qi8ddlmZddlZddlmZddlmZddlmZddl m Z ddl m Z ddl mZdd lmZd d lmZmZmZd d lmZd d lmZmZmZmZmZmZej<eZ GddeZ!GddeZ"GddeZ#dZ$GddeZ%GddeZ&GddeZ'GddeZ(GddeZ)gdZ*y) )CallableN)TransformersKwargs)Cache)RopeParameters)ALL_ATTENTION_FUNCTIONS)Unpack)logging)LlamaPreTrainedModel LlamaRMSNormeager_attention_forward) OlmoConfig) OlmoAttentionOlmoDecoderLayerOlmoForCausalLM OlmoModelOlmoRotaryEmbeddingapply_rotary_pos_embc&LeZdZdZdZddddddddZdgd gfd d gd gfd gd gfd Z d!ded zded zded zded zded zded zded zded zde d zde d zded zded zded zde d zde e ee fzd zde d zde d zded zf$fd Z xZS)" Olmo2Configa This is the configuration class to store the configuration of a [`Olmo2Model`]. It is used to instantiate an OLMo2 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 [allenai/Olmo2-7B-1124-hf](https://huggingface.co/allenai/Olmo2-7B-1124-hf). 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 50304): Vocabulary size of the Olmo2 model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`Olmo2Model`] hidden_size (`int`, *optional*, defaults to 4096): Dimension of the hidden representations. intermediate_size (`int`, *optional*, defaults to 11008): Dimension of the MLP representations. num_hidden_layers (`int`, *optional*, defaults to 32): Number of hidden layers in the Transformer decoder. num_attention_heads (`int`, *optional*, defaults to 32): 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`. 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 2048): 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. 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 1): Padding token id. bos_token_id (`int`, *optional*): Beginning of stream token id. eos_token_id (`int`, *optional*, defaults to 50279): End of stream token id. tie_word_embeddings (`bool`, *optional*, defaults to `False`): Whether to tie weight embeddings 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`. attention_bias (`bool`, defaults to `False`, *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. rms_norm_eps (`float`, *optional*, defaults to 1e-05): The epsilon used by the rms normalization layers. ```python >>> from transformers import Olmo2Model, Olmo2Config >>> # Initializing a Olmo2 7B style configuration >>> configuration = Olmo2Config() >>> # Initializing a model from the Olmo2 7B style configuration >>> model = Olmo2Model(configuration) >>> # Accessing the model configuration >>> configuration = model.config ``` olmo2colwise_gather_outputrowwise_split_inputcolwiserowwise)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 use_cache pad_token_id bos_token_id eos_token_idtie_word_embeddingsrope_parametersattention_biasattention_dropout rms_norm_epsc t|did|d|d|d|d|d|d|d|d | d | d | d | d | d|d|d|d||||_|`y)Nr$r%r&r'r(r)r*r+r,r-r.r/r0r1r2r3r4)super__init__r5clip_qkv)selfr$r%r&r'r(r)r*r+r,r-r.r/r0r1r2r3r4r5kwargs __class__s Y/opt/pipecat/venv/lib/python3.12/site-packages/transformers/models/olmo2/modular_olmo2.pyr9zOlmo2Config.__init__s,  ! # 0 0  !4  !4  " %< 0   & & & !4 , *! "0% *) M)iii+ r@Nsiluig{Gz?TNigFNFgh㈵>)__name__ __module__ __qualname____doc__ model_typebase_model_tp_planbase_model_pp_planintstrfloatboolrdictr9 __classcell__r=s@r>rr/sENJ%<%<%<%:"+ )"+&(9:#%568IJ!"_$56"'"&(-(**,*.!'.2*.!%#$#'#(+0MQ&+*-#'',$J,4Z,: , : , !4Z ,!4Z,$J,"%t,!4<,$;,Dj,Dj,Dj,"D[, ($sN/B*CCdJ!,"t #,$!4<%,&Dj',,r?rceZdZdZy) Olmo2RMSNormc"|j}|jtj}|j dj dd}|tj ||jzz}|j|zj|S)Nr T)keepdim) dtypetotorchfloat32powmeanrsqrtvariance_epsilonweight)r;r input_dtypevariances r>forwardzOlmo2RMSNorm.forwardsy#)) %((7  $$Q',,R,>% Ht?T?T4T(UU  m+// <rSrSs=r?rSc eZdZy)Olmo2RotaryEmbeddingNrDrErFr7r?r>rdrdr?rdc|dd|jddzf}|d|jddzdf}tj| |fdS)z*Rotates half the hidden dims of the input..NrUr )dim)shaperYcat)xx1x2s r> rotate_halfrnsZ 3"!''"+"" " #B 3 q " " #B 99rc2YB ''r?ceZdZd dededzffd Z ddejdeejejfdejdzde dzd ejdzd e e d eejejdzffd Z xZS)Olmo2AttentionNconfig layer_idxct|||t|j|jz|j |_t|j|jz|j |_y)Nrr) r8r9rSr(head_dimr5q_normr)k_normr;rqrrr=s r>r9zOlmo2Attention.__init__s[ 95"6#=#= #MvObObc "6#=#= #MvObObc r?rposition_embeddingsr past_key_valuescache_positionr<returnc `|jdd}g|d|j}|j|j|} |j |j |} |j |} | j|jdd} | j|jdd} | j|jdd} |\} } t| | | | \} } |'| | |d}|j| | |j|\} } tj|jjt }||| | | |f|j"sdn |j$|j&d|\}}|j(g|dj+}|j-|}||fS)NrUrBr )sincosr{rC)dropoutscaling)rirurvq_projrwk_projv_projview transposerupdaterrr get_interfacerq_attn_implementationrtrainingr4rreshape contiguouso_proj)r;rryr rzr{r< input_shape hidden_shape query_states key_states value_statesrr~ cache_kwargsattention_interface attn_output attn_weightss r>rbzOlmo2Attention.forwards$))#2. 88b8$--8 {{4;;}#=> [[]!;< {{=1 #((6@@AF __\2<rprpsd{dsTzd)-26 -)||-)#5<<#=>-) t+ -)  -) ((4/ -)+,-) u||U\\D00 1-)r?rpc"eZdZdedeffd Z ddejdejdzdejdzde dzd e dzd ejdzd e ejejfdzd e e d ejfdZxZS)Olmo2DecoderLayerrqrrct|||t|j|j|_t|j|j|_t|||_|` y)Nrteps)rqrr) r8r9rSr%r5post_attention_layernormpost_feedforward_layernormrp self_attninput_layernormrxs r>r9zOlmo2DecoderLayer.__init__ s_ 95(4V5G5GVM`M`(a%*6v7I7IvObOb*c''vK  r?Nrr position_idsrzr-r{ryr<r|c |} |jd|||||||d|\}} |j|}| |z}|} |j|}|j|}| |z}|S)N)rr rrzr-r{ryr7)rrmlpr) r;rr rrzr-r{ryr<residual_s r>rbzOlmo2DecoderLayer.forwards!)4>>  ')%+) 3     q55mD  =0 !/ 77 F  =0 r?)NNNFNN)rDrErFrrKr9rYrrrrNrr rrbrPrQs@r>rrs!{!s!/304(,!&26HL|| t+&&-    $; ((4/#5<<#=>E+, r?rc eZdZy)Olmo2PreTrainedModelNrer7r?r>rr1rfr?rc$eZdZdeffd ZxZS) Olmo2Modelrqc t||t|j|j|_t jt|jDcgc]}t||c}|_ ycc}w)Nr) r8r9rSr%r5r#nn ModuleListranger'rr"rxs r>r9zOlmo2Model.__init__8s^   !3!39L9LM mmCHIaIaCb ci vy 1 c  csA=)rDrErFrr9rPrQs@r>rr7s {  r?rc eZdZy)Olmo2ForCausalLMNrer7r?r>rrArfr?r)rrrr)+collections.abcrrYtorch.nnrtransformers.utils.genericr cache_utilsrmodeling_rope_utilsrmodeling_utilsrprocessing_utilsr utilsr llama.modeling_llamar r rolmo.configuration_olmorolmo.modeling_olmorrrrrr get_loggerrDloggerrrSrdrnrprrrr__all__r7r?r>rs(% 9 15&^^0   H %D*DR=<= . (3)]3)r&(&R /      r?