^jbddlZddlmZddlmZddlZddlmcmZ ddlmZddl m Z ddl mZddlmZmZdd lmZdd lmZmZdd lmZdd lmZmZdd lmZmZddlm Z m!Z!ddl"m#Z#m$Z$ddl%m&Z&ddl'm(Z(m)Z)m*Z*ddl+m,Z,m-Z-m.Z.ddl/m0Z0ddl1m2Z2GddejfZ4edGddejjZ6GddejjZ7GddejjZ8dejrd e:d!ejrfd"Z; ded-d=d.Z?Gd/d0ejjZ@Gd1d2eZAe)Gd3d4e$ZBe)Gd5d6eBZCe)Gd7d8eBeZDGd9d:eeBZEgd;ZFy)>N)Callable)Optional)nn)initialization)ACT2FN)Cache DynamicCache)GenerationMixin)use_kernel_forward_from_hubuse_kernel_func_from_hub)create_causal_mask) GenericForSequenceClassificationGradientCheckpointingLayer)BaseModelOutputWithPastCausalLMOutputWithPast)ROPE_INIT_FUNCTIONSdynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)TransformersKwargsauto_docstringcan_return_tuple)is_flash_attention_requestedmaybe_autocastmerge_with_config_defaults)capture_outputs)MiniCPM3Configc ZeZdZdZd dedededeffd Zdejffd Z xZ S) MiniCPM3ScaledWordEmbeddingz\ This module overrides nn.Embeddings' forward by multiplying with embeddings scale. num_embeddings embedding_dim padding_idx embed_scalect||||||_|jdt j |dy)Nr&F persistent)super__init__scalar_embed_scaleregister_buffertorchtensor)selfr#r$r%r& __class__s y/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/transformers/models/minicpm3/modeling_minicpm3.pyr+z$MiniCPM3ScaledWordEmbedding.__init__3s;  D"- ]ELL,ERWX input_idsct|||jj|jj zSN)r*forwardr&toweightdtype)r0r4r1s r2r7z#MiniCPM3ScaledWordEmbedding.forward8s2wy)D,<,<,?,? @Q@Q,RRRr3)?) __name__ __module__ __qualname____doc__intfloatr+r.Tensorr7 __classcell__r1s@r2r"r".sGYsY3YSY_dY SSSr3r"RMSNormcheZdZddeddffd Zdej dej fdZdZxZ S) MiniCPM3RMSNormepsreturnNct|tjt j ||_||_y)z> MiniCPM3RMSNorm is equivalent to T5LayerNorm N)r*r+r Parameterr.onesr9variance_epsilon)r0 hidden_sizerHr1s r2r+zMiniCPM3RMSNorm.__init__>s1 ll5::k#:; #r3 hidden_statesc"|j}|jtj}|j dj dd}|tj ||jzz}|j|j|zS)NT)keepdim) r:r8r.float32powmeanrsqrtrMr9)r0rO input_dtypevariances r2r7zMiniCPM3RMSNorm.forwardFsy#)) %((7  $$Q',,R,>% Ht?T?T4T(UU {{]--k:::r3c^t|jjd|jS)Nz, eps=)tupler9shaperM)r0s r2 extra_reprzMiniCPM3RMSNorm.extra_reprMs* ))*+6$2G2G1HIIr3)gư>) r<r=r>rAr+r.rBr7r]rCrDs@r2rGrG<s7$$$$;U\\;ell;Jr3rGc eZdZUejed<d deffd Ze d dedzde dde dzde d e ffd Z ejed ZxZS)MiniCPM3RotaryEmbeddinginv_freqNconfigct||j|_|j|_||_|j j d|_|j}|jdk7rt|j}||j |\}|_ |jd|d|jd|jdy)N rope_typedefaultr`Fr(original_inv_freq) r*r+max_position_embeddingsmax_seq_len_cachedoriginal_max_seq_lenrarope_parametersrccompute_default_rope_parametersrattention_scalingr-clone)r0radevice rope_init_fnr`r1s r2r+z MiniCPM3RotaryEmbedding.__init__Ts "("@"@$*$B$B! 44[A!%!E!E >>Y &.t~~>L+7 V+L($( ZeD 0(..2BuUr3rmz torch.deviceseq_lenrIz torch.Tensorc|jd}t|ddxs|j|jz}d}d|t j d|dtj j|tj|z zz }||fS) a Computes the inverse frequencies according to the original RoPE implementation Args: config ([`~transformers.PreTrainedConfig`]): The model configuration. device (`torch.device`): The device to use for initialization of the inverse frequencies. seq_len (`int`, *optional*): The current sequence length. Unused for this type of RoPE. Returns: Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE). rope_thetahead_dimNr;rrQr:)rmr:) rigetattrrNnum_attention_headsr.arangeint64r8rA)rarmrobasedimattention_factorr`s r2rjz7MiniCPM3RotaryEmbedding.compute_default_rope_parametersds&%%l3fj$/c63E3EIcIc3c U\\!S!5;;?BB&X]XcXcBdgjj k )))r3cN|jddddfjj|jdddj |j }|dddddfj}t |j jtr/|j jdk7r|j jnd}t|d5|j|jzjdd}tj||fd }|j|jz}|j|jz} dddj |j  j |j fS#1swYAxYw) NrrRrmpscpuF) device_typeenabledrQryrs)r`rAexpandr\r8rm isinstancetypestrr transposer.catcosrksinr:) r0x position_idsinv_freq_expandedposition_ids_expandedr~freqsembrrs r2r7zMiniCPM3RotaryEmbedding.forwardsR!MM$4-8>>@GG HZHZ[\H]_acdehhijiqiqr ,QaZ 8 > > @'1!((--'E!((--[`J`ahhmmfk  U C 5&,,.1F1L1L1NNYYZ[]^_E))UEN3C'')d444C'')d444C  5 vvAGGv$cff177f&;;;  5 5s BFF$r6)NNN)r<r=r>r.rB__annotations__r r+ staticmethodrr@r[rArjno_gradrr7rCrDs@r2r_r_QsllV~V (,+/"*%*(*t* ~u$ % **:U]]_ < r+r7rCrDs@r2rrs 0r3rrOn_reprIc|j\}}}}|dk(r|S|dddddddddfj|||||}|j|||z||S)z This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) rN)r\rreshape)rOrbatchnum_key_value_headsslenrrs r2 repeat_kvrso 2?1D1D.E h z!!Qa"23::5BUW\^bdlmM  (;e(CT8 TTr3modulequerykeyvalueattention_maskscalingdropoutkwargsc t||j}t||j} tj||j dd|z} || |z} t j j| dtjj|j} t j j| ||j} tj| | } | j ddj} | | fS)NrQrrR)ryr:)ptrainingr)rnum_key_value_groupsr.matmulrr functionalsoftmaxrTr8r:rr contiguous) rrrrrrrr key_states value_states attn_weights attn_outputs r2eager_attention_forwardrs3 ; ;