gddlZddlmZmZmZmZddlZddlmZddl Zddl m Z ddl m Z mZmZddlmZddlmZddlmZmZmZmZdd lmZdd lmZmZmZmZm Z m!Z!d d l"m#Z#erdd l$m%Z%dd l$m%Z%Gddej&Z'Gddej&Z(Gddej&Z)e j*e+Z,dZ-d4dZ.dej/de0dej/fdZ1Gddej&Z2Gdde2Z3Gdde2Z4e2e3e4d Z5Gd!d"ej&Z6d#Z7ed$e7Gd%d&eZ8d'Z9d(Z:ed$e7Gd)d*e8Z;Gd+d,e8eZ<ed-e7Gd.d/e8Z=ed0e7Gd1d2e8Z>gd3Z?dS)5N)ListOptionalTupleUnion)ACT2FN)Cache DynamicCache StaticCache)GenerationMixin)AttentionMaskConverter)BaseModelOutputWithPastCausalLMOutputWithPast SequenceClassifierOutputWithPastTokenClassifierOutput)PreTrainedModel)add_start_docstrings%add_start_docstrings_to_model_forwardis_flash_attn_2_available#is_flash_attn_greater_or_equal_2_10loggingreplace_return_docstrings) GlmConfig)_flash_attention_forwardc,eZdZdfd ZdZdZxZS) GlmRMSNormư>cttjt j||_||_dS)z9 GlmRMSNorm is equivalent to T5LayerNorm N)super__init__nn Parametertorchonesweightvariance_epsilon)self hidden_sizeeps __class__s `/var/www/html/ai-engine/env/lib/python3.11/site-packages/transformers/models/glm/modeling_glm.pyr!zGlmRMSNorm.__init__:sD l5:k#:#:;; #c|j}|tj}|ddd}|tj||jzz}|j||zS)NT)keepdim) dtypetor$float32powmeanrsqrtr'r&)r( hidden_states input_dtypevariances r,forwardzGlmRMSNorm.forwardBs|#) %((77  $$Q'',,R,>>% Ht?T4T(U(UU {]--k::::r-cHt|jjd|jS)Nz, eps=)tupler&shaper'r(s r, extra_reprzGlmRMSNorm.extra_reprIs& )**II$2GIIIr-)r)__name__ __module__ __qualname__r!r;r@ __classcell__r+s@r,rr9sb$$$$$$;;;JJJJJJJr-rcReZdZdfd ZejddZxZS)GlmRotaryEmbedding'Nc:t||_||_||_d|jt jd|jdt j|jz zz }| d|ddS)Ng?rr/r2inv_freqF)tensor persistent) r r!dimmax_position_embeddingsbaser$arangeint64floatregister_buffer)r(rOrPrQdevicerLr+s r,r!zGlmRotaryEmbedding.__init__Ns '>$ $) Q!5;(W(W(W(](](_(_bfbj(jkl ZUKKKKKr-c"|j|j|jddddf|jddd}|dddddf}|jj}t|tr|dkr|nd}tj |d5||z dd}tj ||fd }| } |} dddn #1swxYwY| |j | |j fS) Nrr0rmpscpuF) device_typeenabledr/rOrK)rLr3rVrTexpandr>type isinstancestrr$autocast transposecatcossinr2) r(x position_idsseq_leninv_freq_expandedposition_ids_expandedrZfreqsembrdres r,r;zGlmRotaryEmbedding.forwardXs """ M$4-8>>@@GG HZ[\H]_acdee ,QQQaaaZ 8 > > @ @hm %/ S%A%AekUZFZFZkk`e ^ U C C C  &,,..1F1L1L1N1NNYYZ[]^__E)UEN333C''))C''))C                 vvAGv$$cff17f&;&;;;sA>EEE)rHrINN)rArBrCr!r$no_gradr;rDrEs@r,rGrGMskLLLLLLU]__<<<_<<<<V:V]bccc6#;V=OV[\\\#F$56r-r8returnc||}|dd\}}|||z}||S)Nr/r0r\)rwchunkrzrx)r(r8 up_statesgates r,r;zGlmMLP.forwardtsX%%m44 #//!/44i 2 24 8 88 ~~i(((r-)rArBrCr!r$ FloatTensorr;rDrEs@r,rprpjs`77777)U%6)5;L))))))))r-rpc|ddddf}|ddddf}tj| |fddS) z*Rotates half the hidden dims of the input..rNr/rr0r\)r$stackflatten)rfx1x2s r, rotate_halfrsQ 319B 319B ;Ryb ) ) ) 1 1" 5 55r-c||}||}|dd|jddzfdd}|dd|jddzfdd}|dd|jddzf|d|jddzdf}}|dd|jddzf|d|jddzdf}}||zt||zz}||zt||zz} t j||gd}t j| |gd} || 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. .Nr0r/r\) unsqueezer>repeat_interleaverr$rc) qkrdrerg unsqueeze_dimq_passk_passq_embedk_embeds r,apply_rotary_pos_embrs( -- & &C -- & &C c'SYr]a''' ( : :1" : E EC c'SYr]a''' ( : :1" : E EC#))))*Ac172;!3C3E3E.E,FvA#))))*Ac172;!3C3E3E.E,FvA3w;q>>C/0G3w;q>>C/0Gi&)r222Gi&)r222G G r-r8n_repr|c|j\}}}}|dkr|S|dddddddddf|||||}||||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>r]reshape)r8rbatchnum_key_value_headsslenhead_dims r, repeat_kvrs 2?1D.E h zz!!!!QQQaaa"23::5BUW\^bdlmmM  (;e(CT8 T TTr-cReZdZdZddedeeffd Z ddej deej d eej d ee d e d e d eej dee ej ej fde ej eej ee ej ffdZxZS) GlmAttentionz=Multi-headed attention from 'Attention Is All You Need' paperNrt layer_idxct||_||_|(td|jjd|j|_|j |_ |j |_ |j |j z|_ |j |_ |j |j z|_d|_dt!j|j z |_|j |j z|j kr t'd|j d|j dt)j|j |j |j z|j|_t)j|j |j |j z|j|_t)j|j |j |j z|j|_t)j|j |j d |_dS) NzInstantiating z without passing a `layer_idx` is not recommended and will lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` when creating this class.Trz?hidden_size must be divisible by num_heads (got `hidden_size`: z and `num_heads`: z).rrF)r r!rtrlogger warning_oncer+rAattention_dropoutr)num_attention_heads num_headsrrnum_key_value_groups is_causalmathsqrtscaling ValueErrorr"ruattention_biasq_projk_projv_projo_projr(rtrr+s r,r!zGlmAttention.__init__s  "     ,!8,,,    "(!9!-3(DN: #)#= $(Nd6N$N!49T]333 MDN *t/? ? ?8RVRb88%)^888  i 0$.4=2PW]Wlmmm i 0$2JT]2Zagavwww i 0$2JT]2Zagavwww i 0$2BOOO r-Fr8attention_maskrgpast_key_valueoutput_attentions use_cachecache_positionposition_embeddingsr|c  |\} } } ||} ||}||}| | | |j|jdd} || | |j|jdd}|| | |j|jdd}|\}}t| |||\} }|&|||d}| |||j |\}}t||j }t||j }tj| |dd|jz}|$|ddddddd|jdf}||z}t$j|dtj| j}t$j||j|j}tj||}|| |j| |jfkr5t7d | |j| |jfd ||dd}|| | d}||}|sd}|||fS) Nrr/rerdrrrr0)rOr2)ptrainingz `attn_output` should be of size z , but is )sizerrrviewrrrbrrupdaterrrr$matmulrr>r" functionalsoftmaxr4r3r2dropoutrrr contiguousr)r(r8rrgrrrrrkwargsbszq_len_ query_states key_states value_statesrdre cache_kwargs attn_weights causal_mask attn_outputs r,r;zGlmAttention.forwards&**,, UA{{=11 [[// {{=11 #((eT^T]SS]]^_abcc __S%1I4=YYccdeghii #((eT5Mt}]]gghiklmm &S#7 jRUWZ#[#[ j  %#&snUUL'5'<'>     #t~udm!L L L)CPTP]3^))$$&&))  "++Aq11<<>> !&&sE266 kk+..   LL.88r-rmNNNFFNN)rArBrC__doc__rrintr!r$Tensor LongTensorr boolrr;rDrEs@r,rrs=GGPPyPXc]PPPPPPD2637*."'59KO:9:9|:9!.:9u/0 :9 ! :9  :9:9!!12:9&eEL%,,F&GH:9 u|Xel3XeEL>Q5RR S:9:9:9:9:9:9:9:9r-rc:eZdZdZfdZ ddejdeejdeejdee d e d e d eejd ee ejejfd e ejeejee ejffdZ xZ S)GlmFlashAttention2aB Glm flash attention module. This module inherits from `GlmAttention` as the weights of the module stays untouched. The only required change would be on the forward pass where it needs to correctly call the public API of flash attention and deal with padding tokens in case the input contains any of them. cbtj|i|t |_dSrm)r r!r_flash_attn_uses_top_left_mask)r(argsrr+s r,r!zGlmFlashAttention2.__init__"s9$)&))) 3V2W2W.W+++r-NFr8rrgrrrrrr|c d}|\} } } ||} ||} ||}| | | |j|jdd} | | | |j|jdd} || | |j|jdd}|\}}t| | ||\} } |&|||d}| | ||j |\} }| dd} | dd} |dd}|j r|j nd}| j}|tjkrtjrtj}n3t'|jdr |jj}n|jjj}t.d|d| |} | |} ||}t5| | ||| |||jt9|d d|j|j }|| | d  }|!|}|sd}|||fS) NFrr/r_pre_quantization_dtypezThe input hidden states seems to be silently casted in float32, this might be related to the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in .sliding_window)rgr softmax_scaleruse_top_left_maskrr0)"rrrrrrrrbrrrrrrr2r$r4is_autocast_enabledget_autocast_gpu_dtypehasattrrtrr&rrr3rrgetattrrrrrr)r(r8rrgrrrrrrrrrrrrdrer dropout_rater9 target_dtyperrs r,r;zGlmFlashAttention2.forward*s"%**,, UA{{=11 [[// {{=11 $((eT^T]SS]]^_abcc __S%1I4=YYccdeghii #((eT5Mt}]]gghiklmm &S#7 jRUWZ#[#[ j  %#&snUUL'5'<'Q5RR SX9X9X9X9X9X9X9X9r-rc4eZdZdZ ddejdeejdeejdeede d e d eejd ee ejejfd e ejeejee ejfffd Z xZ S)GlmSdpaAttentionz Glm attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from `GlmAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to SDPA API. NFr8rrgrrrrrr|c |rCtdt||||||||S|\} } } ||} ||}||}| | | |j |j  dd} || | |j |j  dd}|| | |j |j  dd}|\}}t| |||\} }|&|||d}||||j|\}}t!||j}t!||j}|}||ddddddd|jdf}| jjdkr>|<| } |}|}|| dkrdnd }t,jj| ||||jr|jnd ||j }| dd}|| | d }||}|d|fS) NaGlmModel is using GlmSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.r8rrgrrrrrrr/rrcudaTFr) attn_mask dropout_prscaler0)rrr r;rrrrrrrrbrrrrrrr>rVr^rr$r"rscaled_dot_product_attentionrrrr)r(r8rrgrrrrrrrrrrrrrdrerrrrr+s r,r;zGlmSdpaAttention.forwards     [   77??+-)-"3#-$7#   &**,, UA{{=11 [[// {{=11 #((eT^T]SS]]^_abcc __S%1I4=YYccdeghii #((eT5Mt}]]gghiklmm &S#7 jRUWZ#[#[ j  %#&snUUL'5'<'> !&&sE266 kk+.. D.00r-r) rArBrCrr$rrrr rrr;rDrEs@r,rrs 2637*."'59KOP1P1|P1!.P1u/0 P1 ! P1  P1P1!!12P1&eEL%,,F&GHP1 u|Xel3XeEL>Q5RR SP1P1P1P1P1P1P1P1P1P1r-r)eagerflash_attention_2sdpac\eZdZddedeeffd Z ddejdeejdeej d ee d ee d ee d eej d ee ejejfde ej ee ej ej fffdZxZS)GlmDecoderLayerNrtrcJt|j|_t|j|||_t ||_t|j|j |_ t|j|j |_ dS)N)rtrr*) r r!r)GLM_ATTENTION_CLASSES_attn_implementation self_attnrpmlpr rms_norm_epsinput_layernormpost_attention_layernormrs r,r!zGlmDecoderLayer.__init__s !-.v/JKSYenooo&>>)&*<&BUVVV(263E6K^(_(_(_%%%r-Fr8rrgrrrrrr|c |} ||}|jd||||||||d| \}} } | |z}|} ||}||}| |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_size, sequence_length)` if flash attention is used or `(batch_size, 1, query_sequence_length, key_sequence_length)` if default attention is used. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. 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`). past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*): Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`, with `head_dim` being the embedding dimension of each attention head. kwargs (`dict`, *optional*): Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code into the model r)rrrr)r(r8rrgrrrrrrresidualself_attn_weightspresent_key_valueoutputss r,r;zGlmDecoderLayer.forwardsD!,,];; ?Mdn ? ')%)/) 3 ? ?  ? ? ; (*;!=0 !55mDD //  =0  "  , )+ +G  , )+ +Gr-rmr)rArBrCrrrr!r$rrr rrrr;rDrEs@r,rrsQ``y`Xc]``````2637*.,1$)59KOBB|B!.Bu/0 B ! B $D> BD>B!!12B&eEL%,,F&GHB u (51BEDU1U+V"WW XBBBBBBBBr-raH This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.) This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior. Parameters: config ([`GlmConfig`]): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights. zQThe bare Glm Model outputting raw hidden-states without any specific head on top.c@eZdZeZdZdZdgZdgZdZ dZ dZ dZ dZ dZdS)GlmPreTrainedModelmodelTrpast_key_valuesc|jj}t|tjrJ|jjd||j |jj dSdSt|tj rS|jjd||j -|jj|j  dSdSdS)Nr)r6std) rtinitializer_ranger_r"rur&datanormal_rszero_ Embedding padding_idx)r(modulers r, _init_weightsz GlmPreTrainedModel._init_weightsXsk+ fbi ( ( ? M  & &CS & 9 9 9{&  &&((((('&  - - ? M  & &CS & 9 9 9!- "6#56<<>>>>> ? ?--r-N)rArBrCr config_classbase_model_prefixsupports_gradient_checkpointing_no_split_modules_skip_keys_device_placement_supports_flash_attn_2_supports_sdpa_supports_cache_class_supports_quantized_cache_supports_static_cacherrr-r,rrHse L&*#*+#4"5!N  $! ? ? ? ? ?r-rra 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` or `tuple(tuple(torch.FloatTensor))`, *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`. Two formats are allowed: - a [`~cache_utils.Cache`] instance, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache); - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy cache format. The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the legacy cache format will be returned. 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. c eZdZdZdeffd ZdZdZee  dde j de e j d e e j d e eeee jfd e e jd e ed e ede ede ede e j deeeffdZde j de j de j d ed ef dZede j dedede jde jde j defdZxZS)GlmModelz Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`GlmDecoderLayer`] Args: config: GlmConfig rtctj|_j|_t jjj|j|_t j fdtj D|_ tjj|_t!jdzjj|_d|_|dS)Nc0g|]}t|Sr)r).0rrts r, z%GlmModel.__init__..s# a a aI_VY / / a a ar-rr/)rOrPrQF)r r! pad_token_idr vocab_sizer"r r) embed_tokens ModuleListrangenum_hidden_layerslayersrrnormrGrrP rope_theta rotary_embgradient_checkpointing post_initr{s `r,r!zGlmModel.__init__s    !. +L):F\cict   ',# r-c|jSrmr#r?s r,get_input_embeddingszGlmModel.get_input_embeddingss   r-c||_dSrmr.r(values r,set_input_embeddingszGlmModel.set_input_embeddingss!r-N input_idsrrgr inputs_embedsrroutput_hidden_states return_dictrr|c ||n |jj}||n |jj}||n |jj}| | n |jj} |du|duz rt d|jr%|jr|rt dd}|| |}d} |rVt|tsAd} |t}n.tj|}t d| B||nd} t!j| | |jdz|j} || d}|||| ||} |}|||}|rd nd}|rd nd}d}|jD]r}|r||fz }|jr*|jr#||j|| ||||| | }n||| ||||| | }|d}|r ||rd nd}|r ||dfz }s||}|r||fz }|r|nd}| r|}| st9d ||||fDSt;|||| S)Nz:You must specify exactly one of input_ids or inputs_embedszX`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`.FTzWe detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class (https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)rrrVr)rrgrrrrrr/c3K|]}||V dSrmr)rvs r, z#GlmModel.forward..Is(ttqfgfsfsfsfsfsttr-)last_hidden_staterr8 attentions)rtrr6ruse_return_dictrr+rrrr#r_r r from_legacy_cacheget_seq_lengthr$rRr>rVr_update_causal_maskr*r'_gradient_checkpointing_func__call__r(to_legacy_cacher=r)r(r4rrgrr5rrr6r7rreturn_legacy_cachepast_seen_tokensrr8rall_hidden_statesall_self_attnsnext_decoder_cache decoder_layer layer_outputs next_caches r,r;zGlmModel.forwards2C1N--TXT_Tq$8$D $+Jj "+!6IIDK?L]  & #oom\JJ#7@BBD0:d!![" 6" 6M# 6!m%55!* t}  $ A A!*! #%"' ! ! !. !#.!-#2&7'#1(; ! ! ! *!,M R%28I3P11q%Q"  6=#3"55 -00   2 -!1 1 +4>''$  6#3355J utt]J@QSa$bttttt t&+&+%     r- input_tensorc |jjdkr |d|vr|SdS||nd}t|t}|jjdkr#|s!|st j||||jrdS|j|j } }|j d} |r| } n/t|tj r |j dn|| zdz} ||| | || ||j d} |jjdkrB|@|j jd kr0|s.tj|j} t j| | } | S) Nrrrr)r5past_key_values_length is_trainingrr0)sequence_length target_lengthr2rVr batch_sizer)rtrrAr_r r _ignore_causal_mask_sdparr2rVr>get_max_cache_shaper$r5_prepare_4d_causal_attention_mask_with_cache_positionr^finfomin_unmask_unattended)r(rrNrrrrGusing_static_cacher2rVrRrSr min_dtypes r,rBzGlmModel._update_causal_maskQs ; +/B B B)c^.C.C%%4 @O?Z?99;;;`a'EE ; +v 5 5>P 5Yj 5%>*'7 M   t$*L,?v&,Q/  +??AAMMnel;;<$R((%7!; PP +')#)!,Q   K , 6 6*%*f44%5  E**.I0CKQZ[[Kr-rRrSr2rVrTc ||dkr|}n+tj|j} tj||f| ||}|dkrtj|d}|tj|||ddkz}|ddddddf|ddd}|| }|j d} |ddddddd| f|ddddddfz} | dk} |ddddddd| f | | |ddddddd| f<|S) a 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 plcae 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. N) fill_valuer2rVr)diagonalr9r0r) rOr$rXrYfulltriurRrr]cloner> masked_fill) rrRrSr2rVrrTrrr\ mask_length padding_masks r,rWz>GlmModel._prepare_4d_causal_attention_mask_with_cache_positionsB  %.*<*<*>*>!*C*C(KK E**.I* -0Ye\bK!###jqAAA 5< fEEEH^H^_acdHeHee eK%dD!!!QQQ&67>>z1bRTUUK))//11 ,226 *111aaaL[L+@ANSTSTSTVZ\`bcbcbcScDdd +q0 5@AAAqqq,;,AV5W5c5c )66 AAAqqq!!!\k\12r- NNNNNNNNNN)rArBrCrrr!r/r3rGLM_INPUTS_DOCSTRINGr$rrrrr rrrrrr;rB staticmethodrr2rVrWrDrEs@r,rrsE y$!!!"""+*+?@@'+1537KO59$(,0/3&*59x x #x !.x u/0 x "%tE4E/F(F"GH x   12 x D>x $D>x 'tnx d^x !!12x  u-- .x x x A@x t? ?l? ?  ?  ????B5 555{ 5  5  5555\55555r-rceZdZdgZdeffd ZdZdZdZdZ dZ d Z e e eee  dd ejdeejdeejdeeeeejfdeejdeejdeedeedeedeedeejdedeeeffdZxZS)GlmForCausalLMzlm_head.weightrtct|t||_|j|_t j|j|jd|_| dSNFrr) r r!rrr"r"rur)lm_headr,r{s r,r!zGlmForCausalLM.__init__sj    f%%  +y!3V5FUSSS  r-c|jjSrmrr#r?s r,r/z#GlmForCausalLM.get_input_embeddings z&&r-c||j_dSrmrpr1s r,r3z#GlmForCausalLM.set_input_embeddings"' r-c|jSrmrnr?s r,get_output_embeddingsz$GlmForCausalLM.get_output_embeddingss |r-c||_dSrmru)r(new_embeddingss r,set_output_embeddingsz$GlmForCausalLM.set_output_embeddingss % r-c||_dSrmr)r(decoders r, set_decoderzGlmForCausalLM.set_decoders  r-c|jSrmr{r?s r, get_decoderzGlmForCausalLM.get_decoders zr-) output_typerNrr4rrgrr5labelsrrr6r7rnum_logits_to_keepr|c ||n |jj}| | n |jj} | | n |jj} ||||||||| | |  }|d}||dd| dddf}d}||j|||jfi| }| s|f|ddz}||f|zn|St|||j |j |j S)a Args: labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., 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, ..., config.vocab_size]`. num_logits_to_keep (`int`, *optional*): Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that token can save memory, which becomes pretty significant for long sequences or large vocabulary size. Returns: Example: ```python >>> from transformers import AutoTokenizer, GlmForCausalLM >>> model = GlmForCausalLM.from_pretrained("google/glm-7b") >>> tokenizer = AutoTokenizer.from_pretrained("google/glm-7b") >>> prompt = "What is your favorite condiment?" >>> 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] "What is your favorite condiment?" ```N) r4rrgrr5rrr6r7rrrlosslogitsrr8r>) rtrr6r?rrn loss_functionr"rrr8r>)r(r4rrgrr5rrrr6r7rr loss_kwargsrr8rroutputs r,r;zGlmForCausalLM.forwardsK`2C1N--TXT_Tq$8$D $+Jj &1%T $D>T 'tnT d^T !!12T  T  u,, -T T T a`A@T T T T T r-rka The Glm Model transformer with a sequence classification head on top (linear layer). [`GlmForSequenceClassification`] uses the last token in order to do the classification, as other causal models (e.g. GPT-2) do. Since it does classification on the last token, it requires to know the position of the last token. If a `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in each row of the batch). c|eZdZdeffd ZdZdZee dde e j de e j de e j d e e eee jfd e e jd e e j d e ed e ede ede ede eeffdZxZS)GlmForSequenceClassificationrtct||j|_t||_t j|j|jd|_| dSrm) r r! num_labelsrrr"rur)scorer,r{s r,r!z%GlmForSequenceClassification.__init__Rsi     +f%% Yv14?OOO  r-c|jjSrmrpr?s r,r/z1GlmForSequenceClassification.get_input_embeddings[rqr-c||j_dSrmrpr1s r,r3z1GlmForSequenceClassification.set_input_embeddings^rsr-Nr4rrgrr5rrrr6r7r|c "| | n |jj} ||||||||| |  } | d} || } ||jd}n |jd}|jj|dkrt d|jjd}nv|rtj||jj ddz }||jdz}| | j }nd}| tj || j |f}d}||| |||j}| s|f| ddz}||f|zn|St||| j| j| j S)  labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). Nrrgrr5rrr6r7rrz=Cannot handle batch sizes > 1 if no padding token is defined.r0r9)rr pooled_logitsrtr)rtr?rrr>r!rr$eqrargmaxr3rVrRrrrr8r>)r(r4rrgrr5rrrr6r7transformer_outputsr8rrTsequence_lengthsrrrs r,r;z$GlmForSequenceClassification.forwardas(&1%E $D>E 'tnE d^E  u66 7E E E A@E E E E E r-rz The Glm Model transformer with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. cleZdZdeffd ZdZdZee dde e j de e j de e j d e e e jd e e jd e e j d e ed e ede ede edeeeffdZxZS)GlmForTokenClassificationrtct||j|_t||_t |dd|j}nt |dd|j}nd}tj ||_ tj |j |j|_ |dS)Nclassifier_dropouthidden_dropoutg?)r r!rrrrrrr"Dropoutrrur)rr,)r(rtrr+s r,r!z"GlmForTokenClassification.__init__s     +f%% 6/ 6 6 B!'!:   V-t 4 4 @!'!6  !$ z"455 Yv163DEE  r-c|jjSrmrpr?s r,r/z.GlmForTokenClassification.get_input_embeddingsrqr-c||j_dSrmrpr1s r,r3z.GlmForTokenClassification.set_input_embeddingsrsr-Nr4rrgrr5rrrr6r7r|c p| | n |jj} ||||||||| |  } | d} || } || } d}||| ||j}| s| f| ddz}||f|zn|St || | j| jS)rNrrr/)rrr8r>) rtr?rrrrrr8r>)r(r4rrgrr5rrrr6r7rsequence_outputrrrs r,r;z!GlmForTokenClassification.forwards(&1%1 $D>1 'tn1 d^1  u++ ,1 1 1 A@1 1 1 1 1 r-r)rrrkrr)Nr)@rtypingrrrrr$torch.nnr"torch.utils.checkpoint activationsr cache_utilsr r r generationr modeling_attn_mask_utilsr modeling_outputsrrrrmodeling_utilsrutilsrrrrrrconfiguration_glmrmodeling_flash_attention_utilsrModulerrGrp get_loggerrArrrrrrrrrrrGLM_START_DOCSTRINGrrrhrrkrr__all__rr-r,rs5, //////////// !!!!!!;;;;;;;;;;))))))>>>>>> .-----)(((((KJJJJJJFFFFFFJJJJJJJJ(<<<<<<<<:)))))RY)))&  H % %666&&&&R UU\ U# U%, U U U U\9\9\9\9\929\9\9\9~g9g9g9g9g9g9g9g9TX1X1X1X1X1|X1X1X1x+ MMMMMbiMMM`"W??????? ?0HVWRRRRR!RR Rjt t t t t 't t t n V V V V V #5V V V r I I I I I 2I I I X   r-