a h[I @sVddlmZmZmZmZmZddlZddlmm Z ddlmZddl m Z ddl mZddlmZddlmZdd lmZdd lmZmZmZdd lmZmZdd lmZmZdd lm Z ddl!m"Z"m#Z#m$Z$m%Z%ddl&m'Z'ddl(m)Z)m*Z*ddl+m,Z,e*r$ddl-m.Z.ddl/m0Z0m1Z1ndZ.e)rBddl2m3Z3m4Z4nd\Z4Z3e$rjddl5m6Z6ddl7m8Z8e%9e:Z;ddZe?ej>dddZ@dSejAej>ej>ej>eej>eBeBd!d"d#ZCGd$d%d%ejAZDGd&d'd'ZEej>e?d(d)d*ZFd+d,ZGd-d.ZHeIe.e3e4fZJd/d0ZKGd1d2d2ejAZLGd3d4d4ejjAZMGd5d6d6ejAZNGd7d8d8ed9d:ZOGd;d<dd>ejAZQGd?d@d@ejAZRGdAdBdBejAZSGdCdDdDeZTe"GdEdFdFeZUGdGdHdHejAZVe"GdIdJdJeUZWdTeej>eXej>dfee?eej>eej>e?fdLdMdNZYGdOdPdPeUeZZgdQZ[dS)U)AnyCallableOptional TypedDictUnionN)nn)ACT2FN)Cache)GenerationMixin)AttentionMaskConverter)GradientCheckpointingLayer)BaseModelOutputWithPastMoeCausalLMOutputWithPastMoeModelOutputWithPast)ROPE_INIT_FUNCTIONSdynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)auto_docstringcan_return_tupleis_torch_flex_attn_availablelogging)deprecate_kwarg)is_causal_conv1d_availableis_mamba_2_ssm_available)GraniteMoeHybridConfig)selective_state_update)mamba_chunk_scan_combined mamba_split_conv1d_scan_combined)causal_conv1d_fncausal_conv1d_updateNN) BlockMask)make_flex_block_causal_maskcCsH|dd|jddf}|d|jdddf}tj| |fddS)z*Rotates half the hidden dims of the input..Ndim)shapetorchcat)xx1Zx2r0z/var/www/html/assistant/venv/lib/python3.9/site-packages/transformers/models/granitemoehybrid/modeling_granitemoehybrid.py rotate_halfAsr2cCsD||}||}||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. ) unsqueezer2)qkcossin position_idsZ unsqueeze_dimZq_embedZk_embedr0r0r1apply_rotary_pos_embHs   r9) hidden_statesn_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)r+expandreshape)r:r;batchnum_key_value_headsslenhead_dimr0r0r1 repeat_kvcs 0rC)modulequerykeyvalueattention_maskscalingdropoutc Kst||j}t||j} t||dd|} |durf|ddddddd|jdf} | | } tjj| dtj d |j } tjj | ||j d} t| | } | dd} | | fS)Nr(r r')r*dtype)ptrainingr)rCnum_key_value_groupsr,matmul transposer+r functionalsoftmaxfloat32torMrKrO contiguous) rErFrGrHrIrJrKkwargs key_states value_states attn_weights causal_mask attn_outputr0r0r1eager_attention_forwardos  & r^cseZdZdZeedfdd Zedddddej e ej e ej e e e e ej e eej ej feej e ej e eej fd d d ZZS)GraniteMoeHybridAttentionz=Multi-headed attention from 'Attention Is All You Need' paperconfig layer_idxcs t||_||_|dur4td|jjd|j|_|j |_ |j |_ |j |j |_ |j |_ |j |j |_d|_|j|_|j |j |j krtd|j d|j dtj|j |j |j |jd|_tj|j |j |j |jd|_tj|j |j |j |jd|_tj|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.Tz?hidden_size must be divisible by num_heads (got `hidden_size`: z and `num_heads`: z).bias)super__init__rarblogger warning_once __class____name__attention_dropout hidden_sizenum_attention_heads num_headsrBr@rPZ is_causalZattention_multiplierrJ ValueErrorrLinearZattention_biasq_projk_projv_projo_projselfrarbrir0r1rfs2  z"GraniteMoeHybridAttention.__init__past_key_valuepast_key_values4.58new_nameversionNF)r:rIr8ry use_cachecache_positionposition_embeddingsr<cKsF|\} } } ||} ||} ||}| | | |j|jdd} | | | |j|jdd} || | |j|jdd}|dur|nd\}}|durt | | ||\} } |dur|||d}| | ||j |\} }t }|j jdkrt|j j}||| | ||f|jsdn|j|jd|\}}|| | d}||}||fS) Nrr(r$)r7r6reagerrD)rKrJr')sizerqrrrsviewrnrBrRr@r9updaterbr^ra_attn_implementationrrOrkrJrt)rvr:rIr8ryr~rrrXbszZq_len_Z query_statesrYrZr6r7 cache_kwargsZattention_interfacer]r[r0r0r1forwards>         z!GraniteMoeHybridAttention.forward)NNNFNN)rj __module__ __qualname____doc__rintrfrr,Tensorr LongTensorr booltupler __classcell__r0r0rwr1r_s& r_c@seZdZdZdZejdfedddZdej ej e e e e efeej ej fddd Zejd d d Zde e e dddZdS) HybridMambaAttentionDynamicCachea A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the mamba cache (which has a constant shape regardless of seq_len). This cache has two sets of lists of tensors: `key_cache` and `value_cache` for attention cache and `conv_states` and `ssm_states` for mamba cache. Each of these lists has `num_layers` tensors. The expected shape for each tensor For attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, num_heads, seq_len, head_dim)`, while `conv_states` and `ssm_states` have a shape of `(batch_size, 0)` (empty tensors). For mamba layers, `key_cache` and `value_cache` have a shape of `(batch_size, 0)` (empty tensors), while `conv_states` represents the convolution state and has a shape of `(batch_size, d_inner, d_conv)`, and `ssm_states` represents the ssm state and has a shape of `(batch_size, d_inner, d_state)`. FNrac s0|j|_d|_|j}|j}g|_g|_g|_t|jD]}|j|dkr|jt j |j |j d|j |||dg7_|jt j |j|j||dg7_q6|jt jggdg7_|jt jggdg7_|j|q6fddt|jD|_fddt|jD|_dS) NFmambar(devicerMrcs g|]}tjggdqSrr,tensor.0r batch_sizerr0r1 z=HybridMambaAttentionDynamicCache.__init__..cs g|]}tjggdqSrrrrr0r1rr)layers_block_typehas_previous_state mamba_d_conv mamba_d_state conv_states ssm_statestransformer_layersrangenum_hidden_layersr,zeros mamba_expandrlmamba_n_groups mamba_n_heads mamba_d_headrappend key_cache value_cache)rvrarrMrconv_kernel_sizessm_state_sizeir0rr1rfsB   z)HybridMambaAttentionDynamicCache.__init__)rYrZrbrr<cCsz|j|jddkr*||j|<||j|<n)rr chunk_sizer0r0r1reshape_into_chunksNs  rcCs|d}|djg||R}tjtj|||jtjddd}||d}tj|dd}tjtj|||jtjddd}||tj }|S)zo More stable segment sum calculation. Uses cumulative sums and masking instead of direct subtractions. r'.NrZdiagonalrrLr)) rr=r,Ztrilonesrr masked_fillcumsuminf)rrmaskZ tensor_segsumr0r0r1 segment_sumbs   rcCsN|durJ|jddkrJ|jddkrJ|j}||dddddf|}|S)zm Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66 Nrr)r+rMrV)r:rIrMr0r0r1apply_mask_to_padding_statesys$ rcseZdZdZeedfdd Zdeje e e ej e eje ej dddZ de e e ej e ejd d d Zde e e ej e eje ej d d dZZS)GraniteMoeHybridMambaLayeruP Compute ∆, A, B, C, and D the state space parameters and compute the `contextualized_states`. A, D are input independent (see Mamba paper [1] Section 3.5.2 "Interpretation of A" for why A isn't selective) ∆, B, C are input-dependent (this is a key difference between Mamba and the linear time invariant S4, and is why Mamba is called **selective** state spaces) The are a few differences between this and Mamba2Mixer: - The variable use_precomputed_states is slightly different due to the hybrid cache structure - There's a few non-obvious bugs fixed with batching in the slow path that exist in main - Some extra variables that our layer doesn't need have been removed - We ported most of the refactors in https://github.com/huggingface/transformers/pull/35154, which is (as of Dec 18, 2024) unmerged r`cst|j|_|j|_|j|_|j|_t |j |j|_ ||_ |j |_|j|_t|j|_|j|_|j|_|j|_|j|_|j|_dtdf|_d|_d|_ |j d|j|j|_!t"j#|j!|j!|j |j|j!|jdd|_$|j |j!|j}t"j%|j||jd|_&t"'t()|j|_*t(+d|jd}t"'t(,||_-d |j-_.t/|j |jd |_0t"'t()|j|_1d |j1_.t"j%|j |j|jd|_2t3st45d n t45d dS) NrDrgMbP?g?r(r)Z in_channelsZ out_channelsrdZ kernel_sizegroupspaddingrcTepsaThe fast path is not available because on of `(selective_state_update, causal_conv1d_fn, causal_conv1d_update)` is None. Falling back to the naive implementation. To install follow https://github.com/state-spaces/mamba/#installation and https://github.com/Dao-AILab/causal-conv1dzOThe fast path for GraniteMoeHybrid will be used when running the model on a GPU)6rerfrrnrlrrrrrrintermediate_sizerbZmamba_conv_bias use_conv_bias hidden_act activationractZmamba_proj_biasZuse_bias rms_norm_epsZlayer_norm_epsilonrn_groupsrrBZmamba_chunk_sizerfloattime_step_limitZ time_step_minZ time_step_maxconv_dimrZConv1dconv1drpin_proj Parameterr,rdt_biasarangelogA_logZ_no_weight_decayGraniteMoeHybridRMSNormGatednormDout_projis_fast_path_availablergrh)rvrarbZprojection_sizeArwr0r1rfs\    z#GraniteMoeHybridMambaLayer.__init__N)r: cache_paramsrrIseq_idxcCst||}||}|j\}}} |j|j} |duo|jo|dko|j|jjd|j|jjdkop|kno|duo|ddk} | rR| dj |j |j |j gdd\} } }t| |j|j|jj d|jj|j} tj | |j | | gdd\}}}t|j }|ddddfdddddfd|j|jjtjd}|dddddfdd|j}|jddddfd|j}|jddddfd|j}|||j|jd|j}|||j|jd|j}|||j |j}t|j|j||||||d|dd }|||j |j}||| }| |ddddf}nFt|j }|j!d td fkr|ind |j!i}|j"r|durt#||jj d|jj|j|f|j|j$||j|jj|jj%|j j|j j|j|jd d d |}n|j |j |j |j gdd\} } }|dur^| &dd}t'j()||j*|jddf}|j|j+||jdvr|,|| &dddd|f&dd} n2t-| &dd|jj d|jj|j|d&dd} t| |} tj | |j | | gdd\}}}t.|||d|j||||||jd||||jdf|j$|jd|d|jdd|\}}|durt|durt|j|j+||||d}||| }| |}|S)Nrrr'r).rMT)zr dt_softplusrDrZdt_limitF) rrrrZrmsnorm_weightZ rmsnorm_epsZoutproj_weightZ outproj_biasZheaddimZngroupsZnorm_before_gatereturn_final_statesr()siluZswish)r.weightrdrr)rrrrrrr)/rrr+rrrrrbrsqueezesplitrrrnr#rrrdrr,exprrr=rBrVrUrrrrrrrrOr!rvariance_epsilonrRrrSrrcopy_rr"r )rvr:rrrIrprojected_statesrseq_lenrZgroups_time_state_sizeuse_precomputed_statesgatehidden_states_B_CdtBCrrrZhidden_states_reshapedoutZdt_limit_kwargshidden_states_B_C_transposedr scan_output ssm_stater0r0r1cuda_kernels_forwards            <"   "      $         z/GraniteMoeHybridMambaLayer.cuda_kernels_forward)rrrIc3 s|j\}}}|j}t||}|} | jjjjgdd\} } } |duo|jo|dko|j j jd|j j jdko|kno|duo|ddk} | r^|j j j ddd|j j <| dddddf |j j j|j j dddddf<|j j j jjjd}tj|jjddd} jrR| jj} | } nr|dur| dd}tj|j|jddf}|j j || dddd|fdd} t| |} tj| jjjjjgdd\}}}tj ! }| r|j j j}| dddddfddddf} | dd"|| jdj#} j$d "j$jdj#}tjj%| | | j} t&| j'dj'd} |d "jj#jj tj(d }t| d |j |d}|)|jdddddf}|"|jjj|jd*}|)|d|jd}| d |ddddf}|)|dj#}||d j |d}|j j |j j |||)|jdddddf}|"|jjj|jd*}|)|d|jd}|j j j |j|jd }|+|jj#j}|+|jjd}t,||}|+|jj#}j-d "j-jdj#}||| |j}|)|dddddf}ntj%| j$} t&| j'dj'd} |)||dj#!}|)||dj!}|)||dj!}|j.jjdjd }|j.jjdjd }j/|j/j/j-d t0|}|| d }| |j| }fdd||||fD\}}}}|1dddd}tj2|dd}tt3|} |dddddddddddf|dddddddddddf}!|!jdd}"|"d | 1dddddd }#|#jdd}$|$d |dddddfjdd}%t|ddddddddf|}&||&1ddddd }'|'ddddf|d jdd}(| r0|j j ddddfj |(jd})nt4|(ddddf})tj5|)|(gdd}(tt3tj|dddddddfd}*|*dd}*|*d |(ddddddfjdd}+|+ddddf|+dddf}(},t|}-|ddddf|(ddddddf}.|-1dddd}/|.d|/d }0|%|0}|)|djj#}||}dkr|ddd|ddddf}|)||d}|,dur|dur|j j |,6|| }17|1 |}2|2S)Nr'r)rr)Zshiftsdimsrr(.r).NNrr)r* output_sizecsg|]}t|jqSr0)rr)rtrrvr0r1r rz.r rrL)rr)8r+rMrrrrrrnrrrbrZrollrVrrrr,sumrrrdrrRrrSrrrrrrrrr=rBrZsoftplusclamprrUr>rWrZbmmrZrepeat_interleaverrZpermuterrZ zeros_liker-rr)3rvZ input_statesrrrIrrrrMrrrrrrrr:rrrZ cache_devicerZdAZdBZdBxrZssm_states_reshapedZ C_reshapedyrZ D_residualZA_cumsumLZG_intermediateGZM_intermediateMZY_diagZ decay_statesZB_decayZstatesZprevious_statesZ decay_chunkZ new_statesrZstate_decay_outZC_times_statesZstate_decay_out_permutedZY_offrZcontextualized_statesr0rr1 torch_forwards      @    ,  $"$ $  $P  &*"&0 (& *   z(GraniteMoeHybridMambaLayer.torch_forward)rrrIrcKstr&d|jjjjvr&||||||S|dur6td|j}|dur|jddkr|jddkr||dddddf |}| ||||S)Ncudaz\`seq_idx` support requires fast path support. Please install `mamba_ssm` and `causal_conv1d`rr) rrrrtyperNotImplementedErrorrMr+rVr)rvr:rrrIrrXrMr0r0r1rOs $ z"GraniteMoeHybridMambaLayer.forward)NNNN)NNN)NNNN)rjrrrrrrfr,rrrr IntTensorrrrrr0r0rwr1rsB F . Srcs(eZdZdfdd ZdddZZS) rư>cs&ttt||_||_dS)Nrerfrrr,rrrrvrlrrwr0r1rfgs z%GraniteMoeHybridRMSNormGated.__init__NcCsj|j}|tj}|dur2|tj|tj}|djddd}|t ||j }|j ||SNr(r'T)Zkeepdim) rMrVr,rUrrSrpowmeanrsqrtrr)rvr:r input_dtypevariancer0r0r1rls z$GraniteMoeHybridRMSNormGated.forward)r)N)rjrrrfrrr0r0rwr1rfsrcs:eZdZdZedfdd ZejejdddZZ S)GraniteMoeHybridMLPz~ MLP layer for shared experts Args: config: Configuration object with model hyperparameters. rcsZt|j|_|j|_t|j|_tj |j|jddd|_ tj |j|jdd|_ dS)Nr(Frc) rerfrl input_sizeZshared_intermediate_sizerrrrrp input_linear output_linearrvrarwr0r1rfs   zGraniteMoeHybridMLP.__init__)r:r<cCs<||}|jddd}||d|d}||}|S)Nr(r'r)rr)rchunkrr)rvr:chunked_hidden_statesr0r0r1rs   zGraniteMoeHybridMLP.forward) rjrrrrrfr,rrrr0r0rwr1rxs rc@s@eZdZUdZejed<ejed<eed<eed<ejed<dS)GraniteFlashAttentionKwargsa Keyword arguments for advanced Flash Attention, causal-conv1d, and mamba_ssm kernel usage. Use cases include padding-free training and fewer `torch.compile` graph breaks. Attributes: cu_seq_lens_q (`torch.LongTensor`) Gets cumulative sequence length for query state. cu_seq_lens_k (`torch.LongTensor`) Gets cumulative sequence length for key state. max_length_q (`int`): Maximum sequence length for query state. max_length_k (`int`): Maximum sequence length for key state. seq_idx (`torch.IntTensor): Index of each packed sequence. Z cu_seq_lens_qZ cu_seq_lens_kZ max_length_qZ max_length_krN) rjrrrr,r__annotations__rrr0r0r0r1r#s   r#F)totalcs.eZdZdfdd ZddZddZZS) GraniteMoeHybridRMSNormrcs&ttt||_||_dS)zF GraniteMoeHybridRMSNorm is equivalent to T5LayerNorm Nrrrwr0r1rfs z GraniteMoeHybridRMSNorm.__init__cCsJ|j}|tj}|djddd}|t||j}|j||Sr) rMrVr,rUrrrrr)rvr:rrr0r0r1rs  zGraniteMoeHybridRMSNorm.forwardcCst|jjd|jS)Nz, eps=)rrr+r)rvr0r0r1 extra_reprsz"GraniteMoeHybridRMSNorm.extra_repr)r)rjrrrfrr'rr0r0rwr1r&sr&cs0eZdZeeeddfdd ZddZZS)GraniteMoeHybridParallelExpertsN) num_expertsrrr<cs6ttt||||_||_||_||_ dS)a Initialize the GraniteMoeHybridParallelExperts module. The experts weights are stored in [num_experts, output_size, input_size] format. Such that it's compatible with many MoE libraries, such as [Megablock](https://github.com/databricks/megablocks) and [ScatterMoE](https://github.com/shawntan/scattermoe), as well as the [MoE kernel](https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/fused_moe/fused_moe.py) used in vllm. Args: num_experts (int): Number of experts. input_size (int): Size of the input. output_size (int): Size of the output. N) rerfrrr,emptyrr)rr)rvr)rrrwr0r1rfs  z(GraniteMoeHybridParallelExperts.__init__cCsP|j|dd}g}t|jD] }|t|||j|qtj|dd}|S)a Forward pass of the GraniteMoeHybridParallelExperts module. Args: inputs (Tensor): Input tensor. expert_size: Expert size information. Returns: Tensor: Output tensor. rr)) rrr)rFZlinearrr,r-)rvinputs expert_sizeZ input_listZ output_listrresultsr0r0r1rs z'GraniteMoeHybridParallelExperts.forwardrjrrrrfrrr0r0rwr1r(sr(cs.eZdZeeedfdd ZddZZS)GraniteMoeHybridTopKGatingrr)top_kcs2t||_||_||_tj||dd|_dS)a Initialize the top-k gating mechanism. Args: input_size (`int`): Size of the input. num_experts (`int`): Number of experts. top_k (`int`): Number of top experts to select. FrcN)rerfr)rr2rrplayer)rvrr)r2rwr0r1rfs z#GraniteMoeHybridTopKGating.__init__cCs||}|j|jdd\}}tj|dd|}tj|d|j g|j |j d}| d|d}| d}|}|} | d\} } | j|jdd} |}|| } | | | ||fS)Nrr)rrMrtrunc)Z rounding_mode)r3rtopkr2r,rTZtype_asrrr)rMrZscatterlongrtolistflattensortdiv)rvr:logitsZ top_k_logitsZ top_k_indicesZ top_k_gatesrZgatesr-Z top_k_expertsrZindex_sorted_experts batch_index batch_gatesr0r0r1rsz"GraniteMoeHybridTopKGating.forwardr/r0r0rwr1r0sr0cs.eZdZdZedfdd ZddZZS)GraniteMoeHybridMoEz A Sparsely gated mixture of experts layer with 1-layer Feed-Forward networks as experts. Args: config: Configuration object with model hyperparameters. rcslt|j|_|j|_t|j|_t|j |j|jd|_ t|j |j|j|_ t |j|j |j d|_dS)Nr(r1)rerfrlrrrrrr(num_local_expertsrrr0num_experts_per_tokrouterr rwr0r1rf%s   zGraniteMoeHybridMoE.__init__cCs|\}}}|d|}||\}}}}} ||} || |} | jddd} || d| d} || |} | |dddf} tj|||j f| j | j d}| d|| }| |||j }|| fS)a Forward pass of the mixture of experts layer. Args: layer_input (Tensor): Input tensor. Returns: Tensor: Output tensor. Tensor: Router logits. r'r(r)rrNr4)rr>rBrr!rrr,rrrMrZ index_addr)rvZ layer_inputrlengthZemb_sizerr=r>r- router_logitsZ expert_inputsr:r"Zexpert_outputsrZ layer_outputr0r0r1r8s   zGraniteMoeHybridMoE.forward)rjrrrrrfrrr0r0rwr1r?sr?cseZdZeedfdd Zeddddd eje eje e e e e e e ej e e e e ejejfeee eje e ejejffd d d ZZS)GraniteMoeHybridDecoderLayerr`cst|j|_d|_|jdkr,t||_t|j|jd|_ t|j|jd|_ |j |_ t ||_ d|_|j|dkrt|||_n t|||_|j||_t|dddk|_dS)Nrrrr@)rerfrl self_attnr@r?block_sparse_moer&rinput_layernormpost_attention_layernormresidual_multiplierr shared_mlprrrr_ layer_typegetattr has_expertsrurwr0r1rfYs      z%GraniteMoeHybridDecoderLayer.__init__rxryrzr{NF) r:rIryoutput_attentionsr~routput_router_logitsrrXr<c  Ks|} ||}|jdur8|jf||||d| }d} n$|jf|||||||d| \}} | ||j}|} ||}|jr||\} } | ||}n||}d} | ||j}|f}|r|| f7}|r|| f7}|S)aC 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. past_key_values (`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. 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 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. 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.Can be used to provide `GraniteFlashAttentionKwargs` for padding-free training and/or improve torch.compile performance. N)r:rrrI)r:rIryrOr~rr)rHrrFrJrIrNrGrK)rvr:rIryrOr~rrPrrXZresidualZself_attn_weightsZmoe_hidden_statesrDoutputsr0r0r1rpsL&       z$GraniteMoeHybridDecoderLayer.forward)NNFFNFN)rjrrrrrfrr,rrr rrrrr# FloatTensorrrr0r0rwr1rEXs* rEcsJeZdZUeed<dZdZdgZdgZdZ dZ dZ dZ fddZ ZS) GraniteMoeHybridPreTrainedModelramodelTrEryFcst|t|tr,|jjjd|jjdt|t rp|j j dt t d|jd|j_|jj dnt|tr|jj ddS)NrD)rZstdg?r)re _init_weights isinstancer(rdataZnormal_raZinitializer_rangerrZfill_r,rrrnrrr)rvrErwr0r1rUs    z-GraniteMoeHybridPreTrainedModel._init_weights)rjrrrr$Zbase_model_prefixZsupports_gradient_checkpointingZ_no_split_modulesZ_skip_keys_device_placementZ_supports_flash_attnZ_supports_sdpaZ_can_compile_fullgraphZ _is_statefulrUrr0r0rwr1rSs rScsDeZdZUejed<dedfdd Zee ddZ Z S) GraniteMoeHybridRotaryEmbeddinginv_freqNrcstt|dr:t|jtr:|jd|jd|_nd|_|j|_ |j|_ ||_ t |j|_ | |j |\}|_|jd|dd|j|_dS)N rope_scaling rope_typerdefaultrYF) persistent)rerfhasattrrVrZrgetr[max_position_embeddingsZmax_seq_len_cachedZoriginal_max_seq_lenrarZ rope_init_fnattention_scalingZregister_bufferrYZoriginal_inv_freq)rvrarrYrwr0r1rfs  z(GraniteMoeHybridRotaryEmbedding.__init__c Cs|jddddf|jddd|j}|dddddf}t|jjtrl|jjdkrl|jjnd}t j |ddV|| dd}t j ||fdd }| |j}||j} Wdn1s0Y|j|jd | j|jd fS) Nrr'rZmpscpuF) device_typeZenabledr(r)r)rYrr=r+rVrrVrrr,ZautocastrRr-r6rar7rM) rvr.r8Zinv_freq_expandedZposition_ids_expandedrcZfreqsZembr6r7r0r0r1rs0&,z'GraniteMoeHybridRotaryEmbedding.forward)N) rjrrr,rr$rrfZno_gradrrrr0r0rwr1rXs  rXcseZdZedfdd Zeedeje ej e eje e e e ejfe eje ee ee ee ee ee ejeee eefd ddZde ej d fej ej e ed d d Zeej eeejej ed ddZddZZS)GraniteMoeHybridModelrcstj|_j|_tjj|j|_t fddt j D|_ t jjd|_d|_j|_j|_j|_|j|j|_j|_j|_j|_|jdkrtnd|_|dS)Ncsg|]}t|qSr0)rE)rrbrr0r1rrz2GraniteMoeHybridModel.__init__..rFZrope)rerfZ pad_token_idZ padding_idx vocab_sizerZ Embeddingrl embed_tokensZ ModuleListrrlayersr&rrgradient_checkpointingembedding_multiplierrmrnrBr`Z rope_thetaZposition_embedding_typerX rotary_emb post_initr rwrr1rfs$ zGraniteMoeHybridModel.__init__N) input_idsrIr8ry inputs_embedsr~rOoutput_hidden_statesrP return_dictrrXr<c  KsJ|dur |n|jj}|dur |n|jj}|dur4|n|jj}| durH| n|jj} |du|duArhtd|jr|jr|rt dd}|dur| |}||j }|r|durt d| dur|dur| nd} t j| | |jd|jd} |dur| d}|||| ||}||| }|}d}|jdurB|||}|rLdnd}|rZdnd}| rhdnd}|jD]}|jd kr|n|}|r||f7}||f||||| | |d | }|d}|r|ddur||df7}| rr|d durr||d f7}qr||}|r$||f7}|r8|js8d |_t|||||d S)Nz:You must specify exactly one of input_ids or inputs_embedszX`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`.FzGraniteMoeHybrid requires an initialized `HybridMambaAttentionDynamicCache` to return a cache. Because one was not provided, no cache will be returned.rrrr0r)rIryrOr~rrPrr'T)Zlast_hidden_stateryr: attentionsrD)rarOrnr~use_return_dictrorhrOrgrhrfrirr,rr+rr3_update_causal_mask_update_mamba_maskrjrgrLrrr)rvrlrIr8ryrmr~rOrnrProrrXpast_seen_tokensr\ mamba_maskr:rZall_hidden_statesZall_self_attnsZall_router_logitsZ decoder_layerZ layer_maskZ layer_outputsr0r0r1r!s               zGraniteMoeHybridModel.forwardFr%)rIrrryrOc CsB|jjdkr(|dur$|dkr$|SdS|jjdkrLt|tjrHt|}|S|dur\|nd}|durn|jnd}|jjdkr|s|st j ||||j drdS|j }|j d} |r|} n"t|tjr|j d n || d} |j|| | |||j dd } |jjdkr>|dur>|jjd vr>|s>t|j} t | | } | S) NZflash_attention_2rDZflex_attentionrFZsdpa)rmZpast_key_values_lengthZ is_trainingrr')sequence_length target_lengthrMrr)rZxpuZnpu)raranyrVr,rr&rrr Z_ignore_causal_mask_sdparOrMr+Zget_max_cache_shape5_prepare_4d_causal_attention_mask_with_cache_positionrrfinfominZ_unmask_unattended) rvrIrrryrOrtZusing_compilable_cacherMrvrwr\ min_dtyper0r0r1rrsZ           z)GraniteMoeHybridModel._update_causal_mask)rIrvrwrMrrc KsF|dur|dkr|}n&t|j}tj||f|||jd}|dkrVtj|dd}|tj||jd|ddk9}|ddddddf |ddd}|durB| }|j d} |ddddddd| f|ddddddf |j} | dk} |ddddddd| f | ||ddddddd| f<|S) aM 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. cache_position (`torch.Tensor`): Indices depicting the position of the input sequence tokens in the sequence. batch_size (`torch.Tensor`): Batch size. Nr)Z fill_valuerMrrrrr'r)r*r,rzr{fullrZtriurr>r=cloner+rVr) rIrvrwrMrrrXr\r|Z mask_lengthZ padding_maskr0r0r1rys*  $  6  zKGraniteMoeHybridModel._prepare_4d_causal_attention_mask_with_cache_positioncCs.|}|ddks&|dur*t|dkr*d}|S)zv No need for zeroing states when 1. Cached forward 2. Attending to all inputs rNr)r,all)rvrIrrur0r0r1rss"z(GraniteMoeHybridModel._update_mamba_mask) NNNNNNNNNNN)F)rjrrrrfrrr,rrrrr listrRrrr#rrrrr staticmethodrrMryrsrr0r0rwr1rds\ {  D6rdr() gate_logitsr)rIr<cs|dust|tsdSt|trF|djtjfdd|Ddd}tjjj|dd}tj||dd\}}tjj ||}|durtj | dd} tj |dd} n|j \} } |j d| | } |dddddddf | | | ||fd||}tj| |ddtj|dd} |ddddddf | | | |j dfd|j d}tj||ddtj|dd} |jjdur|jjnd}|j dt|}t| dd|||j df| d}||S)a Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch. See Switch Transformer (https://huggingface.co/papers/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. Nrcsg|]}|qSr0)rV)rZ layer_gateZcompute_devicer0r1rArz,load_balancing_loss_func..r)r'r)rVrrr,r-rrSrTr6Zone_hotrrr+r=r>rVrindexrr3)rr)r2rIZconcatenated_gate_logitsZrouting_weightsrZselected_expertsZ expert_maskZtokens_per_expertZrouter_prob_per_expertrrvrZexpert_attention_maskZ router_per_expert_attention_maskZ device_indexZrankZ overall_lossr0rr1load_balancing_loss_funcsR     &rcseZdZdgZedfdd Zed eej eej eej ee e e ejfeejeej eeeeeeeeeeeej e eej fe eefddd Zdd d ZZS)GraniteMoeHybridForCausalLMzlm_head.weightrcsXt|t||_|j|_tj|j|jdd|_|j |_ |j |_ |j |_ | dS)NFrc)rerfrdrTrerrprllm_headrouter_aux_loss_coefr@r)rArkr rwr0r1rfxs  z$GraniteMoeHybridForCausalLM.__init__Nr)rlrIr8ryrmlabelsr~rOrnrProrlogits_to_keepr<cKs|dur |n|jj}| dur | n|jj} | dur4| n|jj} | durH| n|jj} |jf|||||||| | | | d |}|d}t| trt| dn| }| |dd|ddf}||jj }d}|dur| }|j ||fd|jj i|}d}| r>t| r |jn|d|j|j|}|dur>||j||j7}| s~|f|dd}| rf|f|}|durz|f|S|St||||j|j|j|jdS)ax 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]`. Example: ```python >>> from transformers import AutoTokenizer, GraniteMoeHybridForCausalLM >>> model = GraniteMoeHybridForCausalLM.from_pretrained("ibm/PowerMoE-3b") >>> tokenizer = AutoTokenizer.from_pretrained("ibm/PowerMoE-3b") >>> 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) rlrIr8ryrmr~rOrnrProrrrer'r)lossaux_lossr<ryr:rprD)rarOrPrnrqrTrVrslicerZlogits_scalingrZ loss_functionrerrDr)rArrVrrryr:rp)rvrlrIr8ryrmrr~rOrnrProrrrXrQr:Z slice_indicesr<rroutputr0r0r1rsx(    z#GraniteMoeHybridForCausalLM.forwardTc Ks|du} | sj|dus&|d|jdkrD|dd|jd df}q|jd|jdkr|dd|f}n |rt|j|jd|j|jd}|dur|dur|dd}||dkd| s|dd|jd df}|dur| rd|i} n d|i} | |||||d| S)Nr'rrrrmrl)r8ryr~rIr) r+rrarMrr7rZ masked_fill_rWr) rvrlryrIrmrr8r~rXZ empty_past_kvZ model_inputsr0r0r1prepare_inputs_for_generations<     z9GraniteMoeHybridForCausalLM.prepare_inputs_for_generation) NNNNNNNNNNNNr)NNNNNT)rjrrZ_tied_weights_keysrrfrrr,rrrr rrRrrrrrrrr0r0rwr1rusN   pr)rrdrS)Nr)rD)Nr(N)\typingrrrrrr,Ztorch.nn.functionalrrSr+Ztransformers.activationsrZ cache_utilsr Z generationr Zmodeling_attn_mask_utilsr Zmodeling_layersr Zmodeling_outputsrrrZmodeling_rope_utilsrrZmodeling_utilsrrZprocessing_utilsrutilsrrrrZutils.deprecationrZutils.import_utilsrrZconfiguration_granitemoehybridrZ+mamba_ssm.ops.triton.selective_state_updaterZ!mamba_ssm.ops.triton.ssd_combinedr r!Z causal_conv1dr"r#Z!torch.nn.attention.flex_attentionr%Zintegrations.flex_attentionr&Z get_loggerrjrgr2r9rrrCModulerr^r_rrrrrrrrrrr#r&r(r0r?rErSrXrdrrr__all__r0r0r0r1s               W`  d-0<q$ V9