o wih7@s&ddlmZmZddlZddlmZddlmZddlm Z ddl m Z ddl m Z mZmZdd lmZdd lmZmZmZmZmZmZe eZGd d d eZGd ddeZddZGdddeZGdddeZ GdddeZ!Gddde Z"GdddeZ#GdddeZ$gdZ%dS))CallableOptionalN)Cache)ALL_ATTENTION_FUNCTIONS)logging)LlamaPreTrainedModel LlamaRMSNormeager_attention_forward) OlmoConfig) OlmoAttentionOlmoDecoderLayerOlmoForCausalLM OlmoModelOlmoRotaryEmbeddingapply_rotary_pos_embcseZdZdZdZddddddddZdgd gfd d gd gfd gd gfd Z                  dfdd 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 ``` olmo2 colwise_rep rowwise_repcolwiserowwise)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_tokenslayersnorm+ Nsilu{Gz?TgF@h㈵>c stjdid|d|d|d|d|d|d|d|d | d | d | d | d | d|d|d|d|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_epsclip_qkv)selfr,r-r.r/r0r1r2r3r4r5r6r7r8r9r:r;r<r=rAkwargs __class__r>d/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/transformers/models/olmo2/modular_olmo2.pyr@vsP      zOlmo2Config.__init__)r r!r"r#r#Nr$r%r&Tr'Nr(Fr)NFr*r+) __name__ __module__ __qualname____doc__ model_typebase_model_tp_planbase_model_pp_planr@ __classcell__r>r>rErGrsFM  rc@seZdZddZdS) Olmo2RMSNormcCsJ|j}|tj}|djddd}|t||j}|j||S)NrT)keepdim) dtypetotorchfloat32powmeanrsqrtvariance_epsilonweight)rCr input_dtypevariancer>r>rGforwards  zOlmo2RMSNorm.forwardN)rHrIrJr^r>r>r>rGrPs rPcCsH|dd|jddf}|d|jdddf}tj| |fddS)z*Rotates half the hidden dims of the input..NrQr)dim)shaperUcat)xx1x2r>r>rG rotate_halfsrecseZdZddedeeffdd Z  ddejde ejejfdeejd ee d eej d e ejeejee ejff d d Z Z S)Olmo2AttentionNconfig layer_idxcs@tj||dt|j|j|j|_t|j|j|j|_dS)Nrh) r?r@rPr0head_dimrAq_normr1k_normrCrgrhrEr>rGr@szOlmo2Attention.__init__rposition_embeddingsrpast_key_valuecache_positionreturncKs<|jdd}g|d|jR}|||} |||} ||} | |dd} | |dd} | |dd} |\} } t | | | | \} } |durc| | |d}| | | |j |\} } t }|j jdkrqt|j j}||| | | |f|js}dn|j|jd|\}}|jg|dR}||}||fS)NrQr'r)sincosrpeagerr*)dropoutscaling)r`rjrkq_projrlk_projv_projview transposerupdaterhr rg_attn_implementationrtrainingr=rvreshape contiguouso_proj)rCrrnrrorprD input_shape hidden_shape query_states key_states value_statesrsrr cache_kwargsattention_interface attn_output attn_weightsr>r>rGr^s>       zOlmo2Attention.forward)N)NN)rHrIrJrrintr@rUTensortupler LongTensorr^rOr>r>rErGrfs" rfcseZdZdedeffdd Z       ddejdeejd eej d ee d ee d ee d eej dee ejejfde ej ee ej ej fffddZZS)Olmo2DecoderLayerrgrhcsJtj||dt|j|jd|_t|j|jd|_t||d|_|` dS)Nrieps)rgrh) r?r@rPr-rApost_attention_layernormpost_feedforward_layernormrf self_attninput_layernormrmrEr>rGr@s zOlmo2DecoderLayer.__init__NFrr position_idsrooutput_attentionsr5rprnrqc Kst|} |jd||||||||d| \}} ||}| |}|} ||}||}| |}|f} |r8| | f7} | S)N)rrrrorr5rprnr>)rrmlpr) rCrrrrorr5rprnrDresidualself_attn_weightsoutputsr>r>rGr^s.     zOlmo2DecoderLayer.forward)NNNFFNN)rHrIrJrrr@rUrrrrboolr FloatTensorr^rOr>r>rErGrs8   rc@ eZdZdS)Olmo2RotaryEmbeddingNrHrIrJr>r>r>rGr'rc@r)Olmo2PreTrainedModelNrr>r>r>rGr+rrcs"eZdZdeffdd ZZS) Olmo2ModelrgcsBttjjd|_tfddtj D|_ dS)Nrcsg|]}t|qSr>)r).0rhrgr>rG 6sz'Olmo2Model.__init__..) r?r@rPr-rArnn ModuleListranger/r)rCrgrErrGr@2s  zOlmo2Model.__init__)rHrIrJrr@rOr>r>rErGr1src@r)Olmo2ForCausalLMNrr>r>r>rGr;rr)rrrr)&typingrrrUtorch.nnr cache_utilsrmodeling_utilsrutilsrllama.modeling_llamar r r olmo.configuration_olmor olmo.modeling_olmor rrrrr get_loggerrHloggerrrPrerfrrrrr__all__r>r>r>rGs*         92