+ j>aRbt_0t^RIt^RIt^RIt^RIt^RIt^RIt^RIHt^RI H t ^RI H t ^RI Ht^RIHt^RIHtHt^RIHt^R IHtHt^R IHt^R IHtHtHtHtH t H!t!H"t"H#t#^R I$H%t%]'d^R I&H't'^RI(H)t)^RI*H+t+] !4'd ^RI,t,^RI-H.t.]P^!]04t1R] R] R] R2t2]PfPiRR4Pk4t6]6]9t7RRRRRRRRRRRRR RR!R"R#R"/ t8R$R%/t9]!4'dW^R&I:H;t;Ht>H?t?H@t@HAtAHBtB^R'I:HCtD^R(I:HEtF^R)I:HGtH^R*I:HItJR+R,ltGR-R.ltIR/R0ltKRsL]^kR1R2ltMR3R4ltNRcR5ltOM2R6t7R7tGR8tIR9tK!R:R;4t=!R<R=4t>!R>R?4tPR@tARAt@RBtORCRDRERF^/RGRDRHRF^/RIRDRJRKRL/RMRDRNRF^//tQ]^k^^^^/tR]^k^tS/tT]^kRORPltURQRRltVRdRSRTlltW]T3RURVlltXRcRWRXlltEReRYRZlltCRcR[R\lltYR6sZ] R]4t[R^R_lt\R`Ralt].RfOt^R#)gNCallable)contextmanager)Path) ModuleType) TYPE_CHECKING)!get_checkpoint_conversion_mapping®ister_checkpoint_conversion_mapping)register_patch_mapping)ENV_VARS_TRUE_VALUESlogging)is_flash_attention_requested)KERNELS_MAX_VERSIONKERNELS_MIN_VERSIONis_kernels_availableis_rocm_platformis_torch_availableis_torchdynamo_compilingis_torchdynamo_exportingresolve_internal_import)flash_attention_forward)PretrainedConfig)PreTrainedModel) KernelConfigzh`kernels` is either not installed or uses an incompatible version. Please install a compatible version (z <= version < z), e.g. `pip install kernels==`USE_HUB_KERNELSYES chunk_kdazops.kdafused_recurrent_kdachunk_gated_delta_rulezops.gated_delta_rule fused_recurrent_gated_delta_rule mamba_split_conv1d_scan_combinedzops.triton.ssd_combinedselective_state_updatez!ops.triton.selective_state_updatemamba_chunk_scan_combinedmamba_inner_fnzops.selective_scan_interfaceselective_scan_fnflazflash-linear-attention)CUDAPropertiesDeviceLayerRepositoryLocalLayerRepositoryModeregister_kernel_mappingreplace_kernel_forward_from_hubuse_kernel_mapping) get_kernel) kernelize)use_kernel_forward_from_hub)use_kernelized_funcc$V^8dQhR\/# layer_namestr)formats"f/home/luca/screen-checking/.venv/lib/python3.14/site-packages/transformers/integrations/hub_kernels.py __annotate__r<is###cn\'d \V4#\PR\ 24R#)Fkernels hub usage is disabled through the environment USE_HUB_KERNELS=cV#Nclss&r;-use_kernel_forward_from_hub..psr=)_kernels_enabled$_kernels_use_kernel_forward_from_hublogger warning_once_TRANSFORMERS_USE_HUB_KERNELSr7s&r;r2r2is5  7 C C   XYvXwx # "r=cPV^8dQhR\\,\,/#)r6 module_names)listr)r:s"r;r<r<rs # #$x.8*C #r=c\'d#\V\4'dV.p\V!#\P R\ 24R#)r?cV#rArBrCs&r;rE%use_kernelized_func..{rGr=)rH isinstancer_kernels_use_kernelized_funcrJrKrL)rOs&r;r3r3rsI  ,11 ,~ /> >   XYvXwx # "r=c$V^8dQhR\/#r5r8)r:s"r;r<r<}s # #S #r=c\'d!\PR4\V4#\PR\ 24R#)z`use_kernel_func_from_hub` is deprecated in transformers v5.16 and will be removed in the future. Please use `use_kernel_forward_from_hub` instead.r?cV#rArBrCs&r;rE*use_kernel_func_from_hub..rGr=)rHrJrKrIrLrMs&r;use_kernel_func_from_hubrZ}sK     f 8 C C   XYvXwx # "r=c$V^8dQhR\/#r6returndict)r:s"r;r<r<sZZ4Zr=c </RR\RR^R7/bR\R\^y^yR7R7\RRRR R 7/bR R\P\R R ^R7\P \R R ^R7//bR R\P\R R ^R7\P \R R ^R7//bRR\P\RR^R7\P \RR^R7//bRR\P\RR^R7\P \RR^R7//bRR\P\R R^R7\P \R R^R7//bRR\P\R R^R7\P \R R^R7//bRR\P\R R^R7\P \R R^R7//bRR\P\R R^R7\P \R R^R7//bRR\P\R R^R7\P \R R^R7//bRR\P \RRRR R 7//bRR\P \P ,\RR^R7\P\P ,\RR^R7//bRR\P \P ,\RR ^R7\P\P ,\RR!^R7//bR"R\P\P ,\RR#^R7//bR$R\P\RR%^R7\P \RR%^R7/R&\P\RR%^R7\P \RR%^R7/R'\P \R(R$^R7/R)\P \R*R$^R7/R+\P\RR%^R7\P \RR%^R7//bR,R\P\RR-^R7\P \RR-^R7//bR.R\P\R/R.^R7\P \R/R.^R7/R&\P\R/R.^R7\P \R/R.^R7/R'\P \R/R.^R7/R0\P \R/R1^R7//R2R\P \P ,\R3R2^R7//R4R\P \P ,\R3R4^R7//R5R\P \P ,\R3R5^R7//R6R\P \P ,\R3R7^R7//R8R\P \P ,\R3R9^R7//R:R\P \P ,\R3R:^R7//R;R\P\RR<^R7\P \RR<^R7//R=R\P\RR>^R7\P \RR>^R7//R?R'\P \R@RA^R7/R\R@RA^R7R&\P \RBRA^R7//RCR\P\P ,\RRD^R7/// CpV#)EMultiScaleDeformableAttentioncudaz!kernels-community/deformable-detr)repo_idr7versionQwen3_5GatedDeltaNet)min_capabilitymax_capability)type propertieszAtlas-Inference/gdn(ef12347fc77d6ddf1cb72c0bd0af1c7d6cc69172T)rcr7revisiontrust_remote_codecausal_conv1d_fnzkernels-community/mamba-ssmcausal_conv1d_updater zkernels-community/flar!recurrent_gated_delta_ruler$r"r%r&r#EsmFold2TriangleMultiplicationzbiohub/esmfold2-trimulESMFold2TriangleMultiplication(9bcafd5b29a6c81645ae299d5364f5b9e503aca8 SwiGLUMLPzkernels-community/liger-kernelsLigerSwiGLUMLPLigerTiledSwiGLUMLPGeGLUMLP LigerGEGLUMLPLigerTiledGEGLUMLPLinear LigerLinearRMSNorm LigerRMSNormrocmxpuzkernels-community/rmsnormmpszkernels-community/mlx_rmsnormnpu RMSNormGatedFusedRMSNormGatedMegaBlocksMoeMLPzkernels-community/megablockscpuCPUMegaBlocksMoeMLPFastGELUzkernels-community/activation QuickGELUNewGELUSiLUSiluGeLUGeluGeluTanhrchunk_kimi_delta_attentionrrecurrent_kimi_delta_attentionrotary_pos_embzkernels-community/rotaryapply_rotary_transformerszkernels-community/aiter-ropeForCausalLMLossLigerForCausalLMLossLayer)r*r)r(r,TRAINING INFERENCE TORCH_COMPILE)_KERNEL_MAPPINGs r;_build_kernel_mappingrsWi +?>.Wi " #-SQTU#15G&*  %#Wi : MM? =#5 !$ NNO =#5 !%  !;Wi V #MM? =#9 !$ NNO =#9 !%  %WWi r %MM? 7#; !$ NNO 7#; !%  'sWi N /MM? 7#? !$ NNO 7#? !%  1OWi j (MM? =#> !$ NNO =#> !%  *kWi F /MM? =#E !$ NNO =#E !%  1GWi b MM? =#3 !$ NNO =#3 !%  cWi ~ MM? =#6 !$ NNO =#6 !%  "Wi Z %MM? =#; !$ NNO =#; !%  '[Wi v -NNO 8#C!K*. % /wWi J NNT%7%77 A#3 !: MMD$6$66 A#8 !9  KWi f NNT%7%77 A#2 !: MMD$6$66 A#7 !9  gWi B MMD$6$66 A#0 !9CWi T MM? A#1 !$ NNO A#1 !% MM? A#1 !$ NNO A#1 !% NNO ;#, !%NNO ?#, !%MM? A#1 !$ NNO A#1 !% Q4UWi ~ MM? 7#6 !$ NNO 7#6 !%  Wi Z MM? >#5 !$ NNO >#5 !% MM? >#5 !$ NNO >#5 !% NNO >#5 !%NNO >#8 !%A'!P NNT%7%77 >#- !: NNT%7%77 >#. !: NNT%7%77 >#, !: NNT%7%77 >6[\: NNT%7%77 >6[\: NNT%7%77 >:_`: MM? 7#? !$ NNO 7#? !%   "MM? 7#C !$ NNO 7#C !%  $ NNO :Gblm% 6C^hiNNO >Kfpq% MMD$6$66 ANist9 a Wi r r=c$V^8dQhR\/#r\r^)r:s"r;r<r<+s%%T%r=c2\f \4s\#)zPReturn the default transformers kernel mapping, building it lazily on first use.)_KERNEL_MAPPING_CACHErrBr=r;get_kernel_mapping_transformersr+s ! ($9$; !$$r=c8Vf \4p\V4R#rA)rr-)mappings&r;$register_kernel_mapping_transformersr2s ?57G(r=Fc RpV#)cV#rArBrCs&r; decorator.use_kernel_forward_from_hub..decorator=Jr=rBargskwargsrs*, r;r2r2< r=c RpV#)cV#rArBrCs&r;r&use_kernelized_func..decoratorCrr=rBrs*, r;r3r3Brr=c RpV#)cV#rArBrCs&r;r+use_kernel_func_from_hub..decoratorIrr=rBrs*, r;rZrZHrr=c,a]tRtRtoRtRtRtVtR#)r*iNc\R4h)zNLayerRepository requires `kernels` to be installed. Run `pip install kernels`. RuntimeErrorselfrrs&*,r;__init__LayerRepository.__init__Osop pr=c\R4h)zMLayerRepository requires `kernels` to be installed. Run `pip install kernels.NotImplementedErrorrs&r;loadLayerRepository.loadRs%&uv vr=rBN__name__ __module__ __qualname____firstlineno__rr__static_attributes____classdictcell__ __classdict__s@r;r*r*Ns q w wr=r*c,a]tRtRtoRtRtRtVtR#)r+iUc\R4h)zSLocalLayerRepository requires `kernels` to be installed. Run `pip install kernels`.rrs&*,r;rLocalLayerRepository.__init__Vstu ur=c\R4h)zRLocalLayerRepository requires `kernels` to be installed. Run `pip install kernels.rrs&r;rLocalLayerRepository.loadYs%d r=rBNrrs@r;r+r+Us v  r=r+c&a]tRtRtoRtRtVtR#)FuncRepositoryi^c\R4h)zMFuncRepository requires `kernels` to be installed. Run `pip install kernels`.rrs&*,r;rFuncRepository.__init___sno or=rBN)rrrrrrrrs@r;rr^s p pr=rc\R4h)z^replace_kernel_forward_from_hub requires `kernels` to be installed. Run `pip install kernels`.rrrs*,r;r.r.bs l  r=c\R4h)zVregister_kernel_mapping requires `kernels` to be installed. Run `pip install kernels`.rrs*,r;r-r-gssttr=c\R4h)zcregister_kernel_mapping_transformers requires `kernels` to be installed. Run `pip install kernels`.rrs*,r;rrjs q  r=zfinegrained-fp8rcz!kernels-community/finegrained-fp8rdz deep-gemmzkernels-community/deep-gemmz sonic-moezkernels-community/sonic-moerkz ep-supportnvfp4zkernels-community/nvfp4-gemmc>V^8dQhR\R,R\/#)r6attn_implementationNr])r9bool)r:s"r;r<r<s3:$r=cLVRJ;'d\P!RV4RJ#)zJCheck whether `attn_implementation` matches a kernel pattern from the hub.Nz'^[^/:]+/[^/:]+(?:@[^/:]+)?(?::[^/:]+)?$)research)rs&r; is_kernelrs3 4' c c II@BU V^b br=c0V^8dQhR\R\/#)r6rcr])r9int)r:s"r;r<r<s((S(S(r=ct\P4Fwr\WR7'gKVu# \#)zpReturn the major version of the hub kernel repo `repo_id` to load, e.g. `3` for `kernels-community/flash-attn2`.)"requested_attention_implementationrd)"_FLASH_ATTN_KERNEL_VERSION_MAPPINGitemsr_DEFAULT_ATTN_KERNEL_VERSION)rcflash_attn_versionkernel_versions& r;get_attn_kernel_versionrs3.P.V.V.X* '7 o o! !/Y ('r=c dV^8dQhR\R\R,R\R\R,/#)r6rattention_wrapperNallow_all_kernelsr])r9rrr)r:s"r;r<r<s;LLL19DL\`L$Lr=c^RIHp^RIHpRV9dVP R4^,MTp\ V4'gR#\ 4'g\\4hRV9d%VP R4wrgVP4pMTpRpVP4pVPR4wrhp VP4pV 'dV P4MRp Rp V f \V4p \WiWR7p R p \T R 4'dTf\ pTpMk\T R 4'd^RIHpTeTMTpRp MH\T R4'd(\T R4'dTeTM T P&pRp MTe \)Y4pTP+TX4TP+YT ,4T # \dh\dp \RT R T R 24hRp ? ii;i)a Load and register the kernel associated to `attn_implementation`. Args: attn_implementation: A string, usually a kernel repo like "kernels-community/flash-mla". attention_wrapper: a callable for the wrapper around the attention implementation. In `transformers` we have a wrapper around the `flash_attn_var_len` call, and the same goes for `sdpa` and `eager`. They just prepare the arguments properly. This is mostly used for continuous batching, where we want the `paged` wrapper, which calls the paged cache. allow_all_kernels (`bool`, optional): Whether to load kernels from unverified hub repos, if it is a custom kernel outside of the `kernels-community` hub repository. )ALL_MASK_ATTENTION_FUNCTIONS)ALL_ATTENTION_FUNCTIONS|N:@rkrdrz-An error occurred while trying to load from 'z': .flash_attention_2flash_attn_varlen_funcsparse_atten_func)msa_attention_forwardsdpaflash_attn_forwardsupports_flash_attn) masking_utilsrmodeling_utilsrsplitrr ImportError_MISSING_KERNELS_MESSAGEstrip partitionrr0 ValueError Exceptionhasattrr msa_attentionrrgetattrregister)rrrrractual_attn_namerc kernel_name_revrdkernelemask_implementationkernel_functionrs&&& r;load_and_register_attn_kernelr s =8.get_devices# & %5%7%7 K;''r=N)inherit_mapping)devicer$T)rrrtrainingr,rrr+rh kernel_configuse_local_kernelr*r/kernel_mapping_kernels_kernelize _use_kernels)r#r$r(r+r*s&& r;r1r1s  ! !233( "'4>>dlT]]X\D  )) *F &#11BBBjjuGZGZGjGj  3 3 B BTc d d u$ ?e d 5d;E e ds 'D D$ c V^8dQhR\R\R,R\\,R,R\R\/#)r6rrkNrdrr])r9rrr)r:s"r;r<r<,sGDj3Y    r=cx^RIHp\4'g\\4hRRRVRV/p\ WW%VR7#)r6) __version__ framework transformersrdrc)rkrd user_agentrl)r4rrrget_kernel_hub)rrkrdrr4r7s&&&& r;r0r0,sD   ! !233~y+yR]^J bs r=cJV^8dQhR\R\R\R,/#)r6rpackage internal_pathNr8)r:s"r;r<r<=s(<<c<C<X[^bXb<r=caaaaa\S4o\PSV4pVfSMV RS 2oVfSMS RV 2oRVVVVV3RllpV#)a The same as `use_kernel_forward_from_hub` but with the optional fallback to an original package if it exists, e.g., FLA for Gated Delta Rule, mamba-ssm for mamba2, etc. This combines all options with kernels, enabling kernels on top of the original package if requested as well. The order of priority is 1. Hf kernels (if requested) 2. Original package 3. Torch only path rc0V^8dQhR\R\/#)r6torch_functionr]r)r:s"r;r<.__annotate__Os(1(1((1x(1r=c<aaaaRo\P!S 4p\VS4oSf+SS 8wd$\P!S4p\VSR4oSfSMSo\ \ P!S4P4oSSJo\P!S4TTTTT T3Rl4pS !T4# \dSoLpi;i SfSoiSoi;i)Nc R<S'd\4'd S !V/VB#S'gF\4'g6\PS S 4p\P RS RV RV R24VP 4UUu/uFwr4VS9gKW4bK pppS!V/VB#uuppi)rzC` is falling back to its reference PyTorch implementation because `z>` is not installed. This is correct but much slower; install `z` for the optimized kernel.)rr_PACKAGE_TO_DISTRIBUTIONrrJrKr) rr distributionkvapplicable_paramsrimplementationis_new_implementationr;r?s *, r;wrappedJuse_kernel_func_from_hub_with_fallback..decorator..wrappedbs%)A)C)C%t6v66(1I1K1K 8;;GWM ## {#$~&$~%@B (.||~P~tq>O9Odad~FP!4262 2Qs  B#B#) rrrrrtupleinspect signature parameters functoolswraps) r?r rJrGrHrIfull_func_pathfull_module_pathrkernel_wrapper_decoratorr;s f @@@r;r9use_kernel_func_from_hub_with_fallback..decoratorOs Z,,W5F4V^LN%*:g*E"001AB!(D!A0>/E^>N"'"3"3N"C"N"NO .n D  ( 3 3 ) 3((00? ,+N ,0>/E^N>Ns$AB:: C C C  C C)r2_KERNELS_INTERNAL_PATH_MAPPINGSr)rr;r<rrRrSrTsff& @@@r;&use_kernel_func_from_hub_with_fallbackrW=sg ;9E477 =QM"/"7Y aPY{=[N"/"7wy-=Y(1(1T r=c#."RsRxRsR# Rsi;i5i)z Context manager used to adjust the value of the global `ALLOW_HUB_KERNELS`. This is needed, as this argument cannot be forwarded directly to the `__init__` of the models, where we set the attention implementation. TNF)rrBr=r;allow_all_hub_kernelsrYs!"  "Esc^V^8dQhR\R\\,R\R\/#)r6 parent_cls child_names kernel_clsr])rhrPr9)r:s"r;r<r<s4c r=caaaVPoVVV3Rlp\RVP 2V3RV/4pRVP 2VnV#)a" Create a new class that inherits from `parent_cls` and fuses the child modules specified in `child_names with the provided `kernel_cls`. The first child in `child_names` will be replaced with the `kernel_cls`, and the rest will be replaced with `nn.Identity()` to keep the same interface. c<S!V.VO5/VBSUu.uFp\W4NK ppS!V!p\VS^,V4SR,F"p\W\P!44K$ R#uupi)r:NNN)rsetattrnnIdentity) rrrnamechildrenkernel_instancer\r] original_inits &*, r; patched_init9make_parent_class_for_kernel_fusion..patched_initsnd,T,V,4?@KDGD'K@$h/k!no6OOD D .$AsA3Fusedr)rrhrr)r[r\r]rh patched_clsrgs&ff @r;#make_parent_class_for_kernel_fusionrlsW''M/z2234zmjR^E_`K!&z'>'>&?@K r=c(V^8dQhRRRRRRRR/#) r6rDztype[PreTrainedModel]configrr-rr]NrB)r:s"r;r<r<s6U/U/ U/ U/"U/ U/r=c v a"\VR4'd\VPR4'g\RVP R24hVPPp/p/pRpVP P 4EF wrx\V\\34'dRV/p\V\4'd.\VP44^8wd\RV 24hM\RV: RV: 24h\\VP444p\V\4'ddVwrV PRR4p V PR R4p V PR R 4;'g\ p V eR V /MRV /p V R V /,p W3pMTp R ^R \ /p W3pV P#R 4wpppV'd V'g\R V : RV: 24hVP$'d\'\)V4VR7pM\+R'RVRV/V BpVP-4pVf\RV: 24h\.P0PVP24pV'd\5VVP R2R4MRpVe?RVP69d.\8P:!VP<4R4pVVn\V\4'd1VfWV&EKiVPVnVWG&WVP&EM\V\4'EdVf(\RVP: RVP R24hVPVnVUu.uF pV^,NK ppVUu.uFpVPAR^4^,NK pp\\CV44^8wd\RV 24hV^,PERR4pVUu.uFpVPAR^4^,NK ppVf0\FPH!R4;_uu_4V!V4pRRR4R pVPK4Fwpo"\LPN!VV4'gK%\P;QJdV"3RlV4F 'dK R M RM!V"3RlV44'g!\RV: R V R!VP: 24hRp\SS"4p\UVVV4VVP&K V'g\R"V: R#VP: R$24h\WVRR%7\VR&4'd<\YV4p \[VP\4p!V e V V!,p!\_VV!RR%7WVP&EK WRnR#uupiuupiuupi +'giEL};i)( config_class model_typezModel z# has no config_class or model_type.NzLExpected exactly one kernel repo regardless of device/mode specificity, got zInvalid hub repo z for layer rkrdrlFrzInvalid kernel repo string ) repo_pathr7rcr7z*Could not load kernel class from hub_repo=LayoutforwardcR#rArBrs&*,r; _noop_forward?register_kernel_replacements_and_fusions.._noop_forwardsr=z Fused kernel z* requires a companion layout class named 'zLayout' in the same module.rzGAll patterns for a fused kernel must share the same parent module, got *z\w+metac3<<"TFp\SV4xK R#5irA)r).0childr s& r; ;register_kernel_replacements_and_fusions..$sK{e76511{sTzModule z* does not have the expected child modules z required for the fused kernel zNo module matched pattern z for fused kernel z3. Provide the full dotted path from the model root.) overwriteconversion_mappingrB)0rrprrrqr/rrTr9rLr_lenvaluesnextiterrrrr.r+rr*rsysmodulesrr__dict__rPrQrtkernel_layer_namersplitsetrtorchr+ named_modulesr fullmatchallrhrlr r rPrr )#rDrnr-rq patch_mapping new_mapping meta_modelr7hub_reporepo_strmetadatarkrdrl final_reporcrlayer_name_in_reporepor] kernel_mod layout_clsrvitem glob_patternspparent_patternsparent_patternr\ matched_anyrd module_clsexisting transformsr s#&&& @r;(register_kernel_replacements_and_fusionsrs 3 ' 'ws7G7G/V/V6#,,/RSTT!!,,J%'MKJ - < < B B D he - -h'H h % %8??$%* bckblm+ 0 K ~VW WX__./0 h & &!) H||J5Hll9d3G ( -@% H ] ]L] /6/B 7+U]H^H ,.?@ @H"-JH!1&9;LMH"-J)1););C)@&&0:8,kR\Q_`a a  ) ) )'w--D #-D YY[  I(VW W[[__Z%:%:; R\WZJ,?,?+@)GNbf  !iz7J7J&J __Z// 0 1 "/J  j# & &!*4J',6+>+>J ((2M %0: ++ , E * *! #J$7$7#:;(1122MO ,6+>+>J (1;<T!WWM<rs $% j418   5602   H %N+>*??]^q]rrsu !# /@% H N N P04HH 94&(>&(AA!:47 #"#;<   L# # # *.-Zx %)    wwpp u  #F STU):IqI):J U i7A F 2qq 6" !799(L^Pf/ d."<@ " " 4U/p r=