T^(hddlmZddlmZmZmZmZmZddlZddl m cm Z ddlm Z ddl mZddlmZmZmZmZddlmZdd lmZdd lmZdd lmZmZmZmZmZm Z m!Z!dd l"m#Z#m$Z$dd l%m&Z&m'Z'ddl(m)Z)ddl*m+Z+m,Z,m-Z-m.Z.m/Z/m0Z0m1Z1ddl2m3Z3ddl4m5Z5e0jle7Z8dZ9dZ:Gdde jvZ<Gdde jvZ=Gdde jvZ>dZ?dFdZ@dejdeBdejfd ZC dGd!e jvd"ejd#ejd$ejd%eejd&eDd'eDfd(ZEGd)d*e jvZFGd+d,e jvZGGd-d.e jvZHd/ZIe-d0eIGd1d2e'ZJd3ZKe-d0eIGd4d5eJZLGd6d7ee+ZM dHd8eejeejdfd9eeBd%eejdeejeBffd:ZNGd;dd?eJZPe-d@eIGdAdBeJZQe-dCeIGdDdEeJZRy)I)partial)CallableListOptionalTupleUnionN)nn)ACT2FN)Cache DynamicCacheSlidingWindowCache StaticCache)GenerationMixin)AttentionMaskConverter)FlashAttentionKwargs)BaseModelOutputWithPastCausalLMOutputWithPastMoeCausalLMOutputWithPastMoeModelOutputWithPastQuestionAnsweringModelOutput SequenceClassifierOutputWithPastTokenClassifierOutput)ROPE_INIT_FUNCTIONSdynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack) LossKwargsadd_code_sample_docstringsadd_start_docstrings%add_start_docstrings_to_model_forwardcan_return_tupleloggingreplace_return_docstrings)deprecate_kwarg) MixtralConfigzmistralai/Mixtral-8x7B-v0.1r(c*eZdZdeffd ZdZxZS)MixtralBlockSparseTop2MLPconfigct||j|_|j|_t j|j |jd|_t j|j|j d|_ t j|j |jd|_ t|j|_ yNFbias)super__init__intermediate_sizeffn_dim hidden_size hidden_dimr Linearw1w2w3r hidden_actact_fnselfr+ __class__s j/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/transformers/models/mixtral/modeling_mixtral.pyr1z"MixtralBlockSparseTop2MLP.__init__Gs //  ,,))DOOT\\F))DLL$//F))DOOT\\FV../ c|j|j||j|z}|j|}|SN)r;r7r9r8)r= hidden_statescurrent_hidden_statess r?forwardz!MixtralBlockSparseTop2MLP.forwardRs> $ DGGM,B CdggmF\ \ $(= >$$r@)__name__ __module__ __qualname__r(r1rE __classcell__r>s@r?r*r*Fs 0} 0%r@r*cZeZdZdZfdZdej dej fdZxZS)MixtralSparseMoeBlocka This implementation is strictly equivalent to standard MoE with full capacity (no dropped tokens). It's faster since it formulates MoE operations in terms of block-sparse operations to accommodate imbalanced assignments of tokens to experts, whereas standard MoE either (1) drop tokens at the cost of reduced performance or (2) set capacity factor to number of experts and thus waste computation and memory on padding. ct||j|_|j|_|j |_|j|_ tj|j|jd|_ tjt|jDcgc] }t|c}|_|j"|_ycc}wr-)r0r1r4r5r2r3num_local_experts num_expertsnum_experts_per_toktop_kr r6gate ModuleListranger*expertsrouter_jitter_noise jitter_noise)r=r+_r>s r?r1zMixtralSparseMoeBlock.__init__ds  ,,// !33// IIdoot/?/?eL }}QVW[WgWgQh%iA&?&G%ij #66&js+CrCreturncf|j\}}}|jrQ|jdkDrB|tj|j d|jz d|jzz}|j d|}|j|}tj|dtj}tj||jd\}}||jddz}|j|j}tj ||z|f|j|j" }tj$j&j)||j* j-d dd} t/|j*D]} |j0| } tj2| | \} } |d | fj5d|}| ||| | d fz}|j7d| |j|j|j5|||}||fS)  rg?r'dimdtyper^T)r^keepdim)r_device) num_classesN)shapetrainingrWtorch empty_likeuniform_viewrRFsoftmaxfloattopkrQsumtor_zerosrbr functionalone_hotrOpermuterTrUwherereshape index_add_)r=rC batch_sizesequence_lengthr5 router_logitsrouting_weightsselected_expertsfinal_hidden_states expert_mask expert_idx expert_layeridxtop_x current_staterDs r?rEzMixtralSparseMoeBlock.forwardss2?2E2E/ OZ ==T..2 U--m<EEcDL]L]F]_beievev_vw wM%**2z:  -0 ))Mq L,1JJ XZ,[))?..2t.DD),,]-@-@A#kk / ): 6m>Q>QZgZnZn  hh))112BPTP`P`1aiijkmnpqr   0 01 dJ<< 3L[%<=JC *$+6>>r:NM$0$?/RWY\^bRbBc$c ! * *1e5J5M5MmNaNa5b c d299*oWab"M11r@) rFrGrH__doc__r1rgTensorrErIrJs@r?rLrLXs(  7&2U\\&2ell&2r@rLc,eZdZdfd ZdZdZxZS)MixtralRMSNormct|tjt j ||_||_y)z= MixtralRMSNorm is equivalent to T5LayerNorm N)r0r1r Parameterrgonesweightvariance_epsilon)r=r4epsr>s r?r1zMixtralRMSNorm.__init__s1 ll5::k#:; #r@c"|j}|jtj}|j dj dd}|tj ||jzz}|j|j|zS)Nrdr\T)ra) r_rprgfloat32powmeanrsqrtrr)r=rC input_dtypevariances r?rEzMixtralRMSNorm.forwardsy#)) %((7  $$Q',,R,>% Ht?T?T4T(UU {{]--k:::r@c^t|jjd|jS)Nz, eps=)tuplerrerr=s r? extra_reprzMixtralRMSNorm.extra_reprs* ))*+6$2G2G1HIIr@)gư>)rFrGrHr1rErrIrJs@r?rrs$;Jr@rc|dd|jddzf}|d|jddzdf}tj| |fdS)z*Rotates half the hidden dims of the input..Nr\rdr`)rergcat)xx1x2s r? rotate_halfrsZ 3"!''"+"" " #B 3 q " " #B 99rc2YB ''r@c|j|}|j|}||zt||zz}||zt||zz}||fS)aApplies Rotary Position Embedding to the query and key tensors. Args: q (`torch.Tensor`): The query tensor. k (`torch.Tensor`): The key tensor. cos (`torch.Tensor`): The cosine part of the rotary embedding. sin (`torch.Tensor`): The sine part of the rotary embedding. position_ids (`torch.Tensor`, *optional*): Deprecated and unused. unsqueeze_dim (`int`, *optional*, defaults to 1): The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. Returns: `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. ) unsqueezer)qkcossin position_ids unsqueeze_dimq_embedk_embeds r?apply_rotary_pos_embrsY( -- &C -- &C3w;q>C/0G3w;q>C/0G G r@rCn_reprYc|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) r'N)reexpandrv)rCrbatchnum_key_value_headsslenhead_dims r? repeat_kvrso 2?1D1D.E h z!!Qa"23::5BUW\^bdlmM  (;e(CT8 TTr@modulequerykeyvalueattention_maskscalingdropoutc Tt||j}t||j} tj||j dd|z} |#|ddddddd|j df} | | z} t jj| dtjj|j} t jj| ||j} tj| | } | j ddj} | | fS)Nrdr r\r])prfr')rnum_key_value_groupsrgmatmul transposerer rrrlrrpr_rrf contiguous) rrrrrrrkwargs key_states value_states attn_weights causal_mask attn_outputs r?eager_attention_forwardrs3 ; ;s r?r1zMixtralAttention.__init__s,  " F4F4F&JdJd4de $*$>$>&B\B\$\!}}d* !'!9!9ii 2 2F4N4NQUQ^Q^4^ejk ii 2 2F4N4NQUQ^Q^4^ejk ii 2 2F4N4NQUQ^Q^4^ejk ii : :T]] JFL^L^ejk r@rCposition_embeddingsrpast_key_valuecache_positionrrYc |jdd}g|d|j}|j|j|j dd} |j |j|j dd} |j |j|j dd} |\} } t| | | | \} } |'| | |d}|j| | |j|\} } t}|jjdk7r^|jjdk(r(|jddrtjd nt |jj}||| | | |f|j"sd n |j$|j&t)|jd dd |\}}|j*g|dj-}|j/|}||fS) Nr\r'rd)rrreagersdpaoutput_attentionsFz`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.sliding_window)rrr)rerrrjrrrrupdaterrr+_attn_implementationgetlogger warning_oncerrfrrrrvrr)r=rCrrrrr input_shape hidden_shape query_statesrrrr cache_kwargsattention_interfacerrs r?rEzMixtralAttention.forward s$))#2. 88b8$--8 {{=166|DNNqRST [[/44\BLLQPQR {{=166|DNNqRST &S#7 jRUWZ#[ j  %#&snUL'5'<'d##L '>dkk>^>^&_#$7      % $}}C$2H2HLL"4;;0@$G %  % ! \*k));;;;FFH kk+. L((r@)NN)rFrGrHrr(intr1rgrrrr LongTensorrrrErIrJs@r?rrsG l} l l&+/59 0)||0)#5<<#=>0)!. 0) ! 0) !!1!12 0)-.0) u||Xell3XeELL>Q5RR S0)r@rceZdZdedeffd Z ddejdeejdeejdee ejdee d ee d ee d eejd ee ejejfd e e de ejee ejejfffdZxZS)MixtralDecoderLayerr+rct||j|_t|||_t ||_t|j|j|_ t|j|j|_ y)Nr) r0r1r4r self_attnrLblock_sparse_moer rms_norm_epsinput_layernormpost_attention_layernormrs r?r1zMixtralDecoderLayer.__init__=sm !--)&)< 5f =-f.@.@fFYFYZ(6v7I7IvObOb(c%r@rCrrrroutput_router_logits use_cacherrrrYc |} |j|}|jd|| ||||||d| \}} | |z}|} |j|}|j|\}} | |z}|f}|r|| fz }|r|| fz }|S)a Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`, *optional*): attention mask of size `(batch, sequence_length)` where padding elements are indicated by 0. past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_router_logits (`bool`, *optional*): Whether or not to return the logits of all the routers. They are useful for computing the router loss, and should not be returned during inference. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence. kwargs (`dict`, *optional*): Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code into the model )rCrrrrrrr)rrrr)r=rCrrrrrrrrrresidualself_attn_weightsrzoutputss r?rEzMixtralDecoderLayer.forwardGsF!,,]; ,:4>> , ' 3)%)/) ,  , ( (!=0 !55mD '+'<'<]'K$ } =0  "  )+ +G   ' 'Gr@)NNNFFFNN)rFrGrHr(rr1rgrrrrboolrr FloatTensorrErIrJs@r?rr<s5d}dd26378<,1/4$)59KOC||C!.Cu//0 C !u||!45 C $D> C'tnCD>C!!1!12C&eELL%,,,F&GHC-.C u  (51B1BEDUDU1U+V"WW XCr@rc^eZdZddeffd Zej edZxZ S)MixtralRotaryEmbeddingr+ct|t|drG|j;|jj d|jj d|_nd|_|j |_|j |_||_ t|j |_ |j|j|\}|_ |jd|d|j|_y)N rope_scaling rope_typetypedefaultinv_freqF) persistent)r0r1hasattrrrrmax_position_embeddingsmax_seq_len_cachedoriginal_max_seq_lenr+r rope_init_fnattention_scalingregister_bufferroriginal_inv_freq)r=r+rbrr>s r?r1zMixtralRotaryEmbedding.__init__s  6> *v/B/B/N#0044[&BUBUBYBYZ`BabDN&DN"("@"@$*$B$B! /?+/+<+>@GG HZHZ[\H]_acdehhijiqiqr ,QaZ 8 > > @'1!((--'E!((--[`J`ahhmmfk ^^ U C 5&,,.1F1L1L1NNYYZ[]^_E))UEN3C'')d444C'')d444C  5 vvAGGv$cff177f&;;;  5 5s BF%%F.rB) rFrGrHr(r1rgno_gradrrErIrJs@r?rrs3/}/"U]]_ < ..r@N)rFrGrHr( config_classbase_model_prefixsupports_gradient_checkpointing_no_split_modules_skip_keys_device_placement_supports_flash_attn_2_supports_sdpa_supports_flex_attn_supports_cache_class_supports_quantized_cache_supports_static_cache_supports_attention_backendr%rr@r?rrsU !L&*#./#4"5!N  $""& ?r@ra$ Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. If `past_key_values` is used, optionally only the last `input_ids` have to be input (see `past_key_values`). If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy. - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids) past_key_values (`Cache`, *optional*): Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values` returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`. It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache). If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, this tensor is not affected by padding. It is used to update the cache in the correct position and to infer the complete sequence length. cDeZdZdZdeffd ZdZdZee e  dde e jde e jde e jd e ee j d e e j d e ed e ed e ede ede e jdeedefdZ dde jde jde jd ed ef dZede jdedede j4de j6de jdeded efdZxZS) MixtralModelz Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`MixtralDecoderLayer`] Args: config: MixtralConfig r+c  t|||j|_|j|_t j |j|j|j|_t jt|jDcgc]}t||c}|_ t|j|j|_t#||_d|_|j)ycc}w)Nr)r+F)r0r1 pad_token_idr$ vocab_sizer r#r4 embed_tokensrSrTnum_hidden_layersrlayersrrnormr rotary_embgradient_checkpointing post_initrs r?r1zMixtralModel.__init__-s  !.. ++LL):):F?L] & #oom\J#7BD0d"6BD![[% :M#!m%55!**t}} $ A AM22H6GH! #%("' ! !.! !#.!-#2&7)='#1(; !( ! *!,M =#3"55#!mB&7%99!K% :N -0  -!1 1 %+++%+   r@ input_tensorc |jjdk(rS|H|F|dddfjj|j dk7}|r t d|d|vr|Sy||j nd}t|t}t|t} |jjdk(r?|s=| s;|s9tj||||jj|jry|j|j} } t!j"| j$} |j&d} | s|r|j)}n1t|t j*r|j&dn|| zdz}|j-|| || | ||j&d|j| }|jjdk(r2|0|jj.d vr|stj0|| }|S) Nflash_attention_2r\rzYou are attempting to perform batched generation with padding_side='right' this may lead to unexpected behaviour for Flash Attention version of Mixtral. Make sure to call `tokenizer.padding_side = 'left'` before tokenizing the input. rr)rEpast_key_values_lengthr is_trainingr')ry target_lengthr_rbrrxr+r)cudaxpu)r+rroitemsizerLrMrrrr_ignore_causal_mask_sdparrfr_rbrgfinfominreget_max_cache_shaper5_prepare_4d_causal_attention_mask_with_cache_positionr_unmask_unattended)r=rrXrrris_padding_rightrRusing_static_cacheusing_sliding_window_cacher_rb min_dtyperyr]rs r?rOz MixtralModel._update_causal_masks ;; + +/B B)o.I#1!R%#8#<#<#>#C#C#EIZIZI\]^I_#_ #$a )c^.C%% @O?Z?99;`a'E%/AS%T" KK , , 6'+E%%>>*'7#{{99 MM $**L,?,?vKK&** &,,Q/ %);+??AM nell;$$R(%7!; PP +')#))!,;;+Q   KK , , 6*%%**o=% 1CCKQZ[Kr@ryr]r_rbrxc p||jdk(r|} | Stj|j} tj||f| ||} tj |||j ddkD} |j]t|tr||kDrHtj |||j dd|jz k} | j| | | z} | ddddddfj|ddd} || j} |jd|kDr |ddd|f}|jd} | ddddddd| f|ddddddfj| jz}|dk(}| ddddddd| fj!|| | ddddddd| f<| S)aV Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing. Args: attention_mask (`torch.Tensor`): A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`. sequence_length (`int`): The sequence length being processed. target_length (`int`): The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet. dtype (`torch.dtype`): The dtype to use for the 4D attention mask. device (`torch.device`): The device to place the 4D attention mask on. cache_position (`torch.Tensor`): Indices depicting the position of the input sequence tokens in the sequence. batch_size (`torch.Tensor`): Batch size. config (`MixtralConfig`): The model's configuration class past_key_values (`Cache`): The cache class that is being used currently to generate N) fill_valuer_rbrIr\r'r)r^rgrcrdfullrNrvrrr bitwise_or_rclonererprb masked_fill)rryr]r_rbrrxr+rrrkdiagonal_attend_masksliding_attend_mask mask_length padding_masks r?rfzBMixtralModel._prepare_4d_causal_attention_mask_with_cache_position sH  %.*<*<*>!*C(K:7 E*..I** -0Ye\bK$)<< f#MP^PfPfgiklPm#m $$0"/3EF/\iJi*/,,}V*T&..r158M8MM+')445HI / /K%dD!Q&67>>z1bRTUK))//1 !''+m;%3A~ ~4E%FN,2226 *1aL[L+@ANSTVZ\`bcScDdDgDg&&E   ,q0 5@Aq,;,AV5W5c5c )6 Aq!\k\12r@) NNNNNNNNNN)F)rFrGrHrr(r1r@rCr#r"MIXTRAL_INPUTS_DOCSTRINGrrgrrrrrrrrrEr rO staticmethodrr_rbrfrIrJs@r?r3r3!s } !"*+CD151537=A59$(,0/3/359p E,,-p !.p u//0 p "$u'8'8"9: p   1 12 p D>p $D>p 'tnp 'tnp !!1!12p $$89p  p Ep p#( Q QllQ Q  Q  QfB BBB{{ B  B  BBBBBr@r3c eZdZy)KwargsForCausalLMN)rFrGrHrr@r?rzrzPsr@rz gate_logitsrOc|t|tsyt|trC|dj}tj|Dcgc]}|j |c}d}tj jjd}tj||d\}} tj jj| |} |>tj| jd} tj|d} n|j\} }|jd| |zz}|dddddddfj|| |||fjd||j }tj | j|zdtj |dz } |ddddddfj|| ||fjd|j |}tj ||zdtj |dz } tj | | j#dz}||zScc}w)a Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch. See Switch Transformer (https://arxiv.org/abs/2101.03961) for more details. This function implements the loss function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between experts is too unbalanced. Args: gate_logits: Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of shape [batch_size X sequence_length, num_experts]. num_experts: Number of experts top_k: The number of experts to route per-token, can be also interpreted as the `top-k` routing parameter. attention_mask (`torch.Tensor`, *optional*): The attention_mask used in forward function shape [batch_size X sequence_length] if not None. Returns: The auxiliary loss. Nrr`r\)rrrbrgrrpr rrrlrnrsrrmrerrvror)r{rOrQrcompute_device layer_gateconcatenated_gate_logitsr{rXr|r~tokens_per_expertrouter_prob_per_expertrxryr8expert_attention_mask router_per_expert_attention_mask overall_losss r?load_balancing_loss_funcrSs9:*[%"@+u%$Q..#(99^i-jPZjmmN.K-jpq#r hh))112JPR1SO**_eDA((%%--.> LK!JJ{'8'8':B"'O!C&4&:&:# O4::1=*B^_ 4AtT1 2 V& OUKX Y WR , R  "IIk&7&7&9>> from transformers import AutoTokenizer, MixtralForCausalLM >>> model = MixtralForCausalLM.from_pretrained("mistralai/Mixtral-8x7B-v0.1") >>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mixtral-8x7B-v0.1") >>> prompt = "Hey, are you conscious? Can you talk to me?" >>> inputs = tokenizer(prompt, return_tensors="pt") >>> # Generate >>> generate_ids = model.generate(inputs.input_ids, max_length=30) >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you." ```N) rDrrrrErrrFrr)lossaux_lossrrrCrKrzr)r+rrrFrrJrrslicer loss_functionr6rrzrOrPrrprbrrrCrK)r=rDrrrrErrrrFrrrrrrC slice_indicesrrrs r?rEzMixtralForCausalLM.forwardsh2C1N-TXT_T_TqTq$8$D $++JjJj  %9$D $++JjJj  +5$** + )%+'/!5!5) +  +  11 8B>SV8W~ot4]k mA}a,?@A  %4%%ffdooPPD /%%  (( H !11HKK 4LLL(#33!//))!//  r@) NNNNNNNNNNNr)"rFrGrH_tied_weights_keys_tp_plan_pp_planr1r@rCrrrrr#r&r"rwr%r_CONFIG_FOR_DOCrrgrrrrrrrrrzrrErIrJs@r?rrs*+=)H_-z:;H '(&)6DTU*+CD+AP_`151537=A59-1$(,0/3/35934d E,,-d !.d u//0 d "$u'8'8"9: d   1 12 d ))*d D>d $D>d 'tnd 'tnd !!1!12d c5<</0d *+d  #d aEVd r@ra The Mixtral Model transformer with a sequence classification head on top (linear layer). [`MixtralForSequenceClassification`] uses the last token in order to do the classification, as other causal models (e.g. GPT-2) do. Since it does classification on the last token, it requires to know the position of the last token. If a `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in each row of the batch). c(eZdZfdZdZdZeee dde e jde e jde e jde e de e jd e e jd e ed e ed e ed efdZxZS) MixtralForSequenceClassificationct|||j|_t||_t j |j|jd|_|jyr-) r0r1 num_labelsr3rr r6r4scorer=r<s r?r1z)MixtralForSequenceClassification.__init__CsS   ++!&) YYv114??O  r@c.|jjSrBrrs r?r@z5MixtralForSequenceClassification.get_input_embeddingsLrr@c&||j_yrBrrBs r?rCz5MixtralForSequenceClassification.set_input_embeddingsOrr@rDrrrrErrrrFrYc |j|||||||| } | j} |j| } ||jd} n|jd} |jj | dk7r t d|jj d}n|||jj k7j| jtj}tj|jd| jtj}||zjd}n.d}tj|jj d| tj| | j |f}d}||j#| |||j }t%||| j&| j(| j* S)  labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). rrrrErrrFNrr'z=Cannot handle batch sizes > 1 if no padding token is defined.r\)rbr_z will not detect padding tokens in `inputs_embeds`. Results may be unexpected if using padding tokens in conjunction with `inputs_embeds.`rI)rr pooled_logitsr+)rrrrCrK)rrJrrer+r5rLrprbrgint32rNargmaxrrr>rFrrrrCrK)r=rDrrrrErrrrFtransformer_outputsrCrrxlast_non_pad_token non_pad_mask token_indicesrrs r?rEz(MixtralForSequenceClassification.forwardRs*8!F!Fr!J !#    >>**+,ZZ  u||Jv}}MOaab   %%VFR_hlhshs%tD/ /??-;;*55   r@ NNNNNNNNN)rFrGrHr1r@rCr#r"rwrrgrrr rrrrErIrJs@r?rr3s '(*+CD151537+/59-1$(,0/3A E,,-A !.A u//0 A "% A   1 12 A ))*A D>A $D>A 'tnA  *A EA r@rz The Mixtral Model transformer with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. cDeZdZfdZdZdZeeee e e e  dde ejde ej de ejde ed e ej$d e ejd e ed e ed e ede fdZxZS)MixtralForTokenClassificationct|||j|_t||_t |dd |j }nt |dd |j}nd}tj||_ tj|j|j|_ |jy)Nclassifier_dropouthidden_dropoutg?)r0r1rr3rrrrr Dropoutrr6r4rr=)r=r+rr>s r?r1z&MixtralForTokenClassification.__init__s   ++!&) 6/ 6 B!'!:!:  V-t 4 @!'!6!6 !$ zz"45 YYv1163D3DE  r@c.|jjSrBrrs r?r@z2MixtralForTokenClassification.get_input_embeddingsrr@c&||j_yrBrrBs r?rCz2MixtralForTokenClassification.set_input_embeddingsrr@) checkpointrr&rDrrrrErrrrFrYc |j|||||||| } | j} |j| } |j| } d} ||j | ||j } t | | | j| jS)rrN)rrrCrK) rrJrrrr+rrCrK)r=rDrrrrErrrrFrsequence_outputrrs r?rEz%MixtralForTokenClassification.forwards4,0:: )%+'/!5,6 , "33,,7O,  %%ffdkkBD$!//))   r@r)rFrGrHr1r@rCr#r"rwr _CHECKPOINT_FOR_DOCrrrrgrrr rrrErIrJs@r?rrs '(*+CD&)$151537+/59-1$(,0/3* E,,-* !.* u//0 * "% *   1 12 * ))** D>* $D>* 'tn*  * E* r@rz The Mixtral Model transformer with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`). cdeZdZdZfdZdZdZeee  dde e jde e jde e jde eeee jfd e e jd e e jd e e jd e ed e edefdZxZS)MixtralForQuestionAnsweringrct||tj|jd|_t ||_|jy)Nrd) r0r1r r6r4 qa_outputsr3rr=r<s r?r1z$MixtralForQuestionAnswering.__init__s@  ))F$6$6:!&)  r@c.|jjSrBrrs r?r@z0MixtralForQuestionAnswering.get_input_embeddingsrr@c&||j_yrBrrBs r?rCz0MixtralForQuestionAnswering.set_input_embeddingsrr@rDrrrrEstart_positions end_positionsrrFrYc |j||||||| } | j} |j| } | jdd\}}|j dj }|j dj }d}|||j ||||fi| }t|||| j| jS)a start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for position (index) of the start of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for position (index) of the end of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence are not taken into account for computing the loss. )rrrrErrFr'r\r`N)r start_logits end_logitsrCrK) rrJrsplitsqueezerrrrCrK)r=rDrrrrErrrrFrrrrrrrs r?rEz#MixtralForQuestionAnswering.forwards4,0:: )%+'/!5,6, "331#)<<r<#: j#++B/::< ''+668   &=+D%4%%lJQ^ibhiD+%!!//))   r@r)rFrGrHr'r1r@rCr#r"rwrrgrrrr rrrrErIrJs@r?rrs  '(*+CD156:37KO596:48,0/33 E,,-3 !!2!233 u//0 3 "%tE4E4E/F(F"GH 3   1 12 3 "%"2"233   0 013 $D>3 'tn3  &3 E3 r@r)Nr')r)NrdN)S functoolsrtypingrrrrrrgtorch.nn.functionalr rrrk activationsr cache_utilsr r rr generationrmodeling_attn_mask_utilsrmodeling_flash_attention_utilsrmodeling_outputsrrrrrrrmodeling_rope_utilsrrmodeling_utilsrrprocessing_utilsrutilsrr r!r"r#r$r%utils.deprecationr&configuration_mixtralr( get_loggerrFrrrModuler*rLrrrrrrrmrrrrMIXTRAL_START_DOCSTRINGrrwr3rzrrrrrrr@r?rs699 !OO)>BLF&10   H %3!% %$A2BIIA2HJRYYJ((6 UU\\ U# U%,, U&% II% <<% % << % U\\* %  %%4A)ryyA)HN"))Nb "& -1 O&u||U5<<%8$>?O&#O&U\\* O&  5<<  O&dK /K \ S '=S S l H $:H H V F "8F F r@