a h7@sJddlmZmZddlZddlmZddlmZddlm Z ddl m Z ddl m Z ddlmZdd lmZd d lmZmZmZd d lmZd d lmZmZmZmZmZmZee Z!GdddeZ"GdddeZ#ddZ$GdddeZ%GdddeZ&GdddeZ'GdddeZ(GdddeZ)GdddeZ*gd Z+dS)!)CallableOptionalN)TransformersKwargs)Cache)ALL_ATTENTION_FUNCTIONS)Unpack)logging)deprecate_kwarg)LlamaPreTrainedModel LlamaRMSNormeager_attention_forward) OlmoConfig) OlmoAttentionOlmoDecoderLayerOlmoForCausalLM OlmoModelOlmoRotaryEmbeddingapply_rotary_pos_embcs`eZdZdZdZddddddddZdgd gfd d gd gfd gd gfd Zdfdd ZZS) 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_theta (`float`, *optional*, defaults to 10000.0): The base period of the RoPE embeddings. rope_scaling (`Dict`, *optional*): Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update `max_position_embeddings` to the expected new maximum. See the following thread for more information on how these scaling strategies behave: https://www.reddit.com/r/LocalLLaMA/comments/14mrgpr/dynamically_scaled_rope_further_increases/. This is an experimental feature, subject to breaking API changes in future versions. 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 ``` Zolmo2Z colwise_repZ rowwise_repZcolwiseZrowwise)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_projZ input_idsZ inputs_embeds hidden_statesattention_mask)Z embed_tokenslayersnorm+ Nsilu{Gz?TgF@h㈵>c sFtjf||||||||| | | | | |||||d|||_|`dS)N) 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_embeddings rope_theta rope_scalingattention_biasattention_dropout)super__init__ rms_norm_epsZclip_qkv)selfr'r(r)r*r+r,r-r.r/r0r1r2r3r4r5r6r7r8r;kwargs __class__c/var/www/html/assistant/venv/lib/python3.9/site-packages/transformers/models/olmo2/modular_olmo2.pyr:zs0zOlmo2Config.__init__)rrrrrNrr r!Tr"Nr#Fr$NFr%r&) __name__ __module__ __qualname____doc__Z model_typeZbase_model_tp_planZbase_model_pp_planr: __classcell__r@r@r>rArsDM   rc@seZdZddZdS) Olmo2RMSNormcCsJ|j}|tj}|djddd}|t||j}|j||S)Nr T)Zkeepdim) ZdtypetotorchZfloat32powmeanZrsqrtZvariance_epsilonweight)r<rZ input_dtypeZvariancer@r@rAforwards  zOlmo2RMSNorm.forwardN)rBrCrDrNr@r@r@rArGsrGcCsH|dd|jddf}|d|jdddf}tj| |fddS)z*Rotates half the hidden dims of the input..NrHr )dim)shaperJcat)xx1Zx2r@r@rA rotate_halfsrTc seZdZd eeedfdd Zeddddd ej e ej ej feej ee eej e ee ej eej fd d d ZZS)Olmo2AttentionNconfig layer_idxcs@tj||dt|j|j|j|_t|j|j|j|_dS)NrX) r9r:rGr+head_dimr;q_normr,k_normr<rWrXr>r@rAr:szOlmo2Attention.__init__past_key_valuepast_key_values4.58new_nameversion)rposition_embeddingsrr_cache_positionr=returncKs<|jdd}g|d|jR}|||} |||} ||} | |dd} | |dd} | |dd} |\} } t | | | | \} } |dur| | |d}| | | |j |\} } t }|j jdkrt|j j}||| | | |f|jsdn|j|jd|\}}|jg|dR}||}||fS)NrHr"r )sincosreeagerr%)Zdropoutscaling)rPrZr[Zq_projr\Zk_projZv_projviewZ transposerupdaterXrrWZ_attn_implementationrZtrainingr8rjZreshape contiguousZo_proj)r<rrdrr_rer=Z input_shapeZ hidden_shapeZ query_statesZ key_statesZ value_statesrhrgZ cache_kwargsZattention_interfaceZ attn_outputZ attn_weightsr@r@rArNs>       zOlmo2Attention.forward)N)NN)rBrCrDrrintr:r rJTensortupler LongTensorrrrNrFr@r@r>rArUs rUc seZdZeedfdd Zeddddd eje eje ej e e e e e ej e e ejejfeeejd d d ZZS)Olmo2DecoderLayerrVcsJtj||dt|j|jd|_t|j|jd|_t||d|_|` dS)NrYepsrV) r9r:rGr(r;post_attention_layernormpost_feedforward_layernormrU self_attnZinput_layernormr]r>r@rAr:s zOlmo2DecoderLayer.__init__r^r_r`raNF) rr position_idsr_r0rerdr=rfc Ks^|} |jf|||||||d|\}} ||}| |}|} ||}||}| |}|S)N)rrrxr_r0rerd)rwruZmlprv) r<rrrxr_r0rerdr=Zresidual_r@r@rArNs&    zOlmo2DecoderLayer.forward)NNNFNN)rBrCrDrrnr:r rJrorrqrboolrprrrNrFr@r@r>rArrs& rrc@s eZdZdS)Olmo2RotaryEmbeddingNrBrCrDr@r@r@rAr{$sr{c@s eZdZdS)Olmo2PreTrainedModelNr|r@r@r@rAr}(sr}cs"eZdZedfdd ZZS) Olmo2ModelrWcsBttjjd|_tfddtj D|_ dS)Nrscsg|]}t|qSr@)rr).0rXrr@rA 3z'Olmo2Model.__init__..) r9r:rGr(r;rnnZ ModuleListranger*r)r<rWr>rrAr:/s  zOlmo2Model.__init__)rBrCrDrr:rFr@r@r>rAr~.sr~c@s eZdZdS)Olmo2ForCausalLMNr|r@r@r@rAr8sr)rrr~r}),typingrrrJZtorch.nnrZtransformers.utils.genericrZ cache_utilsrZmodeling_utilsrZprocessing_utilsrutilsr Zutils.deprecationr Zllama.modeling_llamar r rZolmo.configuration_olmorZolmo.modeling_olmorrrrrrZ get_loggerrBloggerrrGrTrUrrr{r}r~r__all__r@r@r@rAs.           :*