o wiy@sdZddlmZmZddlZddlmZddlmZddl m Z ddl m Z dd l mZdd lmZd d lmZd d lmZmZmZmZmZmZmZddlmZeeZGdddeZ GdddeZ!GdddeZ"GdddeZ#GdddeZ$GdddeZ%gdZ&dS)zPyTorch BitNet model.)CallableOptionalN)Cache)FlashAttentionKwargs)CausalLMOutputWithPast)ALL_ATTENTION_FUNCTIONS)Unpack)logging)GemmaMLP)LlamaAttentionLlamaDecoderLayerLlamaForCausalLM LlamaModel LlamaRMSNormapply_rotary_pos_embeager_attention_forward) BitNetConfigc@ eZdZdS) BitNetRMSNormN__name__ __module__ __qualname__rrf/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/transformers/models/bitnet/modular_bitnet.pyr*rcs*eZdZdeffdd ZddZZS) BitNetMLPconfigcs"t|t|j|jd|_dSN)eps)super__init__rintermediate_size rms_norm_eps ffn_sub_norm)selfr  __class__rrr$/s zBitNetMLP.__init__c Cs*|||||||}|S)N) down_projr'act_fn gate_projup_proj)r(xr+rrrforward3s&zBitNetMLP.forward)rrrrr$r0 __classcell__rrr)rr.srcseZdZdedeffdd Z  ddejdeejejfde ejd e e d e ej d e e d eeje eje eejffd dZZS)BitNetAttentionr layer_idxcs$t||t|j|jd|_dSr!)r#r$r hidden_sizer& attn_sub_norm)r(r r3r)rrr$9szBitNetAttention.__init__N hidden_statesposition_embeddingsattention_maskpast_key_valuecache_positionkwargsreturncKs.|jdd}g|d|jR}|||dd} |||dd} |||dd} |\} } t| | | | \} } |durW| | |d}|| | |j |\} } t }|j j dkret |j j }||| | | |f|jsqdn|j|jd|\}}|jg|dR}||}||}||fS)Nrr )sincosr:eagerg)dropoutscaling)shapehead_dimq_projview transposek_projv_projrupdater3rr _attn_implementationrtrainingattention_dropoutrBreshape contiguousr5o_proj)r(r6r7r8r9r:r; input_shape hidden_shape query_states key_states value_statesr?r> cache_kwargsattention_interface attn_output attn_weightsrrrr0=s:       zBitNetAttention.forward)NN)rrrrintr$torchTensortuplerr LongTensorr rr0r1rrr)rr28s& r2c@r)BitNetDecoderLayerNrrrrrr_krr_c@r) BitNetModelNrrrrrr`orr`cs0eZdZdgZdZdZdeffdd ZZS)BitNetForCausalLMzlm_head.weightNr<c stjdi|S)a$ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should either be in `[0, transformers., config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, transformers., config.vocab_size]`. Example: ```python >>> from transformers import AutoTokenizer, BitNetForCausalLM >>> model = BitNetForCausalLM.from_pretrained("microsoft/bitnet-b1.58-2B-4T") >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/bitnet-b1.58-2B-4T") >>> prompt = f'<|begin_of_text|>User: Hey, are you conscious? Can you talk to me?<|eot_id|>Assistant: ' >>> inputs = tokenizer(prompt, return_tensors="pt") >>> # Generate >>> generate_ids = model.generate(inputs.input_ids, max_length=100) >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] "User: Hey, are you conscious? Can you talk to me?Assistant: No, I'm not conscious. I'm an artificial intelligence designed to assist with information and tasks. How can I help you today?" ```Nr)r#r0)r( super_kwargsr)rrr0xszBitNetForCausalLM.forward) rrr_tied_weights_keys_tp_plan_pp_planrr0r1rrr)rrassra)rar`BitNetPreTrainedModel)'__doc__typingrrr[ cache_utilsrmodeling_flash_attention_utilsrmodeling_outputsrmodeling_utilsrprocessing_utilsr utilsr gemma.modeling_gemmar llama.modeling_llamar rrrrrrconfiguration_bitnetr get_loggerrloggerrrr2r_r`ra__all__rrrrs(        $  3 "