *optional*, defaults to `True`): Whether or not the model should return the last key/values attentions (not used by all models). Only relevant if `config.is_decoder=True`. pad_token_id (`int`, *optional*, defaults to 128004): The id of the padding token. bos_token_id (`int`, *optional*, defaults to 128000): The id of the beginning of sentence token. eos_token_id (`int`, *optional*, defaults to 128001): The id of the end of sentence token. rope_parameters (`RopeParameters`, *optional*): Dictionary containing the configuration parameters for the RoPE embeddings. The dictionary should contain a value for `rope_theta` and optionally parameters used for scaling in case you want to use RoPE with longer `max_position_embeddings`. use_sliding_window (`bool`, *optional*, defaults to `False`): Whether to use sliding window attention. sliding_window (`int`, *optional*): Sliding window attention (SWA) window size. If not specified, will default to `None`. no_rope_layers (`List[int]`, *optional*): List with at least the same length as the number of layers in the model. A `1` at an index position indicates that the corresponding layer will use RoPE, while a `0` indicates that it's a NoPE layer. no_rope_layer_interval (`int`, *optional*, defaults to 4): If `no_rope_layers` is `None`, it will be created using a NoPE layer every `no_rope_layer_interval` layers. layer_types (`list`, *optional*): Attention pattern for each layer. Automatically computed based on sliding window and NoPE settings. attention_bias (`bool`, *optional*, defaults to `False`): Whether to use a bias in the query, key, value and output projection layers during self-attention. attention_dropout (`float`, *optional*, defaults to 0.0): The dropout ratio for the attention probabilities. ```python >>> from transformers import SmolLM3Model, SmolLM3Config >>> # Initializing a SmolLM3 style configuration >>> configuration = SmolLM3Config() >>> # Initializing a model from the SmolLM3 style configuration >>> model = SmolLM3Model(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```smollm3past_key_valuesg>Acolwiserowwise)zlayers.*.self_attn.q_projzlayers.*.self_attn.k_projzlayers.*.self_attn.v_projzlayers.*.self_attn.o_projzlayers.*.mlp.gate_projzlayers.*.mlp.up_projzlayers.*.mlp.down_proj input_ids inputs_embeds hidden_statesattention_mask) embed_tokenslayersnormN vocab_size hidden_sizeintermediate_sizenum_hidden_layersnum_attention_headsnum_key_value_heads hidden_actmax_position_embeddingsinitializer_range rms_norm_eps use_cache pad_token_id bos_token_id eos_token_idrope_parametersuse_sliding_windowsliding_windowno_rope_layersno_rope_layer_interval layer_typesattention_biasattention_dropoutmlp_biastie_word_embeddingsc | |_| |_||_||_||_||_||_||_||_||_ ||_ ||_ ||_ ||}||_ ||_| |_| |_| |_||_||_|1t)|Dcgc]}t+|dz|zdk7c}|_n||_||_|Jg}t)|D]:}|j,|}|r||s|j1d*|j1d<||_t5|j2|j||_t9|tdi|ycc}w)Nrsliding_attentionfull_attention)r1r2r3r=r&r-r<r'r(r)r*r5r6r+r,r.r/r0r:r;rangeintr7r8appendr9rr4super__init__)selfr&r'r(r)r*r+r,r-r.r/r0r1r2r3r4r5r6r7r8r9r:r;r<r=kwargs layer_idxhas_rope __class__s ]/opt/pipecat/venv/lib/python3.12/site-packages/transformers/models/smollm3/modular_smollm3.pyrGzSmolLM3Config.__init__s8)((#6 $'>$  &!2!2#6 "4,  &"5 #6 $!2(",!2  !TYZkTl#GPY]&<<AB#D #1D &<#  K"#45 9 ..y9%.*DX&&':;&&'78  9'd..0F0FG. "6"-#s!E)iii+$siluig{Gz?gư>TiiiNFNNrPNFFT)__name__ __module__ __qualname____doc__ model_typekeys_to_ignore_at_inference default_thetabase_model_tp_planbase_model_pp_planrDstrfloatboolr dictrG __classcell__rLs@rMrr+s@N `J#4"5M&/%.%.%."+ )"+&(9:#%568IJ!"_$56"("&(-(**,*+!'.3*.#'!%#)#)#)MQ*/%)%)-."&&+*- %+/3M#$JM#4ZM#: M# : M# !4Z M#!4ZM#$JM#"%tM#!4<M#DjM#$;M#DjM#DjM#DjM# ($sN/B*CCdJ!M#"!4K#M#$d %M#&d 'M#(!$d )M#*4Z+M#,t -M#.!4*) t+ *)  *) ((4/ *)-.*) u||U\\D00 1*)rbrhc(eZdZdedeffd ZxZS)SmolLM3DecoderLayerrirJcNt||||j||_y)N)rFrGr9attention_typerls rMrGzSmolLM3DecoderLayer.__init__s% +$00;rb)rSrTrUrrDrGr`ras@rMrrs<}<<rs% JB15&   D   H %q#$q#h 1 5)~5)p<+<  1  :  )  'E  $?  ";  rb