U 5AfI@s&dZddlZddlmZmZmZmZddlZddlZddlm Z ddl m Z m Z m Z ddlmZddlmZmZmZmZdd lmZdd lmZmZmZmZdd lmZdd lmZm Z m!Z!m"Z"m#Z#m$Z$d dl%m&Z&e!rddl'm(Z(e#)e*Z+dZ,Gddde j-Z.Gddde j-Z/ddZ0d6ddZ1Gddde j-Z2ej3e4ej3dddZ5Gddde j-Z6Gd d!d!e6Z7Gd"d#d#e6Z8e6e7e8d$Z9Gd%d&d&e j-Z:d'Z;ed(e;Gd)d*d*eZGd.d/d/e<Z?ed0e;Gd1d2d2e<Z@ed3e;Gd4d5d5e<ZAdS)7zPyTorch Mistral model.N)ListOptionalTupleUnion)nn)BCEWithLogitsLossCrossEntropyLossMSELoss)ACT2FN)Cache DynamicCacheSlidingWindowCache StaticCache)AttentionMaskConverter)BaseModelOutputWithPastCausalLMOutputWithPast SequenceClassifierOutputWithPastTokenClassifierOutput)PreTrainedModel)add_start_docstrings%add_start_docstrings_to_model_forwardis_flash_attn_2_available#is_flash_attn_greater_or_equal_2_10loggingreplace_return_docstrings) MistralConfig)_flash_attention_forwardrcs.eZdZdfdd ZddZddZZS) MistralRMSNormư>cs&ttt||_||_dS)z= MistralRMSNorm is equivalent to T5LayerNorm N)super__init__r ParametertorchZonesweightvariance_epsilon)self hidden_sizeeps __class__P/tmp/pip-unpacked-wheel-zw5xktn0/transformers/models/mistral/modeling_mistral.pyr"=s zMistralRMSNorm.__init__cCsJ|j}|tj}|djddd}|t||j}|j||S)NT)Zkeepdim) dtypetor$float32powmeanZrsqrtr&r%)r' hidden_states input_dtypeZvariancer,r,r-forwardEs  zMistralRMSNorm.forwardcCst|jjd|jS)Nz, eps=)tupler%shaper&r'r,r,r- extra_reprLszMistralRMSNorm.extra_repr)r )__name__ __module__ __qualname__r"r7r; __classcell__r,r,r*r-r<srcs.eZdZdfdd ZeddZZS) MistralRotaryEmbedding'Ncs`t||_||_||_d|jtjd|jdtjd ||j}|j d|dddS)Ng?rr.r0inv_freqF) persistent) r!r"dimmax_position_embeddingsbaser$arangeZint64floatr1Zregister_buffer)r'rFrGrHdevicerDr*r,r-r"Qs  0zMistralRotaryEmbedding.__init__c Cs|jddddf|jddd}|dddddf}|jj}t|tr`|dkr`|nd}tj |dd@|| dd}tj ||fdd }| }| } W5QRX|j|jd | j|jd fS) Nrr/rZmpscpuF) device_typeZenabledr.rFrC)rDrJexpandr9rKtype isinstancestrr$Zautocast transposecatcossinr1r0) r'x position_idsZinv_freq_expandedZposition_ids_expandedrMZfreqsZembrUrVr,r,r-r7Zs(zMistralRotaryEmbedding.forward)rArBN)r<r=r>r"r$Zno_gradr7r?r,r,r*r-r@Ps r@cCsH|dd|jddf}|d|jdddf}tj| |fddS)z*Rotates half the hidden dims of the input..Nr/r.rN)r9r$rT)rWx1Zx2r,r,r- rotate_halfnsrZcCsD||}||}||t||}||t||}||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. ) unsqueezerZ)qkrUrVrXZ unsqueeze_dimZq_embedZk_embedr,r,r-apply_rotary_pos_embvs   r^cs$eZdZfddZddZZS) MistralMLPcslt|j|_|j|_tj|j|jdd|_tj|j|jdd|_tj|j|jdd|_t |j |_ dSNFbias) r!r"r(Zintermediate_sizerLinear gate_projup_proj down_projr Z hidden_actact_fnr'configr*r,r-r"s zMistralMLP.__init__cCs ||||||SN)rfrgrdre)r'Z hidden_stater,r,r-r7szMistralMLP.forward)r<r=r>r"r7r?r,r,r*r-r_s r_)r5n_repreturncCs^|j\}}}}|dkr|S|dddddddddf|||||}||||||S)z This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) rN)r9rOreshape)r5rkbatchnum_key_value_headsslenhead_dimr,r,r- repeat_kvs 0rrcseZdZdZd eeedfdd Zd ej eej eej ee e e eej e ej eej ee ej fddd ZZS) MistralAttentionz Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer and "Generating Long Sequences with Sparse Transformers". Nri layer_idxcs t||_||_|dkr4td|jjd|j|_|j |_ |j |_ |j |_ |j |_ |j |j |_|j|_|j|_d|_tj|j |j |j dd|_tj|j |j |j dd|_tj|j |j |j dd|_tj|j |j |j dd|_t|j |j|jd|_dS)NzInstantiating z without passing a `layer_idx` is not recommended and will lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` when creating this class.TFra)rGrH)r!r"rirulogger warning_oncer+r<attention_dropoutr(Znum_attention_heads num_headsrqronum_key_value_groupsrGZ rope_theta is_causalrrcq_projk_projv_projo_projr@ rotary_embr'rirur*r,r-r"s2 zMistralAttention.__init__Fr5attention_maskrXpast_key_valueoutput_attentions use_cachecache_positionrlcCs|\}} } ||} ||} ||} | || |j|jdd} | || |j|jdd} | || |j|jdd} | | |\}}t | | ||\} } |dk r|||d}| | | |j |\} } t | |j} t | |j} t| | ddt|j}|dk r<|ddddddd| jdf}||}tjj|dtjd| j}tjj||j|jd}t|| }|||j| |jfkrtd ||j| |jfd ||dd}||| d}||}|sd}|||fS) Nrr.rVrUrr r/)rFr0)ptrainingz `attn_output` should be of size z , but is ) sizer|r}r~viewryrqrSrorr^updaterurrrzr$matmulmathsqrtr9r functionalZsoftmaxr2r1r0dropoutrxr ValueError contiguousr)r'r5rrXrrrrbszq_len_ query_states key_states value_statesrUrV cache_kwargs attn_weights causal_mask attn_outputr,r,r-r7s>         &  zMistralAttention.forward)N)NNNFFN)r<r=r>__doc__rrintr"r$Tensor LongTensorr boolrr7r?r,r,r*r-rss$#rsc sTeZdZdZfddZd ejeejeejee e e eejdddZ Z S) MistralFlashAttention2aJ Mistral flash attention module. This module inherits from `MistralAttention` 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. cstj||t |_dSrj)r!r"r_flash_attn_uses_top_left_mask)r'argskwargsr*r,r-r"szMistralFlashAttention2.__init__NFr5rrXrrrrc Cs>t|trtdd}|\}} } ||} ||} ||} | || |j|j  dd} | || |j |j  dd} | || |j |j  dd} | j d}|dk r||d7}| | |\}}t| | ||\} } |dk r||jdk}t|jdddk r||jjkr|rd|jj}||jd}||jd}|dddd|dddf}|dddd|dddf}|j d|jjdkrtd|j |dk r|dd|df}tj|t|ddd dfgd d }||d }|| | |j|\} } t| |j} t| |j} |js:d n|j}| j}|tjkrtrft }n"t!|jd r~|jj"}n |jj#j}t$%d|d| &|} | &|} | &|} | dd} | dd} | dd} t'| | | || ||t|jdd|j(|j)d }|*|| |j|j }|+|}|s4d}|||fS)Nz`static` cache implementation is not compatible with `attn_implementation==flash_attention_2` make sure to use `sdpa` in the mean time, and open an issue at https://github.com/huggingface/transformersFrr.rrsliding_windowzepast key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got r/rN)rVrU_pre_quantization_dtypezThe input hidden states seems to be silently casted in float32, this might be related to the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in .)rXrrZuse_top_left_maskr{),rQrrrr|r}r~rryrqrSror9rr^get_seq_lengthrugetattrrirrr$rTZ ones_likerrrrzrrxr0r2Zis_autocast_enabledZget_autocast_gpu_dtypehasattrrr%rvrwr1rrr{rmr)r'r5rrXrrrrrrrrrrZ kv_seq_lenrUrVZcache_has_contentsZslicing_tokensZpast_keyZ past_valuerZ dropout_rater6Z target_dtyperrr,r,r-r7s         $$  (                  zMistralFlashAttention2.forward)NNNFFN) r<r=r>rr"r$rrrr rr7r?r,r,r*r-r s" rcsjeZdZdZdejeejeejeee e eeje ejeejee ejfdfdd Z Z S)MistralSdpaAttentionz Mistral attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from `MistralAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to SDPA API. NFrc s|r(tdtj|||||||dS|\} } } ||} ||} ||}| | | |j |j  dd} | | | |j |j  dd} || | |j |j  dd}| ||\}}t| | ||\} } |dk r|||d}|| ||j|\} }t| |j} t||j}|}|dk r@|ddddddd| jdf}| jjdkrp|dk rp| } | } |}|dkr| dkrdnd }tjjj| | |||jr|jnd |d }| dd}|| | d }||}|d|fS) NaMistralModel is using MistralSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.rrr.rrcudaTFr)Z attn_maskZ dropout_pr{r/)rvrwr!r7rr|r}r~rryrqrSrorr^rrurrrzr9rKrPrr$rrZscaled_dot_product_attentionrrxr)r'r5rrXrrrrrrrrrrrrUrVrrr{rr*r,r-r7s^         &  zMistralSdpaAttention.forward)NNNFFN) r<r=r>rr$rrrr rrr7r?r,r,r*r-rs" r)eagerflash_attention_2sdpacs|eZdZeedfdd Zd ejeejeej ee ee ee eej e ej ee ej ej ffdddZZS) MistralDecoderLayerrtcsXt|j|_t|j||d|_t||_t|j|j d|_ t|j|j d|_ dS)Nrtr)) r!r"r(MISTRAL_ATTENTION_CLASSES_attn_implementation self_attnr_mlpr rms_norm_epsinput_layernormpost_attention_layernormrr*r,r-r"s   zMistralDecoderLayer.__init__NFrc Ks|} ||}|jf|||||||d|\}} } | |}|} ||}||}| |}|f} |rp| | f7} |r~| | f7} | S)a Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`, *optional*): attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1, query_sequence_length, key_sequence_length)` if default attention is used. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence kwargs (`dict`, *optional*): Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code into the model r)rrrr) r'r5rrXrrrrrZresidualZself_attn_weightsZpresent_key_valueoutputsr,r,r-r7s0      zMistralDecoderLayer.forward)NNNFFN)r<r=r>rrr"r$rrrr rr FloatTensorr7r?r,r,r*r-rs" raL 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 ([`MistralConfig`]): 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. zUThe bare Mistral Model outputting raw hidden-states without any specific head on top.c@s:eZdZeZdZdZdgZdZdZ dZ dZ dZ ddZ dS)MistralPreTrainedModelmodelTrpast_key_valuescCs||jj}t|tjr>|jjjd|d|jdk rx|jj n:t|tj rx|jjjd|d|j dk rx|jj|j  dS)Nr)r4std) riZinitializer_rangerQrrcr%dataZnormal_rbZzero_ Embedding padding_idx)r'modulerr,r,r- _init_weightsbs    z$MistralPreTrainedModel._init_weightsN)r<r=r>r config_classZbase_model_prefixZsupports_gradient_checkpointingZ_no_split_modulesZ_skip_keys_device_placementZ_supports_flash_attn_2Z_supports_sdpaZ_supports_cache_classZ_supports_static_cacherr,r,r,r-rSsraL 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 `decoder_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; - 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. cseZdZdZedfdd ZddZddZee de j e e j e e j e eeee jfe e je ee ee ee ee e j eeefd d d Ze j e j e j eeed ddZZS) MistralModelz Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`MistralDecoderLayer`] Args: config: MistralConfig rics~tj|_j|_tjj|j|_t fddt j D|_ j |_ tjjd|_d|_|dS)Ncsg|]}t|qSr,)r).0rurr,r- sz)MistralModel.__init__..rF)r!r" pad_token_idr vocab_sizerrr( embed_tokensZ ModuleListrangeZnum_hidden_layerslayersrrrnormgradient_checkpointing post_initrhr*rr-r"s zMistralModel.__init__cCs|jSrjrr:r,r,r-get_input_embeddingssz!MistralModel.get_input_embeddingscCs ||_dSrjrr'valuer,r,r-set_input_embeddingssz!MistralModel.set_input_embeddingsN) input_idsrrXr inputs_embedsrroutput_hidden_states return_dictrrlc  CsH|dk r |n|jj}|dk r |n|jj}|dk r4|n|jj}| dk rH| n|jj} |dk|dk Arhtd|jr|jr|rt dd}|dkr| |}d} |rt |t s|jst |}d} t d| dkr|dk r|nd} tj| | |jd|jd} |dkr| d}|||| |||} |}|r8d nd}|rFd nd}d}|jD]}|rh||f7}|jr|jr||j|| ||||| }n||| ||||| d }|d}|r||rd nd}|rT||df7}qT||}|r||f7}|r|nd}| r|}| s8td d ||||fDSt||||dS)NzaYou cannot specify both input_ids and inputs_embeds at the same time, and must specify either onezZ`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...FTzWe detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)rrrKr,)rrXrrrrr.css|]}|dk r|VqdSrjr,)rvr,r,r- Esz'MistralModel.forward..)Zlast_hidden_staterr5 attentions)rirrruse_return_dictrrrrvrwrrQr r Zfrom_legacy_cacherr$rIr9rKr[_update_causal_maskrZ_gradient_checkpointing_func__call__rZto_legacy_cacher8r)r'rrrXrrrrrrrZreturn_legacy_cachepast_seen_tokensrr5Zall_hidden_statesZall_self_attnsZnext_decoder_cacheZ decoder_layerZ layer_outputsZ next_cacher,r,r-r7s            zMistralModel.forward)r input_tensorrrrrcCs|jdkr^|dk rF|rF|dddf|dk}|rFtd|dk rZd|krZ|SdS|dk rn|dnd}t|t} t|t} |jjdkr| s| s|st j ||||jj |j drdS|j |j} } t| j} |jd}| rt||jj }n4| r|}n$t|tjr|jdn ||d}|dk r^|d kr^|dkrVtd |}nZtj||f| | | d }tj|| d |ddk}|jj dk r| r||jj kr|tj|| d |dd|jj k||9}|ddddddf|jdddd}|dk r|}|d kr|jd}|ddddddd|f|ddddddf}|dk}|ddddddd|f|| |ddddddd|f<|jjdkr|dk r|jjdkr|st || }|S)Nrr/rzYou are attempting to perform batched generation with padding_side='right' this may lead to unexpected behaviour for Flash Attention version of Mistral. Make sure to call `tokenizer.padding_side = 'left'` before tokenizing the input. rr)rZpast_key_values_lengthrZ is_trainingrzGCustom 4D attention mask should be passed in inverted form with max==0`)Z fill_valuer0rKrr.r) rsumitemrrrQrrrirZ_ignore_causal_mask_sdparrr0rKr$Zfinfominr9maxZget_max_lengthrrFfullrIrmZ bitwise_or_rOcloneZ masked_fillrPZ_unmask_unattended)r'rrrrrrZis_padding_rightrZusing_static_cacheZusing_sliding_window_cacher0rKZ min_dtypeZsequence_lengthZ target_lengthrZ exclude_maskZ mask_lengthZ padding_maskr,r,r-rMs  $         *  8     z MistralModel._update_causal_mask) NNNNNNNNNN)r<r=r>rrr"rrrMISTRAL_INPUTS_DOCSTRINGr$rrrrr rrrrrr7rr?r,r,r*r-rsF xrcseZdZdgZfddZddZddZdd Zd d Zd d Z ddZ e e e eeddejeejeejeeeeejfeejeejeeeeeeeeeejeeefd ddZdddZZS)MistralForCausalLMzlm_head.weightcs@t|t||_|j|_tj|j|jdd|_| dSr`) r!r"rrrrrcr(lm_headrrhr*r,r-r"s   zMistralForCausalLM.__init__cCs|jjSrjrrr:r,r,r-rsz'MistralForCausalLM.get_input_embeddingscCs ||j_dSrjrrr,r,r-rsz'MistralForCausalLM.set_input_embeddingscCs|jSrjrr:r,r,r-get_output_embeddingssz(MistralForCausalLM.get_output_embeddingscCs ||_dSrjr)r'Znew_embeddingsr,r,r-set_output_embeddingssz(MistralForCausalLM.set_output_embeddingscCs ||_dSrjr)r'decoderr,r,r- set_decoderszMistralForCausalLM.set_decodercCs|jSrjrr:r,r,r- get_decoderszMistralForCausalLM.get_decoder) output_typerN) rrrXrrlabelsrrrrrrlc  Cs,|dk r |n|jj}| dk r | n|jj} | dk r4| n|jj} |j|||||||| | | d } | d} || }|}d}|dk r|dddddf}|dddf}|d|jj }|d}| |j }t }|||}| s|f| dd}|dk r|f|S|St ||| j| j| jdS)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]`. Returns: Example: ```python >>> from transformers import AutoTokenizer, MistralForCausalLM >>> model = MistralForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1") >>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-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) rrrXrrrrrrrr.r/rlosslogitsrr5r)rirrrrrrJrrrr1rKrrrr5r)r'rrrXrrrrrrrrrr5rrZ shift_logitsZ shift_labelsloss_fctoutputr,r,r-r7sL*     zMistralForCausalLM.forwardTc Ks|dk rR|dk r.|dd|jd df}n$|jd|jdkrR|dd|f}|dk r|dkr|dd}||dkd|r|dd|jd df}|jtjd}|dk r|ddkrd|i} n d|i} | |||||d| S)Nrrr/)Z memory_formatrr)rXrrrr) r9longZcumsumZ masked_fill_rr$Zcontiguous_formatrr) r'rrrrrrXrrZ model_inputsr,r,r-prepare_inputs_for_generation3s.   z0MistralForCausalLM.prepare_inputs_for_generation) NNNNNNNNNNN)NNNNNT)r<r=r>Z_tied_weights_keysr"rrrrrrrrrr_CONFIG_FOR_DOCr$rrrrr rrrrr7rr?r,r,r*r-rsT   \ra The Mistral Model transformer with a sequence classification head on top (linear layer). [`MistralForSequenceClassification`] 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). cseZdZfddZddZddZeed ee j ee j ee j ee e ee jfee jee j eeeeeeeee eefd d d ZZS) MistralForSequenceClassificationcs@t||j|_t||_tj|j|jdd|_| dSr`) r!r" num_labelsrrrrcr(scorerrhr*r,r-r"ts   z)MistralForSequenceClassification.__init__cCs|jjSrjrr:r,r,r-r}sz5MistralForSequenceClassification.get_input_embeddingscCs ||j_dSrjrrr,r,r-rsz5MistralForSequenceClassification.set_input_embeddingsN rrrXrrrrrrrrlc  Cs:| dk r | n|jj} |j|||||||| | d } | d} || } |dk rV|jd}n |jd}|jjdkr||dkr|td|jjdkrd}nF|dk rt||jj dd}||jd}| | j }nd}| tj || j d|f}d}|dk r| | j }|jjdkrd|jdkr*d|j_n:|jdkr\|jtjksR|jtj kr\d |j_nd |j_|jjdkrt}|jdkr|||}n |||}nN|jjd krt}||d|j|d}n|jjd krt}|||}| s"|f| dd}|dk r|f|S|St||| j| j| jd 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). NrrXrrrrrrrrz=Cannot handle batch sizes > 1 if no padding token is defined.r/rZ regressionZsingle_label_classificationZmulti_label_classificationr)rirrrr9rrr$eqrZargmaxr1rKrIZ problem_typerr0rr Zsqueezerrrrrr5r)r'rrrXrrrrrrrZtransformer_outputsr5rZ batch_sizeZsequence_lengthsZ pooled_logitsrrrr,r,r-r7sr         (    z(MistralForSequenceClassification.forward) NNNNNNNNNN)r<r=r>r"rrrrrr$rrrr rrrrrr7r?r,r,r*r-rcs6  rz The Mistral 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. cseZdZfddZddZddZeed ee j ee j ee j ee e j ee j ee j eeeeeeeeeeefd d d ZZS) MistralForTokenClassificationcs|t||j|_t||_t|dddk r6|j}nt|dddk rN|j}nd}t ||_ t |j |j|_ |dS)Nclassifier_dropouthidden_dropoutg?)r!r"rrrrr rrZDropoutrrcr(rr)r'rir r*r,r-r"s   z&MistralForTokenClassification.__init__cCs|jjSrjrr:r,r,r-rsz2MistralForTokenClassification.get_input_embeddingscCs ||j_dSrjrrr,r,r-rsz2MistralForTokenClassification.set_input_embeddingsNrc  Cs| dk r | n|jj} |j|||||||| | d } | d} || } || } d}|dk rxt}|| d|j|d}| s| f| dd}|dk r|f|S|St|| | j | j dS)r Nr rr/r.)rrr5r) rirrrrrrrrr5r)r'rrrXrrrrrrrrZsequence_outputrrrrr,r,r-r7s8   z%MistralForTokenClassification.forward) NNNNNNNNNN)r<r=r>r"rrrrrr$rrrrrrrrr7r?r,r,r*r-r s6   r )Nr)Brrtypingrrrrr$Ztorch.utils.checkpointrZtorch.nnrrr Z activationsr Z cache_utilsr r rrZmodeling_attn_mask_utilsrZmodeling_outputsrrrrZmodeling_utilsrutilsrrrrrrZconfiguration_mistralrZmodeling_flash_attention_utilsrZ get_loggerr<rvrModulerr@rZr^r_rrrrrsrrrrZMISTRAL_START_DOCSTRINGrrrrrr r,r,r,r-st          a Y KF* l