9i%SrSSKrSSKrSSKrSSKrSSKrSSKrSSKJr SSK J r J r J r J r Jr SSKrSSKJr SSK Jr SSK Jr SSKJr SS KJr SS KJr SS KJrJr SS KJ r SS K!J"r" Sr#\$"5r%\$\&\'S'\(4S\&S\&S\)\*SS4Sjjr+/SQr,\ \\\\\/\4r-\ \\\\/\4r."SS\SS9r/"SS\ 5r0"SS\ 5r1"SS\5r2S\ S\24S jr3SlS\ S!\4\ \5S"4S#\4\ \5S"4S$\\5\6\ \54S%\74 S&jjr8S'\S(\S)\S*\S+\S\4 S,jr9S(\S)\S*\S+\S\4 S-jr:S.r;S/rS0\S1\4S4jr?S5\S6\S7\5S8\54S9jr@"S:S;5rAS<rBS=rC\;\;S4S>\S?\7S\4\\ \44S@jjrDSA\.S\.4SBjrESA\.S\.4SCjrF\;\;4S\4SDjjrG\;\;4SE\4\\ \4SF\ \ SG\4\5\54SH\5SI\5S\A4 SJjjrHSmSL\\-\.4SM\ \5SN\ \5SO\5SP\5SQ\S\4SRjjrISK\;S4SF\.SM\ \5SN\ \5SO\5SP\5SQ\SS\\5\4\5\544S\A4STjjrJSU\SV\S\A4SWjrKSX\\-\.4SY\R SZ\R S[\7S\\-\.44 S\jrLS\;S4SF\.SM\ \5SN\ \5SY\R SZ\ \R SS\\5\4\5\544S\A4S]jjrMSnSU\SV\S^\S_\7S`\ \04 SajjrNSU\SV\S^\4SbjrOSU\SV\S^\4ScjrPSU\SV\S^\4SdjrQSU\SV\S^\S\4\R \R \R 44SejrRSoSSf.SU\SV\S^\Sg\ \-SE\ \ASh\ \SSi\7S_\7Sj\ \/S`\ \0S\\\4\\4\4\\1444Skjjjrg!\a SSKJr GNf=f!\a SSKJr GNf=f)pzIThis module implements the user facing API for flex_attention in PyTorch.N)Enum)AnyCallable NamedTupleOptionalUnion)Tensor) TypedDict) NotRequired)flex_attention)_set_compilation_env)DeviceLikeType))_temp_remove_metadata_torch_function_mode-_temp_remove_pre_dispatch_torch_function_mode)_validate_sdpa_input) tree_map_onlyF_WARNINGS_SHOWN warning_idmessagecategoryreturncnU[;a+[R"XSS9 [RU5 gg)z=Helper to ensure each warning is shown only once per process.) stacklevelN)rwarningswarnadd)rrrs a/var/www/html/land-doc-ocr/venv/lib/python3.13/site-packages/torch/nn/attention/flex_attention.py _warn_oncer<s-( gA6J')) BlockMaskr AuxOutput AuxRequestFlexKernelOptionscreate_block_mask create_maskcreate_nested_block_maskor_masks and_masks noop_maskcJ\rSrSr%Sr\\\S'\\\S'\\\S'\\\S'\\\S'\\\S'\\\S '\\\S '\\\S '\\\S '\\\S '\\\S'\\\S'\\\S'\\\S'\\\S'\\\S'Sr g)r$WabOptions for controlling the behavior of FlexAttention kernels. These options are passed to the underlying Triton kernels to control performance and numerical behavior. Most users will not need to specify these options as the default autotuning provides good performance. The options can be prefixed with ``fwd_`` or ``bwd_`` to apply only to forward or backward pass respectively. For example: ``fwd_BLOCK_M`` and ``bwd_BLOCK_M1``. Note: We currently do not provide any backward compatibility guarantees for these options. That being said most of these have remained pretty stable since their introduction. But We do not consider this part of the public API just yet. We think that some documentation Is better than secret hidden flags, but we may change these options in the future. Example Usage: .. code-block:: python # Using dictionary (backward compatible) kernel_opts = {"BLOCK_M": 64, "BLOCK_N": 64, "PRESCALE_QK": True} output = flex_attention(q, k, v, kernel_options=kernel_opts) # Using TypedDict (recommended for type safety) from torch.nn.attention.flex_attention import FlexKernelOptions kernel_opts: FlexKernelOptions = { "BLOCK_M": 64, "BLOCK_N": 64, "PRESCALE_QK": True, } output = flex_attention(q, k, v, kernel_options=kernel_opts) # Forward/backward specific options kernel_opts: FlexKernelOptions = { "fwd_BLOCK_M": 64, "bwd_BLOCK_M1": 32, "PRESCALE_QK": False, } output = flex_attention(q, k, v, kernel_options=kernel_opts) num_warps num_stagesBLOCK_MBLOCK_NBLOCK_M1BLOCK_N1BLOCK_M2BLOCK_N2 PRESCALE_QKROWS_GUARANTEED_SAFEBLOCKS_ARE_CONTIGUOUSWRITE_DQFORCE_USE_FLEX_ATTENTIONUSE_TMAkpackmatrix_instr_nonkdim waves_per_euN) __name__ __module__ __qualname____firstlineno____doc__r int__annotations__bool__static_attributes__r>r rr$r$Ws7'T3LC  N b b#,#,#,#,T""6&d++Q 't,,A$G*$//- / s 1%c**;c""1r r$)totalc6\rSrSr%SrSr\\S'Sr\\S'Sr g)r#zRequest which auxiliary outputs to compute from flex_attention. Each field is a boolean indicating whether that auxiliary output should be computed. Flse max_scoresr>N) r?r@rArBrCrKrFrErLrGr>r rr#r#s CJr r#cB\rSrSr%SrSr\\\S'Sr \\\S'Sr g)r"z|Auxiliary outputs from flex_attention operation. Fields will be None if not requested, or contain the tensor if requested. NrKrLr>) r?r@rArBrCrKrr rErLrGr>r rr"r"s' !C&  #'J 'r r"c$\rSrSrSrSrSrSrSrg)_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 rr>N) r?r@rArBrC SCORE_MODMASK_MODUNKNOWNrGr>r rrPrPs IHGr rPfnc[S[R"U5RR 555nUS:XdUS:XdeUS:Xa[ R $US:Xa[ R$[ R$)aHGet the type of modification function. This function inspects the number of positional arguments of the function to determine the type of modification function. If the function has 5 positional arguments, it is considered as a score_mod function. If the function has 4 positional arguments, it is considered as a mask function. c3z# UH1nUR[RR:XdM-Sv M3 g7f)rRN)defaultinspect Parameterempty).0params r _get_mod_type..s1>E ==G--33 3 >s,; ;) sumr[ signature parametersvaluesrPrTrUrV)rWnum_positional_argss r _get_mod_typeris&&r*55<<> ! #':a'?? ?a ***  ! ))) (((r prefix.suffixout_dims group_dimc/n/SQnU(aUS/- nUS/- nUHn[R"XU-U-US9nM U$)a[Used 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)NNrNNrNNro)rNNN)in_dimsrl)torchvmap)rWrjrkrlrm dimensionsdimss r_vmap_for_bhqkvrusg2OQJJ  !  J ZZTMF$:X N Ir scorebatchheadtoken_qtoken_kvcU$Nr>)rvrwrxryrzs r _identityr}0s  Lr cTURS[RURS9$)zReturns a noop mask_modr>)sizedtypedevice)new_onesrqrFr)rwrxryrzs rr*r*:s! >>rELL> IIr i@num_blocks_in_row col_indicesc^^^ URSm URSmURSSnURmUUU 4SjnUn[[U55Hn[R "USS9nM U"X5nU$)Nc>URT TS-[RS9n[R"T [RTS9R S5n[R"T[RTS9nX@R S5:n[R "XQT5nURS5X#U4'USS2ST24R5$)NrRrrrrr>) new_zerosrqint32arangerD unsqueezewherer contiguous) kv_num_blocks kv_indices dense_mask row_indices col_range index_mask valid_indicesrnum_colsnum_rowss rcreate_dense_one+_ordered_to_dense..create_dense_oneNs))(HqL )T ll8599VLVV  LL6J !8!8!<<  JHE 2<1D1DR1H -.!YhY,'2244r )rr)rp)shaperrangelenrqrr) rr batch_dimsrcreate_dense_batched_outrrrs @@@r_ordered_to_denserHs  $H  $H"(("-J  % %F5 , 3z? #$zz*>O$ 0 >C Jr c8UR[RS9nURSS9n[R"USSSS9nUR[R[R S9UR[R[R S94$)NrrdimT)r descendingstable) memory_format)torqrrdargsortcontiguous_format)rrrs r_dense_to_orderedrfs|U[[1J"2.-- tDQKU[[8O8OPu{{%2I2IJ r cN[X5n[URSS55$)Nrr)rr transpose)rrdenses r_transpose_orderedrps# / =E U__R4 55r num_blocksindices new_num_rows new_num_colscUSS2SS2SU2SU24nUSS2SS2SU24n[R"X:X5n[R"XSS2SS2SS2S4:SS9R[R5nX4$)Nrr)rqrrdrr)rrrrs r_adjust_num_blocks_and_indicesrus aM\M=L=89GAq-<-/0JZ6 QJ71a %>>BGJJ5;;WJ  r cN\rSrSr%Sr\\\4\S'\\S'\\S'\ \\S'\ \\S'\ \\S'\ \\S '\ \\S '\ \\S '\\\4\S '\ \S 'S\\\4S\S\S\ \S\ \S\ \S \ \S \ \S \ \S \\\4S \ 4Sjr \ SS\ SSS4S\S\S\ \S\ \S \\\\\44S \ \ S\ \\\4S\4Sjj5rS$S\4Sjjr\S5rSrS%SjrSrS\S\4SjrSrS\4SjrS\4SjrS&S jrS!\\R<\4SS4S"jr S#r!g)'r!ia 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_lengthsrrfull_kv_num_blocksfull_kv_indices q_num_blocks q_indicesfull_q_num_blocksfull_q_indices BLOCK_SIZEmask_modc 2UR5S:a [S5eUcS5eUcS5eUSLUSL:XdS5eUSLU SL:XdS5eXlX lX0lX@lXPlX`lXplXl Xl Xl Xl g)Nr)BlockMask must have at least 2 dimensionszkv_num_blocks must be providedzkv_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) r RuntimeErrorrrrrrrrrrrr) selfrrrrrrrrrrrs r__init__BlockMask.__init__s >> a JK K(J*JJ(%D'DD%"d*40GH U H"T)~/EF S F'*$"4.("!2,$ r NTcompute_q_blocksc UR5S:a [S5eUSLUSL:XdS5eU(a([X5upUbUce[X45upO SupOSupSup[U[5(aXU4nUbUO[ nUc-UR SUS-n UR SUS -nX4nU"UUUUUU U U U UUS 9 $) 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. rrNrNNrrrrR) rrrrrrrrrrr)rrr isinstancerDr*r)clsrrrrrrrrrrrrq_length kv_lengths rfrom_kv_blocksBlockMask.from_kv_blockss%< >> a JK K"d*40GH U H &8&S #L!-&2224F&51!>5?1!>&0 #L0: -  j# & &$1J'38  !''+jm;H"((,z!}>#))+  U r rc [U[5(dU4OUn/UQ[S5P[S5P[S5P7SSnURRSSn[S[ X#555nURUnUR UnURb.URceURUnURUnOSnSn[RUUUUURSURURSLS9$)am 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) NrSc3# UHPup[U[5(a2U*Us=::aS:aO O[X-X-S-5O[XS-5OUv MR g7f)rrRN)rrDslice)r^ins rr`(BlockMask.__getitem__..sY +!S!!*+a ! U15!%!) $qa% +sAA)rrrr)rrrrrziprrrr!rrrr)rindexpaddedsizesnew_kv_num_blocksnew_kv_indicesnew_full_kv_num_blocksnew_full_kv_indicess r __getitem__BlockMask.__getitem__bs,T!+5% 8 8e@5@%+@uT{@E$K@!D""((!, F*    !..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=r?r)rrqr rrrrrrrrrrrhasattrrr?)rrs r__repr__BlockMask.__repr__s8 6Xell3 6!!%!3!3!9!9 :;"oo3345&&3D4K4K&L%MN##01E1E#F"GH -d.?.? @AB*4>>:;<%%243I3I%J$KL""/0C0C"D!EF"oo./ % MMOC016=dmmZ6X6XDMM22^b^k^klm  r new_q_len new_kv_lencXRS-S- URS-nX RS-S- URS-n[URURX45upVURb5UR ce[URUR UU5unnOSnSnUR UUUUURUR5$)NrrR)rrrrrrrr) rrrrrrrrrs r_adjustBlockMask._adjusts!OOA$66:tq?QQ "__Q%77!;PQ@RR ,J   - )  " " .''3 33/''$$  &#&* ""& ""   "  OO MM   r c0URnSnU"U5$)zIReturns the number of elements (not accounting for sparsity) in the mask.cN[R"[RUS5$NrR) functoolsreduceoperatormul)xss r_prodBlockMask.numel.._prods##HLL"a8 8r r)rrr s rnumelBlockMask.numels  9U|r c UR5nURR5nURbX RR5- nUR 5UR S-UR S-nX1- nSSU- -$)zEComputes the percentage of blocks that are sparse (i.e. not computed)rrRd)r rrdritemr)r total_sizecomputed_blocks computed_size dense_ratios rrBlockMask.sparsitysZZ\ ,,002  " " . 66::< /nURU5 SR[U55S-nSnSn[SU"TT55n[SU"TT 55n[ STU5HFn[ STU5H-nT n UHn Xn M U"XXu-2XU-245n X+S-- nM/ US- nMH U$)Nz,  cxUR5R5R5nUS:XagUS:Xagg)NrRu█r u░)floatmeanr)section percentages rsummarize_sectionHBlockMask.to_string..create_block_vis..summarize_sections6$]]_11388: ? 1_ r cXS- -U-$rr>)abs rcdiv;BlockMask.to_string..create_block_vis..cdivsU ))r rRrr)appendjoinreversedmaxr) batch_idx descriptorsvisr$r)row_stepcol_steprccur_maskidxcharrmax_colsmax_rowsrrs rcreate_block_vis-BlockMask.to_string..create_block_vis sK   ) .))H[12T9C ! *1d8X67H1d8X67H1h1q(H5A)H(#+= ), Q\!118|3C!CDD!8OC6t 2Jr 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) rrrrD enumerate itertoolsproductrr+r,)r grid_sizelimitrr; total_visrr7r/ block_visrr9r:rrs @@@@@rrBlockMask.to_strings ]]_ *4*:*:'Xx i % % H H "_HH!* Hh  D '   *=*Qa*= > NC|  '  !VW  d()4I   Y ' yy## >s5C< rcl^[[RU4SjURSS95n[ U6$)a>Moves 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 individual tensor attributes may or may not be moved to the specified device, depending on their current device placement. c&>URT5$r|)r)rrs rBlockMask.to..Ns add6lr F)r)rrqr rr!)rrmapped_attributess ` rr BlockMask.to;s6"* LL " MM%M (  +,,r ) rrrrrrrrrrr)T)rr!))rKrc)"r?r@rArBrCrrDrEr r_mask_mod_signaturer classmethod_DEFAULT_SPARSE_BLOCK_SIZErrFrrpropertyrrrrrr r rrrrqrstrrrGr>r rr!r!sS:xsCx  ((f%%6""''V$$c3h!!#!38_#!#! #! %V, #! "&) #!v&#!F##!$F+#!!(#!#s(O#!&#!J 04,02L2615!%G G G %V, G "&) G #uS#X./ G ./G eCHo.G G G R  844E N *  # : '% '&@$D-u||S01-k-r r!c|UR5U:a'URS5nUR5U:aM'U$)Nr)rr)rrs r_broadcast_to_dimrRTs0 %%'C- KKN %%'C- Hr cX-S- U-U-$rr>rmultiples r_round_up_to_multiplerVZs L1  )H 44r maskseparate_full_blocksc UR[R:Xde[US5nSn[RR R USU"URSU5SU"URSU545nURupVpxXq-S:XdeX-S:XdeURXVXq-XU-U5nURSSSSSS 5nURSS/S 9n U(aOX-n X:Hn U S:X:-n U R[RS 9n U R[RS 9n X4$U S:n U R[RS 9n U S4$) Nrcc[X5U- $r|)rVrTs rpadding_needed_for_multiple@_convert_mask_to_block_mask..padding_needed_for_multiplegs$Q1A55r rrrrRrrSrbrr) rrqrFrRnn functionalpadrviewpermuterdrint8) rW Q_BLOCK_SIZE KV_BLOCK_SIZErXr[BHQKVmask_block_sumfull_block_sum full_blockspartial_blockss r_convert_mask_to_block_maskrm^s :: ## # T1 %D6 88   " " ' 2 F ' 2 E   D**KA! q   "" " 99 a]/BM D << 1aAq DXX HN%5$6 (1,1PQ'***<!nn5::n6 **'!+'***<t##r mask_modsc^^[ST55(d[ST35eU4SjnU$)z9Returns a mask_mod that's the union of provided mask_modsc38# UHn[U5v M g7fr|callabler^args rr`or_masks..2 x}} )All inputs should be callable mask_mods: cl>URS[RS9nTHnXE"XX#5-nM U$Nr>r)rrqrFr(hq_idxkv_idxresultrWrns ror_maskor_masks..or_masks9Ruzz2Dd177F r allr)rnrs` rr(r(s3 2 2 2 2FykRSS Nr c^^[ST55(d[ST35eU4SjnU$)z@Returns a mask_mod that's the intersection of provided mask_modsc38# UHn[U5v M g7fr|rqrss rr`and_masks..rvrwrxcl>URS[RS9nTHnXE"XX#5-nM U$rz)rrqrFr{s rand_maskand_masks..and_masks9Bejj1Dd177F r r)rnrs` rr)r)s3 2 2 2 2FykRSS Or cUR5S:XdeURup4pVUR"X!/URQ76nURSSSSSS5R X4XR-Xa-5nU$)NrcrrSrrbrR)rrexpandrareshape) block_maskrdrcrerfrgrhs r_convert_block_mask_to_maskrs} >> q  ""KA!""<RAQAQRJ##Aq!Q15== a 2J r rrrrcrdc UupV[U5nUb [U5nOSn[RUSUSUSUSX44UUS9$)NrrrR)rrr)rr!r) rrrrcrdrlrk partial_bmfull_bms r$_create_sparse_block_from_block_maskrsp#-N">2J=N >   # #1 1    0 $ r cudamod_fnrerfQ_LENKV_LENrc pUcSnUcSn[R"SXS9n[R"SX%S9n[R"SX5S9n[R"SXES9n [U5n SSKJn U "5 U [ R :XaaUn [U SS9n U "[R"XX4US9XgX5n [R"[R"U 5SS 5nUsSSS5 $U [ R:Xa Un[US S9nU"XgX5nUsSSS5 $[e!,(df  g=f) aThis function creates a mask tensor from a mod_fn function. Args: mod_fn (Union[_score_mod_signature, _mask_mod_signature]): Function to modify attention scores. B (int): Batch size. H (int): Number of query heads. Q_LEN (int): Sequence length of query. KV_LEN (int): Sequence length of key/value. device (str): Device to run the mask creation on. Returns: mask (Tensor): A mask tensor with shape (B, H, M, N). NrRr)r)TransformGetItemToIndex)r)rjFTr>) rqrri,torch._dynamo._trace_wrapped_higher_order_oprrPrTruzerosrisneginfrUAssertionError)rrerfrrrr(r|mrmod_typer score_modrrWrs rr&r&s* y y  Q)A Q)A Q-A Q.AV$HT " (22 2I' $?IEKKeFKQSTXC;;u~~c2E4@D # "*33 3H&x;HA!'D # "!  # "s7A,D'-*D'!D'' D5rc [U5nU[R:Xd SU35eUcSnUcSn[U[5(aUn Un OUupU(a>[ R "S[5 [R"[5"XX#XEU5$[XX#XE5n [U U U SS9up[X4UX44U U 5nU$)aThis function creates a block mask tuple from a mask_mod function. Args: mask_mod (Callable): mask_mod function. This is a callable that defines the masking pattern for the attention mechanism. It takes four arguments: b (batch size), h (number of heads), q_idx (query index), and kv_idx (key/value index). It should return a boolean tensor indicating which attention connections are allowed (True) or masked out (False). B (int): Batch size. H (int): Number of query heads. Q_LEN (int): Sequence length of query. KV_LEN (int): Sequence length of key/value. device (str): Device to run the mask creation on. BLOCK_SIZE (int or tuple[int, int]): Block size for the block mask. If a single int is provided it is used for both query and key/value. Returns: BlockMask: A BlockMask object that contains the block mask information. Example Usage: .. code-block:: python def causal_mask(b, h, q_idx, kv_idx): return q_idx >= kv_idx block_mask = create_block_mask(causal_mask, 1, 1, 8192, 8192, device="cuda") query = torch.randn(1, 1, 8192, 64, device="cuda", dtype=torch.float16) key = torch.randn(1, 1, 8192, 64, device="cuda", dtype=torch.float16) value = torch.randn(1, 1, 8192, 64, device="cuda", dtype=torch.float16) output = flex_attention(query, key, value, block_mask=block_mask) z4create-block_mask requires a mask_mod function! Got rRa_compile flag on create_block_mask was originally added to work around a torch.compile limitation. That limitation has since been addressed. So, to compile create_block_mask, we suggest doing torch.compile(create_block_mask). This still works for now, but will be removed in the future.T)rcrdrX)rirPrUrrDrrDeprecationWarningrqcompiler%r&rmr)rrerfrrrr_compilerrcrd mask_tensorpartial_block_maskfull_block_maskrs rr%r%sRX&H (11 1 >xjI 1 y y *c""! " &0#   m  }}./ 6:  h1VDK*E!#! +' 6 -  J r querykeyc URn[R[R"/SQ[R US9[R "/SQ[R US9[SS9$)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. )rRrRrRr)rRrRrRrR)rRrR)rrrr)rr!rrqonesrr_LARGE_SPARSE_BLOCK_SIZE)rrrs r_create_empty_block_maskrTsS \\F  # #jj%++fM;;|5;;vN+ $ r orig_mod_funcq_ntkv_nt is_score_modcB^^^^ ^ SnURm URmU"T URRURS- 5m XLaT mO-U"TURRURS- 5mU(a UUUU U 4SjnU$UUUU U 4SjnU$)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 c[R"XR[RS9n[R"XSS9S- nU$)N)rrT)rightrR)rqrrr searchsorted)offsets total_length range_tensorseq_idxs r_build_seq_idx0_nested_mod_func_adapter.._build_seq_idx}s;|| u{{ $$W$G!Kr rRc >T UnUT T U- nUT T U- nT UT U:Hn[R"UT "XX&U5[S55$)Nz-inf)rqrr )rvr(r|r}r~b_nestedq_nested kv_nestedis_same_sequence kv_offsets kv_seq_idxr q_offsets q_seq_idxs r nt_score_mod._nested_mod_func_adapter..nt_score_modsq 'Hy5)9::HJv,>!??I(/:f3EE ;; eqIFf  r ch>T UnUT T U- nUTT U- nT UT U:HnT "XAXV5U-$r|r>) r(r|r}r~rrrrrrrrrs r nt_mask_mod-_nested_mod_func_adapter..nt_mask_mods] 'Hy5)9::HJv,>!??I(/:f3EE  hBEUU Ur )_offsets_valuesr _ragged_idx) rrrrrrrrrrrs ` @@@@r_nested_mod_func_adapterrcs4  IJy$,,*<*= 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 deviceFrrR)rrr)r ValueErrorr%rrrr)rrerfrrrrs rr'r'sN } {{ell" Z    UK   4++a/0 E--12{{  r value return_lse return_auxc`Uc0O [U5nURSS5 URSS5 URSS5 URSS5 URRS:H=(d9 URRS:H=(d URRS:HnUnSnUbURnUR nSU;deSUS'U(d$[ R"5US'U(aSUS'S U;deXS 'U(aU(a [S 5eU$) Nr5Fr6r7r8TcpuOUTPUT_LOGSUMEXP OUTPUT_MAXz-Returning max scores is not supported on CPU.) dict setdefaultrtyperKrLrqis_grad_enabledNotImplementedError) rrrrkernel_optionsrany_inputs_on_cpu_device output_lse output_maxs r_apply_kernel_optionsrs9*1RtN7KNmU34e<5u=j$/  U" & ::??e # & <<   %JJ^^ **  ^ 33 3)-N%& .3-B-B-D)* #27N- . ~ -- -#-< J""QRR r cURS5URS5:wa0[SURS5SURS5S35eg)NrzJExpect query and key/value to have the same embedding dimension but got E=z and E=.)rrrrrs r_validate_embed_dimrGsU zz"~"%B( ~Q @  &r c1SknURRU;a#[SURRS35eg)zTODO: Remove once non cuda/cpu devices support is added We only need to check query since we have already that q,k,v are on the same device >rhpuxpurzTFlexAttention is only supported on CUDA, CPU or HPU devices. Found input tensors on z device.N)rrr)rrrsupported_devicess r_validate_devicerOsL6 || 11 &&+ll&7&7%8 B  2r cNURUR:wdURUR:wa [S5eUR(a URc<UR(a URcUR(aURb [S5egg)NzFlexAttention does not support mixed nested tensor / non-nested tensor inputs. Please file an issue requesting this if it is important to you.zFlexAttention does not support nested tensors that are non-contiguous with holes. Please file an issue requesting this if it is important to you.) is_nestedr_lengthsrs r_validate_nestednessr[s #--'3==EOO+K N  U^^7 MMcll6 OO : N  !;Or cS[S[4SjnS[S[4Sjn[R[R4nUR nXe;=(an [R RSL=(aK [RRS5S:=(a" [RRS5S:nU(dXU4$U"U5(dUR5nU"U5(dUR5nU"U5(d0URS S 5R5RS S 5nXU4$) a Enforce memory layouts for query, key, and value tensors. For non-FP8 dtypes, no action is taken. For FP8 dtypes, we enforce the following memory layouts: - Query tensor must be in row-major memory layout, as it will be the left-operand in the FP8 GEMM `q @ k.T`. - Key tensor must be in row-major memory layout, as it will be transposed when used as the right-operand in the FP8 GEMM `q @ k.T`, meaning it will correctly be in column-major memory layout for the GEMM. - Value tensor must be in column-major memory layout, as it will be the right-operand in the FP8 GEMM `softmax_scores @ v`. Returns the query, key, and value tensors with the enforced memory layouts. tensorrc.UR5SS:H$)NrrRstriders r is_row_major*_enforce_mem_layouts..is_row_major}}r"a''r c.UR5SS:H$)NrrRrrs r is_col_major*_enforce_mem_layouts..is_col_majorrr Nr) ) rrr) r rFrq float8_e4m3fn float8_e5m2rversionrget_device_capabilityrr)rrrrr fp8_dtypesgemm_precisionshould_enforce_mem_layouts r_enforce_mem_layoutsrns."(V(((V((  J[[N $ ? MM  d * ? JJ , ,V 4 > ? JJ , ,V 4w >  %5       "   nn   B'224>>r2F u r )rrscale enable_gqarc  ^[XU5 [XU5 [XU5 [XU5 [ XU5upnUR 5S:wd(UR 5S:wdUR 5S:wa [ S5eU(dTURS5URS5:wa0[SURS5SURS5S35eU(a<URS5n URS5n X-S:wa[S U SU S 35eURS5URS5:waUc0[S URS5S URS5S 35eURRS5URS5:waL[SURS5SURS5SURRS5S 35eUc[nOXUR(aGURUR5URUR5:XaUOUn [X0U SS9nUc [X5nURS[ :XaURS[ :XaGOUR(a|UR"SUR$RURS- 5:waA['SUR"SUR$RURS- 535eGODUR"Sn UR"SnURS5U :dURS5U:a=[SUR"SURS5SURS5S35eURS5U :aURS5U::d*URS5U ::aRURS5U:a=[SUR"SURS5SURS5S35eURS5U :XdeURS5U:XdeUc(S[(R*"URS55- nUR,URR,:wa0['SUR,SURR,S 35eU(aU b [S5eU(aU c[/S S![0S"9 [3UUUUUU 5nS#[4[6S$[84S%jn[:R<R?5(awXU4HCn[:R@RCUS5 [:R@RCUS5 ME [EUUUUURG5UU5unnnU"UUUXS&9$[H(d [/S'S(S)9 [:R@RK5(d ['S*5eSS+K&J'n S,n[Q5 [:R@RRRU5 [W5 [Y5nU(a U"U5nOS-n[H(aUnO[:RZ"UUSS.9nU"UUUUURG5UU5unnnS S S 5 S S S 5 S S S 5 S S S 5 U"WWWXS&9$!,(df  N/=f!,(df  N8=f!,(df  NA=f!,(df  NJ=f)/a1This 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)`. For FP8 dtypes, should be in row-major memory layout for optimal performance. key (Tensor): Key tensor; shape :math:`(B, Hkv, S, E)`. For FP8 dtypes, should be in row-major memory layout for optimal performance. value (Tensor): Value tensor; shape :math:`(B, Hkv, S, Ev)`. For FP8 dtypes, should be in column-major memory layout for optimal performance. 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. **Deprecated**: Use ``return_aux=AuxRequest(lse=True)`` instead. kernel_options (Optional[FlexKernelOptions]): Options to control the behavior of the underlying Triton kernels. See :class:`FlexKernelOptions` for available options and usage examples. return_aux (Optional[AuxRequest]): Specifies which auxiliary outputs to compute and return. If None, only the attention output is returned. Use ``AuxRequest(lse=True, max_scores=True)`` to request both auxiliary outputs. Returns: output (Tensor): Attention output; shape :math:`(B, Hq, L, Ev)`. When ``return_aux`` is not None: aux (AuxOutput): Auxiliary outputs with requested fields populated. When ``return_aux`` is None (deprecated paths): lse (Tensor): Log-sum-exp of attention scores; shape :math:`(B, Hq, L)`. Only returned if ``return_lse=True``. 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 rcz-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.rRrzMExpect 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=Trrzblock_mask of shape zJ is not compatible with nested tensor input with total sequence length of rz,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 z|Cannot specify both return_lse and return_aux. return_lse is deprecated, please use return_aux=AuxRequest(lse=True) instead.deprecated_return_lsezjreturn_lse is deprecated and will be removed in v2.10. Please use return_aux=AuxRequest(lse=True) instead.)rrrcV[R"S5nU=(d USL=(a URnUSL=(a URnU(aUR 5S:aX-OSnU(aUR 5S:aX%-OSnUb U[ UUS94$U(aX4$U$)zFNormalize stats and build return value (aux-aware, legacy-compatible).g@Nr)rKrL)mathlogrKrLr r") rrKrLrrln2 return_max lse_scaled max_scaleds r_finalize_outputs)flex_attention.._finalize_outputsfshhsmL:T#9#Ljnn t+E 0E0E #-#))+/SY !+ 0@0@0BQ0FJ T   ! %  ? " r )rrflex_attention_performanceaflex_attention called without torch.compile() - this will use an unfused implementation that materializes the full scores matrix instead of generating a fused kernel. SOLUTION: Use torch.compile(flex_attention)(...) If you want to debug your score_mod/mask_mod, you can set: torch.nn.attention.flex_attention._FLEX_ATTENTION_DISABLE_COMPILE_DEBUG = True This will allow you to use print statements or breakpoints. Note: This doesn't work with the backwards pass and may produce incorrect results.)rrz&flex_attention requires dynamo support)+make_eager_backend_with_torch_function_modec[U0UD6$r|)flex_attention_hop)argskwargss r_flex_attention_hop_wrapper3flex_attention.._flex_attention_hop_wrappers!42622r eager)backend fullgraph).rrrrrrrrrrr}rrrrrrrrrr sqrtrr FutureWarningrrr#rFrqcompileris_dynamo_compiling_dynamo mark_staticrr%_FLEX_ATTENTION_DISABLE_COMPILE_DEBUGis_dynamo_supported torch._dynamo.backends.debuggingrr utilsdisable_cache_limitrrr)rrrrrrrrrrHqHkvkvblock_mask_q_lenblock_mask_kv_lenrrrrKrLrr metadata_moderflex_fns rr r sZU+E*U'U+,U?E 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 -e9  a $<<  ! !! $(@ @    B 5==#5#5e6G6G!6K#L L&z'7'7&891161C1CEDUDUXYDY1Z0[]  M &++B/&,,R0 ::b>, , ?P0P>z?O?O>PP_`e`j`jkm`n_oo{|}E}EFH}I|JJKK  JJrN- -#((2,BS2Sjjn 00SXXb\DU5U>z?O?O>PP_`e`j`jkm`n_oo{|}E}EFH}I|JJuu zz"~!1111xx|0000 }dii 2// ||z//666||nE**B*B*I*I)J! M  j, \    * # B"  +    N Z(   : ~~))++e$A MM % %a , MM % %a ,% 2        !   S*! jZ   1 03a  == , , . .CDD 3   ]] 4 4 6>@>@M$G V #*<<"="'--7TX#,3!"++-&,(CjAA7 6  S* 1A@A@7 6 sU,)\ \  [<+A[+ [< \ \+ [9 5[<< \ \  \ \ \,)r>rF)rr|)NNNFFN)TrCrr[r>r rrenumrtypingrrrrrrqr r ImportErrortyping_extensionsr &torch._higher_order_ops.flex_attentionr rtorch._higher_order_ops.utilsr torch._prims_commonr"torch.fx.experimental.proxy_tensorrrtorch.nn.attention._utilsrtorch.utils._pytreerr$setrrPrE UserWarningrWarningr__all___score_mod_signaturerLr$r#r"rPrirrDlistrFrur}r*rNrrrrrr!rRrVrmr(r)rrr&r%rrr'rrrrrr r>r rr@sP == , ."X>.;-&).%ES!>I((!(-1']( (   H& PQ?GHr2 r2j( (  )h)#4)2)+01 ++ (3-$ %+ (3-$ %+Chsm,,- +  +\        J J JJ J  J!"f<U66>%:6&6v6      O-O-d 5 ,,!& +$ +$ +$  68F# #$ +$\ , 1D  - 2E  -+  $33 fhv../x sCx     D$ .! &(;; <.!}.!}.!  .!  .!  .! .!n$.H N!N}N}N  N  N  Nc5c?*+NNb F  I O-/BBCO ,,O <<O O  !4 45 On%).H X!X}X}X ,, X ELL ! X c5c?*+ XXB(, 4 4 4 4 4 $ 4n v F 6   F        V F &8 88'-8 5<<u|| 348~15&*!26b(,b b b b,- b # b E? bbb./b$b 65(% 0A*B BCbk,,++, .--.s#L9M 9 M M  MM