Vh$dZddlZddlZddlZddlZddlZddlZddlmZddl m Z m Z m Z m Z ddlZddlmZddlmZddlmZddlmZdd lmZmZdd lmZmZdd lmZgd Ze eeeeegefZ e eeeegefZ!Gd deZ"de de"fdZ# d[de de$e e%dfde$e e%dfde e%e&e e%fde'f dZ(dedededededef dZ)dededededef dZ*d Z+d!Z,d"ed#efd$Z-de$eeffd%Z.d"ed#efd&Z/d'ed(ed)e%d*e%fd+Z0Gd,d-Z1d.Z2d/Z3e+e+dfd0ed1e'de$ee effd2Z4d3e!de!fd4Z5d3e!de!fd5Z6e+e+fdefd6Z7e+e+fd7e$ee efd8e e d9e$e%e%fd:e%d;e%de1f d<Z8 d\d>e e e!fd?e e%d@e e%dAe%dBe%dCe9defdDZ:d=e+dfd8e!d?e e%d@e e%dAe%dBe%dCe9dEe e%e$e%e%ffde1fdFZ;dGedHede1fdIZdGedHedPedQe'fdRZ?dGedHedPefdSZ@dGedHedPefdTZAdGedHedPefdUZB d]dGedHedPedVe e d7e e1dWe eCdXe'dQe'dYe eDe9e fde ee$eefffdZZy)^zIThis module implements the user facing API for flex_attention in PyTorch.N)Enum)AnyCallableOptionalUnion)Tensor)TransformGetItemToIndex)flex_attention)_set_compilation_env))_temp_remove_metadata_torch_function_mode-_temp_remove_pre_dispatch_torch_function_mode)_supported_head_dim_validate_sdpa_input) tree_map_only) BlockMaskr create_block_mask create_maskcreate_nested_block_maskor_masks and_masks noop_maskceZdZdZdZdZdZy)_ModificationTypezEnum for the type of modification function. - SCORE_MOD: score_mod function which accepts a score as the first argument - mask_mod: mask function which does not accept a score and is only used for generating block mask N)__name__ __module__ __qualname____doc__ SCORE_MODMASK_MODUNKNOWNQ/home/dcms/DCMS/lib/python3.12/site-packages/torch/nn/attention/flex_attention.pyrr)s IHGr%rfnreturnctdtj|jj D}|dk(s|dk(sJ|dk(rt j S|dk(rt jSt jS)a\Get the type of modification function. This function inspects the number of positional arguments of the function to determine the type of modification function. If the function has 5 positional arguments, it is considered as a score_mod function. If the function has 4 positional arguments, it is considered as a mask function. c3nK|]-}|jtjjk(rd/yw)rN)defaultinspect Parameterempty).0params r& z _get_mod_type..<s0  ==G--33 3 s35) sumr, signature parametersvaluesrr!r"r#)r'num_positional_argss r& _get_mod_typer95s&&r*55<<> ! #':a'?? ?a ***  ! ))) (((r%Fprefix.suffixout_dims group_dimcxg}gd}|r|dgz }|dgz }|D] }tj|||z|z|}"|S)aUsed to vmap both score_mods and mask_mods over 4-dimensional/5-dimension inputs. Mapping over the [b, hq, q_idx, kv_idx] or [b, hkv, g, q_idx, kv_idx] dimensions. Args: fn (callable): The function to vmap. prefix (tuple): The prefix of the vmap. For score mod functions, this should be set to (0,). For mask_mods = () suffix (tuple): We need to add (0,) if gradOut is being mapped over, and (None,) * len(other_buffers). out_dims (tuple): For forward cases, keep this as the default 0 since we are only returning 1 output. For backwards, the joint graph returns grads for B, H, Q_idx, KV_idx and other_buffers, so we set this to (0, None, None, None, None) + (None,) * len(other_buffers). Returns: callable: The vmapped function. ))NNNr)NNrNNrNNr?)rNNN)in_dimsr<)torchvmap)r'r:r;r<r= dimensionsdimss r&_vmap_for_bhqkvrEKsp2OQJJ  !  JO ZZFTMF$:X NO Ir%scorebatchheadtoken_qtoken_kvc|SNr$)rFrGrHrIrJs r& _identityrMys  Lr%cZ|jdtj|jS)zReturns a noop mask_modr$)sizedtypedevice)new_onesrAboolrQ)rGrHrIrJs r&rrs! >>rELL> IIr%i@num_blocks_in_row col_indicesc |jd |jd|jdd}|j fd}|}tt|D]}t j |d}|||}|S)Nc|j dztj}tj tjj d}tjtj}||j dk}tj ||}|jd|||f<|dddfjS)NrrPrPrQrYr$) new_zerosrAint32arangeint unsqueezewhererR contiguous) kv_num_blocks kv_indices dense_mask row_indices col_range index_mask valid_indicesrQnum_colsnum_rowss r&create_dense_onez+_ordered_to_dense..create_dense_ones))(HqL )T ll8599VLVV  LL6J !8!8!<<  J HE 2<1D1DR1H ; -.!YhY,'2244r%)rr)r@)shaperQrangelenrArB) rUrV batch_dimsrmcreate_dense_batched_outrQrkrls @@@r&_ordered_to_denserus  $H  $H"(("-J  % %F5 , 3z? #P$zz*>OP 0+ >C Jr%cT|jtj}|jd}tj|ddd}|jtjtj |jtjtj fS)Nr[rYdimT)rx descendingstable) memory_format)torAr^r4argsortcontiguous_format)rfrUrVs r&_dense_to_orderedrs|U[[1J"2.-- tDQKU[[8O8OPu{{%2I2IJ r%cPt||}t|jddS)NrXrY)rur transpose)rUrVdenses r&_transpose_orderedrs% / =E U__R4 55r% num_blocksindices new_num_rows new_num_colsc|ddddd|d|f}|ddddd|f}tj||k||}tj||dddddddfkdjtj}||fS)NrYrw)rArbr4r|r^)rrrrs r&_adjust_num_blocks_and_indicesrs aM\M=L=89GAq-<-/0JZ,6 LQJ7Z1a %>>BGJJ5;;WJ w r%c*eZdZUdZeeefed<eed<eed<eeed<eeed<eeed<eeed<eeed <eeed <eeefed <e ed <deeefdededeedeedeedeed eed eed eeefd e fd Z e dde ddfdededeedeed e eeeeffd ee deeeeffdZd defdZedZdZd!dZdZdedefdZdZdefdZdefdZd"dZde ej:efddfdZy)#ra< BlockMask is our format for representing a block-sparse attention mask. It is somewhat of a cross in-between BCSR and a non-sparse format. Basics ------ A block-sparse mask means that instead of representing the sparsity of individual elements in the mask, a KV_BLOCK_SIZE x Q_BLOCK_SIZE block is considered sparse only if every element within that block is sparse. This aligns well with hardware, which generally expects to perform contiguous loads and computation. This format is primarily optimized for 1. simplicity, and 2. kernel efficiency. Notably, it is *not* optimized for size, as this mask is always reduced by a factor of KV_BLOCK_SIZE * Q_BLOCK_SIZE. If the size is a concern, the tensors can be reduced in size by increasing the block size. The essentials of our format are: num_blocks_in_row: Tensor[ROWS]: Describes the number of blocks present in each row. col_indices: Tensor[ROWS, MAX_BLOCKS_IN_COL]: `col_indices[i]` is the sequence of block positions for row i. The values of this row after `col_indices[i][num_blocks_in_row[i]]` are undefined. For example, to reconstruct the original tensor from this format: .. code-block:: python dense_mask = torch.zeros(ROWS, COLS) for row in range(ROWS): for block_idx in range(num_blocks_in_row[row]): dense_mask[row, col_indices[row, block_idx]] = 1 Notably, this format makes it easier to implement a reduction along the *rows* of the mask. Details ------- The basics of our format require only kv_num_blocks and kv_indices. But, we have up to 8 tensors on this object. This represents 4 pairs: 1. (kv_num_blocks, kv_indices): Used for the forwards pass of attention, as we reduce along the KV dimension. 2. [OPTIONAL] (full_kv_num_blocks, full_kv_indices): This is optional and purely an optimization. As it turns out, applying masking to every block is quite expensive! If we specifically know which blocks are "full" and don't require masking at all, then we can skip applying mask_mod to these blocks. This requires the user to split out a separate mask_mod from the score_mod. For causal masks, this is about a 15% speedup. 3. [GENERATED] (q_num_blocks, q_indices): Required for the backwards pass, as computing dKV requires iterating along the mask along the Q dimension. These are autogenerated from 1. 4. [GENERATED] (full_q_num_blocks, full_q_indices): Same as above, but for the backwards pass. These are autogenerated from 2. seq_lengthsrdrefull_kv_num_blocksfull_kv_indices q_num_blocks q_indicesfull_q_num_blocksfull_q_indices BLOCK_SIZEmask_modc b|jdkr td|Jd|Jd|Jd|Jd|du|duk(sJd|du| duk(sJd||_||_||_||_||_||_||_||_ | |_ | |_ | |_ y) Nr)BlockMask must have at least 2 dimensionszkv_num_blocks must be providedzkv_indices must be providedzq_num_blocks must be providedzq_indices must be providedGfull_kv_num_blocks and full_kv_indices must be both provided or omittedzEfull_q_num_blocks and full_q_indices must be both provided or omitted) rx RuntimeErrorrrdrerrrrrrrr) selfrrdrerrrrrrrrs r&__init__zBlockMask.__init__s >> a JK K(J*JJ(%D'DD%'H)HH'$B&BB$"d* t #  U T U "T) d "  S R S '*$"4.("!2,$  r%Nc t|jdkr td|du|duk(sJdt||\}} ||Jt||\} } nd\} } t|tr||f}||nt }|.|j d|dz} | j d|dz} | | f}|||||||| | | || S) a Creates a BlockMask instance from key-value block information. Args: kv_num_blocks (Tensor): Number of kv_blocks in each Q_BLOCK_SIZE row tile. kv_indices (Tensor): Indices of key-value blocks in each Q_BLOCK_SIZE row tile. full_kv_num_blocks (Optional[Tensor]): Number of full kv_blocks in each Q_BLOCK_SIZE row tile. full_kv_indices (Optional[Tensor]): Indices of full key-value blocks in each Q_BLOCK_SIZE row tile. BLOCK_SIZE (Union[int, tuple[int, int]]): Size of KV_BLOCK_SIZE x Q_BLOCK_SIZE tiles. mask_mod (Optional[Callable]): Function to modify the mask. Returns: BlockMask: Instance with full Q information generated via _transposed_ordered Raises: RuntimeError: If kv_indices has < 2 dimensions. AssertionError: If only one of full_kv_* args is provided. rrNrNNrXrr) rrdrerrrrrrrr)rxrr isinstancer`rrn)clsrdrerrrrrrrrrq_length kv_lengths r&from_kv_blockszBlockMask.from_kv_blocks;s: >> a JK K"d* t #  U T U #5]J"O i  )". ..0B"O1 - ~1; - ~ j# &$j1J'38  !''+jm;H!+jm;I#Y/K#'!1+%/)!  r%flattenc|r=|jd|jdf}|jd|jdf}n|jf}|jf}g||j|j|j|j |j |j|j|j||jS)z Returns a tuple of the attributes of the BlockMask. Args: flatten (bool): If True, it will flatten the tuple of (KV_BLOCK_SIZE, Q_BLOCK_SIZE) rr) rrrdrerrrrrrr)rr block_sizers r&as_tuplezBlockMask.as_tuples //!,dooa.@AJ++A.0@0@0CDK//+J++-K         OO    # #         NN    " "          MM  r%cf|jj^}}}t||jzSrL)rerntupler)rrqrss r&rnzBlockMask.shapes. OO11QZ 4#3#333r%cd|jd|jdd}|jj}||z }|dz }|S)NzBlockMask(shape=z , sparsity=.2fz%,  ))rnsparsity to_stringstrip)rsmask_strs r&__str__zBlockMask.__str__sOtzzl+dmmoc5J% P>>#))+ X  U r%r(c |j|}|j|}|j-|jJ|j|}|j|}nd}d}tj |||||j d|jS)a Returns a new BlockMask instance by getting the mask for the given index position. Args: index: Index to apply to all attributes. Example Usage: .. code-block:: python def causal_mask(b, h, q_idx, kv_idx): return q_idx >= kv_idx block_mask = create_block_mask(causal_mask, 4, 2, 512, 512, device="cuda") assert block_mask.kv_num_blocks.shape == (4,2,4) assert block_mask.kv_indices.shape == (4,2,4,4) # Index on batch dimension new_block_mask = block_mask[0] assert new_block_mask.kv_num_blocks.shape == (2,4) assert new_block_mask.kv_indices.shape == (2,4,4) # Index on batch and head dimension new_block_mask = block_mask[0, 1] assert new_block_mask.kv_num_blocks.shape == (4,) assert new_block_mask.kv_indices.shape == (4,4) # slicing on batch and head dimension new_block_mask = block_mask[0:2, 1:2] assert new_block_mask.kv_num_blocks.shape == (2,1,4) assert new_block_mask.kv_indices.shape == (2,1,4,4) # slicing on batch, head, and query dimension new_block_mask = block_mask[0:2, 1:2, torch.tensor([1], dtype=torch.int32)] assert new_block_mask.kv_num_blocks.shape == (2,1,1) assert new_block_mask.kv_indices.shape == (2,1,1,4) Nrrr)rdrerrrrrr)rindexnew_kv_num_blocksnew_kv_indicesnew_full_kv_num_blocksnew_full_kv_indicess r& __getitem__zBlockMask.__getitem__sJ!..u5/  " " .''3 33%)%<%.shape_or_nonesm177 5 5r%zBlockMask( kv_num_blocks=z, kv_indices=z, full_kv_num_blocks=z, full_kv_indices=z, q_num_blocks=z, q_indices=z, full_q_num_blocks=z, full_q_indices=z, BLOCK_SIZE=z , shape=z, sparsity=rz%, mask_mod=rr)rrArrdrnrerrrrrrrrhasattrrr)rrs r&__repr__zBlockMask.__repr__sL 6Xell3 6!!%!3!3!9!9 :;"oo3345&&3D4K4K&M%NO##01E1E#F"GH -d.?.? @AB*4>>:;<%%243I3I%J$KL""/0C0C"D!EF"oo./ % MMOC016=dmmZ6XDMM22lm  _c^k^klm  r% new_q_len new_kv_lenc||jdzdz |jdz}||jdzdz |jdz}t|j|j||\}}|j4|j Jt|j|j ||\}}nd}d}|j |||||j|jS)Nrr)rrrdrerrrr) rrrrrrrrrs r&_adjustzBlockMask._adjusts!DOOA$66:tq?QQ "T__Q%77!;PQ@RR ,J   |- )>  " " .''3 33/''$$  &#&* ""& ""   "  OO MM   r%c0|j}d}||S)zIReturns the number of elements (not accounting for sparsity) in the mask.cLtjtj|dSNr) functoolsreduceoperatormul)xss r&_prodzBlockMask.numel.._prods##HLL"a8 8r%r)rrnrs r&numelzBlockMask.numels  9U|r%c"|j}|jj}|j||jjz }|j |j dz|j dz}||z }dd|z zS)zEComputes the percentage of blocks that are sparse (i.e. not computed)rrd)rrdr4ritemr)r total_sizecomputed_blocks computed_size dense_ratios r&rzBlockMask.sparsitysZZ\ ,,002  " " . t66::< .create_block_vis..summarize_sectionGs6$]]_11388: ? 1_ r%c||dz z|zSrr$)abs r&cdivz;BlockMask.to_string..create_block_vis..cdivPsQU ))r%rrr)appendjoinreversedmaxro) batch_idx descriptorsvisrrrow_stepcol_steprccur_maskidxcharrfmax_colsmax_rowsrkrls r&create_block_visz-BlockMask.to_string..create_block_vis@sK   ) .))H[12T9C ! *1d8X67H1d8X67H1h1 q(H5$A)H(1#+C=1, Q\!11q8|3C!CDD4!8OC$t  Jr%z...z3To print out more, set BlockMask.to_string(limit=N)zNYou can also index (BlockMask[batch, head]) to choose a specific batch or headr) rrnrr` enumerate itertoolsproductrorr)r grid_sizelimitrqr total_visirr block_visrfrrrkrls @@@@@r&rzBlockMask.to_string0s ]]_ *4*:*:'Xx i % H H "_HH!* Hh  D '   *=Qa= > (NCe|  '  !VW  d()4I   Y ' (yy## >s-C0 rQcpttjfd|jd}t |S)aMoves the BlockMask to the specified device. Args: device (torch.device or str): The target device to move the BlockMask to. Can be a torch.device object or a string (e.g., 'cpu', 'cuda:0'). Returns: BlockMask: A new BlockMask instance with all tensor components moved to the specified device. Note: This method does not modify the original BlockMask in-place. Instead, it returns a new BlockMask instance where invidual tensor attributes may or may not be moved to the specified device, depending on their current device placement. c&|jSrL)r|)rrQs r&zBlockMask.to..sadd6lr%F)r)rrArrr)rrQmapped_attributess ` r&r|z BlockMask.tors6"* LL " MM%M (  +,,r%)T)r(r))rr3) rrrr rr`__annotations__rr_mask_mod_signaturer classmethod_DEFAULT_SPARSE_BLOCK_SIZErrrSrpropertyrnrrrrrrrrrrArQstrr|r$r%r&rrsD:xsCx  ((f%%6""''V$$c3h!!%!38_%!%! %! %V, %! "&) %!v&%!F#%!$F+%!!(%!#s(O%!&%!N 04,02L2615B B B %V, B "&) B #uS#X./ B ./B eCHo.B B H  8446 p *  # : '% '&@$D-u||S01-k-r%rcv|j|kr%|jd}|j|kr%|S)Nr)rxra)rrxs r&_broadcast_to_dimr s0 %%'C- KKN %%'C- Hr%c||zdz |z|zSrr$rmultiples r&_round_up_to_multiplers L1  )H 44r%maskseparate_full_blocksc |jtjk(sJt|d}d}tjj j |d||jd|d||jd|f}|j\}}}}||zdk(sJ||zdk(sJ|j||||z|||z|}|jdddddd }|jddg } |rY||z} | | k(} | dkD| | kz} | jtj } | jtj } | | fS| dkD} | jtj } | dfS) Nr3c t|||z SrL)rr s r&padding_needed_for_multiplez@_convert_mask_to_block_mask..padding_needed_for_multiples$Q1A55r%rrYrXrrrr2rwr[) rPrArSr nn functionalpadrnviewpermuter4r|int8) r Q_BLOCK_SIZE KV_BLOCK_SIZErrBHQKVmask_block_sumfull_block_sum full_blockspartial_blockss r&_convert_mask_to_block_maskr$s :: ## # T1 %D6 88   " " ' 2 F ' 2 E   D**KAq!R | q   "" " 99 1a<r]/BM D << 1aAq DXX HN% 5$6 (1,.1PQ'***<!nn5::n6 {**'!+'***<t##r% mask_modscRtdDstdfd}|S)z9Returns a mask_mod that's the union of provided mask_modsc32K|]}t|ywrLcallabler/args r&r1zor_masks..2x}2)All inputs should be callable mask_mods: ct|jdtj}D]}||||||z}|SNr$r[)r]rArSrhq_idxkv_idxresultrr%s r&or_maskzor_masks..or_masksBRuzz2 8Dd1a77F 8 r%allr)r%r6s` r&rrs0 2 2 2FykRSS Nr%cRtdDstdfd}|S)z@Returns a mask_mod that's the intersection of provided mask_modsc32K|]}t|ywrLr(r*s r&r1zand_masks..r,r-r.ct|jdtj}D]}||||||z}|Sr0)rRrArSr1s r&and_maskzand_masks..and_masksBBejj1 8Dd1a77F 8 r%r7)r%r<s` r&rrs0 2 2 2FykRSS Or%c|jdk(sJ|j\}}}}|j||g|j}|jddddddj ||||z||z}|S)Nr3rrrr2r)rxrnexpandrreshape) block_maskrrrrrrs r&_convert_block_mask_to_maskrAs >> q  ""KAq!R"""<RAQAQRJ##Aq!Q15== 1a,] 2J r%r@rrrrc |\}}t|}| t|}nd}tj|d|d|d|d||f||S)Nrrrr)rrr) r@rrrrr#r" partial_bmfull_bms r&$_create_sparse_block_from_block_maskrEst#-NK">2J=N >   # #1 1    -0 $ r%cudamod_fnrrQ_LENKV_LENrQc |d}|d}tjd||}tjd||}tjd||}tjd||} t|} t5| tj k(rh|} t | d} | tj||||||||| } tjtj| dd} | cdddS| tjk(r%|}t |d }||||| } | cdddSt#1swYyxYw) aThis function creates a mask tensor from a mod_fn function. Args: mod_fn (Union[_score_mod_signature, _mask_mod_signature]): Function to modify attention scores. B (int): Batch size. H (int): Number of query heads. Q_LEN (int): Sequence length of query. KV_LEN (int): Sequence length of key/value. device (str): Device to run the mask creation on. Returns: mask (Tensor): A mask tensor with shape (B, H, M, N). Nrr)rQ)r)r:FTr$) rAr_r9r rr!rEzerosrbisneginfr"AssertionError)rGrrrHrIrQrr2mnmod_type score_modrtrrs r&rr s5* y y  Q&)A Q&)A Qf-A Qv.AV$H " ! (22 2I' $?IEKK1eVFKQPQSTVWXC;;u~~c2E4@D ! !*33 3H&x;HAq!Q'D ! !!  ! !s>A2D8:.D82D88Erc t|}|tjk(s Jd||d}|d}t|tr|} |} n|\} } |r?t j dttjt|||||||St||||||} t| | | d\} } t| | f|||f| | }|S)aThis function creates a block mask tuple from a mask_mod function. Args: mask_mod (Callable): mask_mod function. This is a callable that defines the masking pattern for the attention mechanism. It takes four arguments: b (batch size), h (number of heads), q_idx (query index), and kv_idx (key/value index). It should return a boolean tensor indicating which attention connections are allowed (True) or masked out (False). B (int): Batch size. H (int): Number of query heads. Q_LEN (int): Sequence length of query. KV_LEN (int): Sequence length of key/value. device (str): Device to run the mask creation on. BLOCK_SIZE (int or tuple[int, int]): Block size for the block mask. If a single int is provided it is used for both query and key/value. Returns: BlockMask: A BlockMask object that contains the block mask information. Example Usage: .. code-block:: python def causal_mask(b, h, q_idx, kv_idx): return q_idx >= kv_idx block_mask = create_block_mask(causal_mask, 1, 1, 8192, 8192, device="cuda") query = torch.randn(1, 1, 8192, 64, device="cuda", dtype=torch.float16) key = torch.randn(1, 1, 8192, 64, device="cuda", dtype=torch.float16) value = torch.randn(1, 1, 8192, 64, device="cuda", dtype=torch.float16) output = flex_attention(query, key, value, block_mask=block_mask) z4create-block_mask requires a mask_mod function! Got ra_compile flag on create_block_mask was originally added to work around a torch.compile limitation. That limitation has since been addressed. So, to compile create_block_mask, we suggest doing torch.compile(create_block_mask). This still works for now, but will be removed in the future.T)rrr)r9rr"rr`warningswarnDeprecationWarningrAcompilerrr$rE)rrrrHrIrQr_compilerPrr mask_tensorpartial_block_maskfull_block_maskr@s r&rr8s PX&H%...I =hZHI.y y *c"! " &0# m  m  0u}}./ aE66:  h1eVVDK*E!#! +' 6 _-  J r%querykeyc|j}tjtjgdtj |tj gdtj |tdS)zDefault block mask for flex attention. If users don't specify any block sparse mask info, we create this empty block sparse mask. Which creates a BlockMask with 1 block that is the full length of the query and key tensors. )rrrr\)rrrr)rr)rdrerr)rQrrrAonesr^rK_LARGE_SPARSE_BLOCK_SIZE)r[r\rQs r&_create_empty_block_maskr`sS \\F  # #jj%++fM;;|5;;vN+ $ r% orig_mod_funcq_ntkv_nt is_score_modc4 d}|j |j| |jj|jdz  ||ur n-||jj|jdz |r  fd}|S fd}|S)aAdapter to convert a score_mod / mask_mod to be NJT-compatible. The given mod func should be written as if operating over a single sequence at a item. This adapter will handle conversion from indices operating over a "stacked sequence" of length ``sum(S)`` for sequence length ``S`` in the NJT to "sequence relative" indices in range ``[0, S)``. Args: orig_mod_func (Callable): Function to modify attention scores. It takes four or five arguments, depending on whether a mask_mod or score_mod func is passed. q_nt (torch.Tensor): Jagged layout nested tensor (NJT) that defines the sequence length structure for query. kv_nt (torch.Tensor): Jagged layout nested tensor (NJT) that defines the sequence length structure for key / value. is_score_mod (bool): Indicates whether the mod function is a score_mod. Returns: nt_score_mod: An NJT-compatible version of orig_score_mod ctj||jtj}tj||ddz }|S)N)rQrPT)rightr)rAr_rQr^ searchsorted)offsets total_length range_tensorseq_idxs r&_build_seq_idxz0_nested_mod_func_adapter.._build_seq_idxs?|| u{{ $$Wl$G!Kr%rc  |}| |z }| |z } | |k(}tj| |||||tdS)Nz-inf)rArbr)rFrr2r3r4b_nestedq_nested kv_nestedis_same_sequence kv_offsets kv_seq_idxra q_offsets q_seq_idxs r& nt_score_modz._nested_mod_func_adapter..nt_score_modsu 'Hy5)9::HJv,>!??I(/:f3EE ;; eXq(IFf  r%cl |}| |z }| |z } | |k(} |||||zSrLr$) rr2r3r4rorprqrrrsrtrarurvs r& nt_mask_modz-_nested_mod_func_adapter..nt_mask_modsa 'Hy5)9::HJv,>!??I(/:f3EE  1h BEUU Ur%)_offsets_valuesrn _ragged_idx) rarbrcrdrmrwryrsrtrurvs ` @@@@r&_nested_mod_func_adapterr}s4  IJy$,,*<*= kv_idx block_mask = create_nested_block_mask(causal_mask, 1, 1, query, _compile=True) output = flex_attention(query, key, value, block_mask=block_mask) .. code-block:: python # shape (B, num_heads, seq_len*, D) where seq_len* varies across the batch query = torch.nested.nested_tensor(..., layout=torch.jagged) key = torch.nested.nested_tensor(..., layout=torch.jagged) value = torch.nested.nested_tensor(..., layout=torch.jagged) def causal_mask(b, h, q_idx, kv_idx): return q_idx >= kv_idx # cross attention case: pass both query and key/value NJTs block_mask = create_nested_block_mask(causal_mask, 1, 1, query, key, _compile=True) output = flex_attention(query, key, value, block_mask=block_mask) zLcreate_nested_block_mask(): Expected q_nt and kv_nt to be on the same deviceFrdr)rQrrW)rQ ValueErrorrr}r{rnr|)rrrrbrcrrWs r&rrs~ } {{ell" Z    4UK   4++a/0 E--12{{  r%value return_lsec|in t|}|jdd|jdd|jdd|jddd|vsJd|d<|sotj}||d<|jj dk(xs4|jj dk(xs|jj dk(}|rd|d<|S) N PRESCALE_QKFROWS_GUARANTEED_SAFEBLOCKS_ARE_CONTIGUOUSWRITE_DQTOUTPUT_LOGSUMEXPcpu)dict setdefaultrAis_grad_enabledrQtype)r[r\rrkernel_optionsoutput_logsumexpany_inputs_on_cpu_devices r&_apply_kernel_optionsr9s*1RtN7KNmU34e<5u=j$/ ^ 33 3)-N%& !002-=)* LL   & *zz%' *||  E) ! $27N- . r%c|jd|jdk7r0td|jdd|jddy)NrYzJExpect query and key/value to have the same embedding dimension but got E=z and E=.)rOrrr[r\rs r&_validate_embed_dimrYsU zz"~"%B( ~Q @   r%c|jjdk7r=|jjdk7r#td|jjdyy)zTODO: Remove once non cuda/cpu devices support is added We only need to check query since we have already that q,k,v are on the same device rFrzOFlexAttention is only supported on CUDA or CPU devices. Found input tensors on z device.N)rQrrrs r&_validate_devicerksZ ||F"u||'8'8E'A &&+ll&7&7%8 B  (B"r%c&|j|jk7s|j|jk7r td|jr |j0|jr |j|jr|j tdyy)NzFlexAttention does not support mixed nested tensor / non-nested tensor inputs. Please file an issue requesting this if it is important to you.zFlexAttention does not support nested tensors that are non-contiguous with holes. Please file an issue requesting this if it is important to you.) is_nestedr_lengthsrs r&_validate_nestednessrvs #--'3==EOO+K N  U^^7 MMcll6 OO : N  !;Or%rQscale enable_gqarc  t|||t|||t|||t||||j dk7s&|j dk7s|j dk7r t d|sS|j d|j dk7r0td|j dd|j dd|r<|j d} |j d} | | zdk7rtd | d| d |j d|j dk7r|0td |j dd |j dd |jj d|j dk7rLtd|j dd|j dd|jj dd |t}nV|jrJ|j |j|j |jk(r|n|} t||| d}| t||}|jdtk(r|jdtk(rn|jrz|j d|j"j |jdz k7r}t%d|j d|j"j |jdz |j d} |j d} |j d| kDs|j d| kDr=td|j d|j dd|j dd|j d| kr|j d| ks(|j d| krQ|j d| kr=td|j d|j dd|j dd|j d| k(sJ|j d| k(sJ|'dt'j(|j dz }|j*|jj*k7r0t%d|j*d|jj*d t-|||||}t.j0j3r|||fD]B}t.j4j7|dt.j4j7|dDt9|||||j;||\}}|r||t'j<dzfS|St.j4j?s t%d dd!l m!}d"}tE5t.j4jFjI5tK5tM5}|r ||}nd#}t/jN||d$|||||j;||\}}|r>||t'j<dzfcd d d cd d d cd d d cd d d S|cd d d cd d d cd d d cd d d S#1swYnxYw d d d n #1swYnxYwd d d n #1swYnxYwd d d y #1swYy xYw)%a# This function implements scaled dot product attention with an arbitrary attention score modification function. This function computes the scaled dot product attention between query, key, and value tensors with a user-defined attention score modification function. The attention score modification function will be applied after the attention scores have been calculated between the query and key tensors. The attention scores are calculated as follows: The ``score_mod`` function should have the following signature: .. code-block:: python def score_mod( score: Tensor, batch: Tensor, head: Tensor, q_idx: Tensor, k_idx: Tensor ) -> Tensor: Where: - ``score``: A scalar tensor representing the attention score, with the same data type and device as the query, key, and value tensors. - ``batch``, ``head``, ``q_idx``, ``k_idx``: Scalar tensors indicating the batch index, query head index, query index, and key/value index, respectively. These should have the ``torch.int`` data type and be located on the same device as the score tensor. Args: query (Tensor): Query tensor; shape :math:`(B, Hq, L, E)`. key (Tensor): Key tensor; shape :math:`(B, Hkv, S, E)`. value (Tensor): Value tensor; shape :math:`(B, Hkv, S, Ev)`. score_mod (Optional[Callable]): Function to modify attention scores. By default no score_mod is applied. block_mask (Optional[BlockMask]): BlockMask object that controls the blocksparsity pattern of the attention. scale (Optional[float]): Scaling factor applied prior to softmax. If none, the default value is set to :math:`\frac{1}{\sqrt{E}}`. enable_gqa (bool): If set to True, enables Grouped Query Attention (GQA) and broadcasts key/value heads to query heads. return_lse (bool): Whether to return the logsumexp of the attention scores. Default is False. kernel_options (Optional[Dict[str, Any]]): Options to pass into the Triton kernels. Returns: output (Tensor): Attention output; shape :math:`(B, Hq, L, Ev)`. Shape legend: - :math:`N: \text{Batch size} ... : \text{Any number of other batch dimensions (optional)}` - :math:`S: \text{Source sequence length}` - :math:`L: \text{Target sequence length}` - :math:`E: \text{Embedding dimension of the query and key}` - :math:`Ev: \text{Embedding dimension of the value}` .. warning:: `torch.nn.attention.flex_attention` is a prototype feature in PyTorch. Please look forward to a more stable implementation in a future version of PyTorch. Read more about feature classification at: https://pytorch.org/blog/pytorch-feature-classification-changes/#prototype r3z-NYI: query, key, and value must be 4D tensorszGExpect query and key/value to have the same number of heads but got Hq=z and Hkv=z&. Try setting enable_gqa=True for GQA.rrzMExpect number of query heads to be a multiple of kv heads for GQA but got Hq=rNzlExpect query and key/value to have the same batch size, or non-none block_mask, but got block_mask=None, Bq=z , and Bkv=zxExpect query and key/value to have the same batch size, or block_mask and query to have the same batch size, but got Bq=z, Bkv=z, B_block_mask=TrrXzblock_mask of shape zJ is not compatible with nested tensor input with total sequence length of rYz,block_mask was created for block_mask.shape=z but got q_len=z and kv_len=zz. As the block mask was created for a smaller length than you're using it for, you likely need to create a new block mask.ad. As the block mask was created for a larger length than you're using it for, you can either 1. create a new block mask with the correct length, or 2. 'adjust' the existing block mask to the correct length by calling block_mask._adjust(q_len, kv_len). This essentially 'crops' the block mask to the upper left corner, which does not work for all mask_mods!g?z=Expect q/k/v and block_mask to be on the same device but got z and rz&flex_attention requires dynamo support)+make_eager_backend_with_torch_function_modect|i|SrL)flex_attention_hop)argskwargss r&_flex_attention_hop_wrapperz3flex_attention.._flex_attention_hop_wrapper@s!42622r%eager)backend fullgraph)(rrrrrxNotImplementedErrorrOrrdrMrr|r}r`rr_rnr{rmathsqrtrQrrAcompileris_dynamo_compiling_dynamo mark_staticrrlogis_dynamo_supported torch._dynamo.backends.debuggingrr utilsdisable_cache_limitr r rV)r[r\rrQr@rrrrHqHkvkvblock_mask_q_lenblock_mask_kv_lenrrtlserr metadata_moders r&r r s@U+sE*UC'U+ yy{a3779>UYY[A-=!"QRR EJJrNchhrl:**R.)388B<.A3 4   ZZ]hhqk 8q= T3%q2  zz!} #  //4zz!}oZQR }TUW   # # ( ( +uzz!} <#jjm_F388A;-zOgOgOlOlmnOoNppqs   zz%++,9J9J0KK  -YrPTU -eS9  a $<<  ! !! $(@ @    B 5==#5#5e6G6G!6K#L L&z'7'7&891161C1CEDUDUXYDY1Z0[]  &++B/&,,R0 ::b>, , ?P0P>z?O?O>PP_`e`j`jkm`n_oo{|}E}EFH}I|JJKK  JJrN- -#((2,BS2Sjjn 00SXXb\DU5U>z?O?O>PP_`e`j`jkm`n_oo{|}E}EFH}I|JJuu zz"~!1111xx|0000 }dii 2// ||z//666||nE**B*B*I*I)J! M  +    N ~~))+e$ -A MM % %a , MM % %a , -& 3y**=*=*? S dhhqk)) )J == , , .CDD 3  # ]] 4 4 6 #>@ #>@#M$"M)##* u}}3WPT !"++-& HC""C$((1+$55'## # # # ###0 #+## # # # ###### # # # # # ####sy&)[ Z9 Z$%AZ  Z$ Z9 [(Z ) Z$2 Z9; [Z Z$ Z9$Z- )Z90 [9[ >[[)r$rF)rF)NNNFFN)Er rr,rrrrSenumrtypingrrrrrAr,torch._dynamo._trace_wrapped_higher_order_opr &torch._higher_order_ops.flex_attentionr rtorch._higher_order_ops.utilsr "torch.fx.experimental.proxy_tensorr r torch.nn.attention._utilsrrtorch.utils._pytreer__all___score_mod_signaturerrr9rr`listrSrErMrrr_rurrrrr rr$rrrArErrrr`r}rrrrrrrr$r%r&rsP 11 PW>P-   H& PQ?GH  )h)#4)2)+01 ++ (3-$ %+ (3-$ %+Chsm,,- +  +\        J J JJ J  J!"f<U66>%:6&6v6      }-}-@ 5 ,,!& +$ +$ +$  68F# #$ +$\ , 1D  - 2E  -+  $33 fhv../x sCx     D ,! &(;; <,!},!},!  ,!  ,!  ,! ,!j.H M!M}M}M  M  M  Mc5c?*+MM` F  I L-/BBCL ,,L <<L L  !4 45 Lh%).H P!P}P}P ,, P ELL ! P c5c?*+ PPf '-;?@ v F 6 $ F     V F .15&*!/3R# R# R# R#,- R# # R# E? R#R#R#T#s(^,R# 65( ()R#r%