tjUdZddlmZddlZddlZddlZddlZddlZddlZddl Z ddl m Z m Z ddl mZmZmZddlmZmZmZddlmZmZddlmZdd lmZmZdd lmZmZm Z m!Z!ddl"Z#d d l$m%Z%d dl&m'Z'm(Z(m)Z)er ddl*Z*ddl*m+Z+e!dZ,e%j-e.Z/dZ0e(rdZ0e1Z2de3d<d~dZ4dZ5e'rdZ5ddZ6ddZ7dZ8dd Z9dd!Z:dd"Z;dd#Z ddd-Z?d.Z@dd/ZA ddd3ZBdd6ZCd7ZDd8ZEdd:ZFGd;d1tyonyestruer>0fnnoofffalserzinvalid truth value )lower ValueError)vals r5 strtoboolrI\sJ ))++C 222q 333q 3C33 4 44r7 str | Nonectt|}|drdS|drdS|drdSdS)z Tries to guess the framework of an object `x` from its repr (brittle but will help in `is_tensor` to try the frameworks in a smart order, without the need to import the frameworks). z 2 2ur7ctttdt|gng}dkr|d|fdDfd|DS)z Returns an (ordered since we are in Python 3.7+) dictionary framework to test function, which places the framework we can guess from the repr first, then Numpy, then the others. )rLrMrNNrMc g|] }|dfv| S)rM).0rApreferred_frameworks r5 z1_get_frameworks_and_test_func..s*\\\QqATVZ@[7[7[q7[7[7[r7c"i|] }|| SrWrW)rXrAframework_to_tests r5 z1_get_frameworks_and_test_func..s! 8 8 8A # 8 8 8r7)is_torch_tensoris_numpy_array is_mlx_arrayrTappendextend)rR frameworksr\rYs @@r5_get_frameworks_and_test_funcrdxs  4A66*29L8MJd""$\\\\"3\\\]]] 8 8 8 8Z 8 8 88r7boolct|}|D]}||rdSt|rdSdS)z{ Tests if `x` is a `torch.Tensor`, `np.ndarray` or `mlx.array` in the order defined by `infer_framework_from_repr` TF)rdvaluesr)rRframework_to_test_func test_funcs r5 is_tensorrjse ;1==+2244 9Q<< 44 t 5r7c6t|tjS)z/ Tests if `x` is a numpy array or not. ) isinstancerMndarrayrRs r5r_r_s a $ $$r7cFtsdSddl}t||jS)z] Tests if `x` is a torch tensor or not. Safe to call even if torch is not installed. FrN)r&r'rlTensorrRr's r5r^r^, uLLL a & &&r7cFtsdSddl}t||jS)z] Tests if `x` is a torch device or not. Safe to call even if torch is not installed. FrN)r&r'rldevicerqs r5is_torch_devicerurrr7ctsdSddl}t|tr#t ||rt ||}ndSt||jS)z\ Tests if `x` is a torch dtype or not. Safe to call even if torch is not installed. FrN)r&r'rlrOhasattrgetattrdtyperqs r5is_torch_dtyperzsf uLLL!S 5!   q!!AA5 a % %%r7c4t|rdSt|rdSt|ttt t jfrdSt|ttfr*t|dkrdSt|dSdS)zA Check if a value is array-like (includes ragged arrays) TrF) r_r^rlr8floatrerMnumberlisttuplelen_is_tensor_or_array_like)values r5rrsetut%#udBI677t%$''2 u::??4'a111 5r7 device_typerOrytorch.dtype | Noneenabled cache_enabled bool | Nonectstdddl}|dkrtS|j|s|r|j||||StS)a Context manager that only autocasts if: - `autocast` is already enabled in this context - Or this call to `maybe_autocast` has `enabled=True` This prevents `autocast` being added to the graph when it is effectively a no-op. Which makes graph splitting in `torch.compile` more flexible as it removes the requirement that partition IDs be monotonically increasing. z2`maybe_autocast` requires PyTorch to be installed.rNmeta)ryrr)r& ImportErrorr'r is_autocast_enabledautocast)rryrrr's r5maybe_autocastrs| PNOOOLLLf}} u --u~kWdeeee}}r7c8ddlm}t||jS)Nr)mlx.corecorerlarray)rRmxs r5_is_mlxrs& a " ""r7c2tsdnt|S)zZ Tests if `x` is a mlx array or not. Safe to call even when mlx is not installed. F)_is_mlx_availablerrns r5r`r`s* 955wqzz9r7"requested_attention_implementationversionint | list[int] | Nonec||td||jn|dS|3t|tr|g}t fd|DSdvS)a_ Checks whether some flavor of flash attention is requested or not. Optionally, checks for specific versions of flash attention. This is checked against one of the two arguments, i.e. either the `config` or the directly passed value `requested_attention_implementation`. Otherwise, an error will be raised (ambiguity). The different versions of flash attention are usually - Implementations based on the original flash attention repo: https://github.com/Dao-AILab/flash-attention - Kernels implementations such as: https://huggingface.co/kernels-community/vllm-flash-attn3 NzRequested attention implementation is ambiguous: Please pass either the config or the name of the attention implementation, not both.Fc3fK|]+}tjdt|zduV,dS)z .*flash.*N)rematchrO)rXv checked_attention_implementations r5 z/is_flash_attention_requested..+sBrref28L3q6613STT\``rrrrrrr7flash)rG_attn_implementationrlr8any)configrrrs @r5is_flash_attention_requestedr s@L c   +1+F((+M((/u gs # # iGrrrrjqrrrrrr 6 66r7implementationtuple[bool, str | None]cb|dS|d}||dfS)z Split the optional `paged|` prefix from an attention implementation string. Note that `None` means using the default attention implementation, which is either torch's native `sdpa` or `eager` (if `sdpa` is not implemented for that model). N)FNzpaged|)rQ removeprefix)ris_pageds r5split_attention_implementationr1s; {((22H ^00:: ::r7cNt|ttfr|St|ttfrd|DSt|t tfr4td|Drt |Sd|DSddd}t|}|D]#\}}||r|||cS$t|tj r| S|S)zP Convert a PyTorch tensor, Numpy array or python list to a python list. c4i|]\}}|t|SrW to_py_objrXkrs r5r]zto_py_obj..Es$888DAq9Q<<888r7c3dK|]+}t|tttjfV,dSN)rlr8r|rMr})rXrRs r5rzto_py_obj..Hs5CC!z!c5")455CCCCCCr7c,g|]}t|SrWr)rXos r5rZzto_py_obj..Ls*** ! ***r7c*|Srtolistobjs r5zto_py_obj..O#**,,r7c*|Srrrs r5rzto_py_obj..Prr7rLrM) rlr8r|dictritemsr~rallrdrMr}r)rframework_to_py_objrh frameworkris r5rr>sN#U|$$ + C$) * *+88CIIKK8888 C$ ' '+ CCsCCC C C 99 +*c****'&&& ;3?? 6 < < > >77 9 9S>> 71&y1#66 6 6 6 7#ry!!zz|| r7cvddd}t|ttfrd|DSt|tt frt j|St|}|D]#\}}||r|||cS$|S)zP Convert a PyTorch tensor, Numpy array or python list to a Numpy array. cr|Sr)detachcpunumpyrs r5rzto_numpy..fs&#**,,**,,2244r7c|SrrWrs r5rzto_numpy..gs#r7rc4i|]\}}|t|SrW)to_numpyrs r5r]zto_numpy..ks$77741a8A;;777r7) rlrrrr~rrMrrd)rframework_to_numpyrhrris r5rr`s 54o #h'((77399;;7777 C$ ' 'x}};3?? 6 < < > >66 9 9S>> 60%i055 5 5 5 6 Jr7 json_filec t|d5}|}dddn #1swxYwYtj|}n&#tj$rt d|dwxYw|S)zeA helper to load safe config files and raise a proper error message if it wasn't serialized correctlyzutf-8)encodingNz"It looks like the config file at 'z' is not a valid JSON file.)openreadjsonloadsJSONDecodeErrorOSError)rreadertext config_dicts r5safe_load_json_filerxsc )g . . . !&;;==D ! ! ! ! ! ! ! ! ! ! ! ! ! ! !j&&  ccca9aaabbbc s&A4 A8A8A#A7cxeZdZdZddZfdZfdZdZdZd Z d Z d Z fd Z fd Z fdZddZxZS) ModelOutputa Base class for all model outputs as dataclass. Has a `__getitem__` that allows indexing by integer or slice (like a tuple) or strings (like a dictionary) that will ignore the `None` attributes. Otherwise behaves like a regular python dictionary. You can't unpack a `ModelOutput` directly. Use the [`~utils.ModelOutput.to_tuple`] method to convert it to a tuple before. r r!c$t|dS)zRegister subclasses as pytree nodes. This is necessary to synchronize gradients when using `torch.nn.parallel.DistributedDataParallel` with `static_graph=True` with modules that output `ModelOutput` subclasses. N)r6)clss r5__init_subclass__zModelOutput.__init_subclass__s +3/////r7ctj|i|tt||jt k}|r3t |s&t|jd|jj ddSdS)Nr#z` is not a dataclass. This is a subclass of ModelOutput and so must use the @dataclass decorator.) super__init__r6rP __class__rr TypeErrorr0r1)selfargskwargsis_modeloutput_subclassrs r5rzModelOutput.__init__s$)&)))*4::666 #'.K"? " <+=+= ?__T^%<___     r7c ttt}t|st jjdtd|ddDst jjdt|dj }tfd|ddD}|rt|swt|tr| }d}n% t|}d}n#t$rd }YnwxYw|r t!|dj dt#|dj t'|D]\}}t|t(t*fr.t|d kst|dt,s,|dkr||dj <nt d |d n9t!|d|d|d|d|d<dSdS|||dj <dSdS|D]-}j|j } | | |j <.dS) zeCheck the ModelOutput dataclass. Only occurs if @dataclass decorator has been used. z has no fields.c3(K|] }|jduVdSr)defaultrXfields r5rz,ModelOutput.__post_init__..s)GGU5=D(GGGGGGr7rNz. should not have more than one required field.rc3ZK|]%}j|jduV&dSr)__dict__getnamerXrrs r5rz,ModelOutput.__post_init__..s:#h#heDM$5$5ej$A$AT$I#h#h#h#h#h#hr7TFrzCannot set key/value for z&. It needs to be a tuple (key, value).)r6rPr rrGrr1rrxrrjrlrriterrsetattrr __delitem__ enumerater~rrOrr) r class_fields first_fieldother_fields_are_noneiteratorfirst_field_iteratoridxelementrrrs ` r5 __post_init__zModelOutput.__post_init__s +4::666d|| <   J 7HHHII IGGl1226FGGGGG i 7ggghh hdLO$899 ##h#h#h#hWcdedfdfWg#h#h#h h h % );)?)?% )+t,, 1&,,..'+$$1#K00H+/(( 111+0(((1 $ 9l1o2D999##LO$8999$-h$7$7 6 6LC%ge}== WQRARARZdelmneoqtZuZuAR!889DDa!566#- kG k k k##D'!*gaj999qz-+21:WQZ( 6 6(-8\!_)***)(& ) )M%%ej11='(D$ ) )sD&& D54D5c<td|jjd)Nz$You cannot use ``__delitem__`` on a instance. Exceptionrr1rrrs r5rzModelOutput.__delitem__s!bt~?Vbbbcccr7c<td|jjd)Nz#You cannot use ``setdefault`` on a rrrs r5 setdefaultzModelOutput.setdefaults!adn>Uaaabbbr7c<td|jjd)NzYou cannot use ``pop`` on a rrrs r5popzModelOutput.pops!Zt~7NZZZ[[[r7c<td|jjd)NzYou cannot use ``update`` on a rrrs r5updatezModelOutput.updates!]$.:Q]]]^^^r7ct|tr)t|}||S||Sr)rlrOrrto_tuple)rr inner_dicts r5 __getitem__zModelOutput.__getitem__sF a   &djjll++Ja= ==??1% %r7cdt|D}||vr$|"t||t||dS)Nch|] }|j SrW)rrs r5 z*ModelOutput.__setattr__..s<<.s/II5WT5:..IIIIIIr7)r r __reduce__rr )rcallable_args remainingrrs` r5rzModelOutput.__reduce__sD!! (77%%'' '&+gg&8&8&:&:#%)IIIIF4LLIIIII) )))r7rc^tfdDS)za Convert self to a tuple containing all the attributes/keys that are not `None`. c3(K|] }|V dSrrW)rXrrs r5rz'ModelOutput.to_tuple.. s'22T!W222222r7)rkeysrs`r5r zModelOutput.to_tuples02222diikk222222r7)r r!)r r)r1r0 __qualname____doc__rrrrrrrr rrrr  __classcell__)rs@r5rrs&  0000 6)6)6)6)6)pdddccc\\\___&&&)))))((((( *****33333333r7routputtuple[list[Any], list[str]]ct|t|fSr)r~rgr)r!s r5r.r.s-   $v{{}}"5"5 55r7rg Iterable[Any]context list[str]type[ModelOutput] | Nonec H|ditt||S)NrW)rzip)rgr%rs r5r/r/s, ; 4 4c'62233 4 44r7c(eZdZdZedZdS) ExplicitEnumzC Enum with more explicit error message for missing values. c t|d|jdt|j)Nz is not a valid z, please select one of )rGr1r~_value2member_map_r)rrs r5 _missing_zExplicitEnum._missing_ sG p pcl p p4PSPfPkPkPmPmKnKn p p   r7N)r1r0rr classmethodr.rWr7r5r+r+s9  [   r7r+ceZdZdZdZdZdZdS)PaddingStrategyz Possible values for the `padding` argument in [`PreTrainedTokenizerBase.__call__`]. Useful for tab-completion in an IDE. longest max_length do_not_padN)r1r0rrLONGEST MAX_LENGTH DO_NOT_PADrWr7r5r1r1's) GJJJJr7r1ceZdZdZdZdZdZdS) TensorTypez Possible values for the `return_tensors` argument in [`PreTrainedTokenizerBase.__call__`]. Useful for tab-completion in an IDE. rLrMrNN)r1r0rrPYTORCHNUMPYMLXrWr7r5r9r92s) G E CCCr7r9c&eZdZdZddZdZdZdS) ContextManagersz Wrapper for `contextlib.ExitStack` which enters a collection of context managers. Adaptation of `ContextManagers` in the `fastcore` library. context_managerslist[AbstractContextManager]c:||_t|_dSr)r?r stack)rr?s r5rzContextManagers.__init__Cs 0[[ r7cN|jD]}|j|dSr)r?rB enter_context)rcontext_managers r5 __enter__zContextManagers.__enter__Gs7#4 6 6O J $ $_ 5 5 5 5 6 6r7c*|jj|i|dSr)rB__exit__rs r5rHzContextManagers.__exit__Ks" T,V,,,,,r7N)r?r@)r1r0rrrrFrHrWr7r5r>r>=sP !!!!666-----r7r>ctj|j}|jD]}|dkr|j|jdurdS dS)zr Check if a given model can return loss. Args: model_class (`type`): The class of the model. return_lossTF)inspect signatureforward parametersr) model_classrLps r5can_return_lossrQOsV!+"566I  !  )"6q"9"AT"I"I44 5r7c|j}tj|j}d|vrd|jDSd|jDS)zq Find the labels used by a given model. Args: model_class (`type`): The class of the model. QuestionAnsweringc"g|] }d|vs|dv | S)label)start_positions end_positionsrWrXrPs r5rZzfind_labels..js+mmma7a<<1HlClClClClClr7cg|]}d|v| S)rUrWrXs r5rZzfind_labels..ls@@@a7a<<<<dd}t||||S)z/Flatten a nested dict into a single level dict.r\r#c3K|D]s\}}|r"t||zt|zn|}|r@t|tr+t |||Ed{Vm||fVtdS)N)r_)rrOrlr flatten_dict)r]r^r_rrrs r5 _flatten_dictz#flatten_dict.._flatten_dictrsGGII  DAq:DK#j//I-A66!C Z>22 '3)DDDJJLLLLLLLLLL1f   r7r\r#)r)r]r^r_rcs r5rbrbos4  aY77 8 88r7ct|rtj||St|r||jn |j|St dt|d)z< Framework-agnostic version of transpose operation. )axesNz"Type not supported for transpose: r#)r_rM transposer^rpermuterGrP)rrfs r5rgrg}steN|E----   N,uwwMEM4,@@Ld5kkLLLMMMr7ct|rtj||St|r |j|St dt |d)z: Framework-agnostic version of reshape operation. z Type not supported for reshape: r#)r_rMreshaper^rGrP)rnewshapes r5rjrjsgeLz%***   Lu}h''JDKKJJJKKKr7ct|rtj||St|r,||n||St dt |d)z: Framework-agnostic version of squeeze operation. )axisNdimz Type not supported for squeeze: r#)r_rMsqueezer^rGrPrrms r5rprpseLz%d++++   L"&,u}}EMMdM4K4KKJDKKJJJKKKr7ct|rtj||St|r||St dt |d)z> Framework-agnostic version of expand_dims operation. rnz$Type not supported for expand_dims: r#)r_rM expand_dimsr^ unsqueezerGrPrqs r5rsrssleP~eT***   P4(((NU NNNOOOr7ct|rtj|St|r|St dt |d)z7 Framework-agnostic version of size operation. z$Type not supported for tensor_size: r#)r_rMsizer^numelrGrP)rs r5 tensor_sizerxsbePwu~~   P{{}}NU NNNOOOr7ctst|Sddl}|jr/t ||jr||jnt|S)zk Casts an input to a torch int64 tensor if we are in a tracing context, otherwise to a Python int. rN) r&r8r'jit is_tracingrlrptoint64rqs r5 torch_intr~sg 1vv LLL % 4 4 6 6 b:a;V;V b144   \_`a\b\bbr7ctst|Sddl}|jr/t ||jr||jnt|S)zo Casts an input to a torch float32 tensor if we are in a tracing context, otherwise to a Python float. rN) r&r8r'rzr{rlrpr|float32rqs r5 torch_floatrsg 1vv LLL"')"6"6"8"8 dZ5<=X=X d144   ^abc^d^ddr7extra list | Nonec8|pg}t|fd}|S)aI Decorator to filter out named arguments that are not in the function signature. This decorator ensures that only the keyword arguments that match the function's signature, or are specified in the `extra` list, are passed to the function. Any additional keyword arguments are filtered out and a warning is issued. Parameters: extra (`Optional[list]`, *optional*): A list of extra keyword argument names that are allowed even if they are not in the function's signature. Returns: Callable: A decorator that wraps the function and filters out invalid keyword arguments. Example usage: ```python @filter_out_non_signature_kwargs(extra=["allowed_extra_arg"]) def my_function(arg1, arg2, **kwargs): print(arg1, arg2, kwargs) my_function(arg1=1, arg2=2, allowed_extra_arg=3, invalid_arg=4) # This will print: 1 2 {"allowed_extra_arg": 3} # And issue a warning: "The following named arguments are not valid for `my_function` and were ignored: 'invalid_arg'" ``` ctj}t|j}|d|vd|vd_tfd}|S)NrrTc`i}i}|D]\}}| vr|||<|||<|rwd|D}d|} r|djjdz}n r|djdz}nd}t jd|jd|t d |i|S) Ncg|]}d|d S)'rW)rXrs r5rZzWfilter_out_non_signature_kwargs..decorator..wrapper..s 'I'I'IQA'I'I'Ir7z, rr#r\z1The following named arguments are not valid for `z` and were ignored: r) stacklevel)rjoinrr1warningswarn UserWarning) rr valid_kwargsinvalid_kwargsrrinvalid_kwargs_names cls_prefixfuncis_class_methodis_instance_methodvalid_kwargs_to_passs r5wrapperzCfilter_out_non_signature_kwargs..decorator..wrappers LN  * *1,,,&'LOO()N1%% 'I'I.'I'I'I$'+yy1E'F'F$&$!%a!2!;c!AJJ$$!%a!1C!7JJ!#J A ATXTaAA*>AA 4... .r7)rKrLsetrNrunion _filter_out_non_signature_kwargsr)rsigfunction_named_argsrrrrextra_params_to_passs` @@@r5 decoratorz2filter_out_non_signature_kwargs..decorators%%!#."5"5"7"7882889MNN$'::#6615- t / / / / / / /  />r7)r)rrrs @r5filter_out_non_signature_kwargsrs<6 KREu::,,,,,\ r7ceZdZUdZded<ded<ded<ded<ded <ded <d ed <d ed <ded<ded<ded<dS)TransformersKwargsa Keyword arguments to be passed to the forward pass of a `PreTrainedModel`. Attributes: num_items_in_batch (`Optional[torch.Tensor]`, *optional*): Number of items in the batch. It is recommended to pass it when you are doing gradient accumulation. output_hidden_states (`Optional[bool]`, *optional*): Most of the models support outputting all hidden states computed during the forward pass. output_attentions (`Optional[bool]`, *optional*): Turn this on to return the intermediary attention scores. output_router_logits (`Optional[bool]`, *optional*): For MoE models, this allows returning the router logits to compute the loss. cu_seq_lens_q (`torch.LongTensor`, *optional*) Gets cumulative sequence length for query state. cu_seq_lens_k (`torch.LongTensor`, *optional*) Gets cumulative sequence length for key state. max_length_q (`int`, *optional*): Maximum sequence length for query state. max_length_k (`int`, *optional*): Maximum sequence length for key state. position_ids (`torch.LongTensor`, *optional*) Indices of positions of each input sequence tokens. is_causal (`bool`, *optional*) Can be set to False to enable bi-directional attention, i.e. use decoder Attention modules as encoders. seq_idx (`torch.IntTensor`, *optional*): Sequence index for each token in a flattened packed batch. ztorch.Tensor | Nonenum_items_in_batchroutput_hidden_statesoutput_attentionsoutput_router_logitsztorch.LongTensor | None cu_seq_lens_q cu_seq_lens_kz int | None max_length_q max_length_k position_ids is_causalztorch.IntTensor | Noneseq_idxN)r1r0rr__annotations__rWr7r5rr s8,+++%%%%""""%%%%********))))######r7r)totalrdict[str, Any]c d|vS)z3Checks whether a config dict is a timm config dict.pretrained_cfgrW)rs r5is_timm_config_dictrJs { **r7pretrained_model_pathc|dSt|}tj|}tj|}|r_|drJt |5}tj|}dddn #1swxYwYt|S|rtj tj |drht tj |d5}tj|}dddn #1swxYwYt|SdS)zA Checks whether a checkpoint is a timm model checkpoint. NFz.jsonz config.json) rOospathisfileisdirendswithrrloadrexistsr)ris_fileis_dirrArs r5is_timm_local_checkpointrOs$u 566gnn233G W]]0 1 1F0(11'::0 ' ( ( 'A)A,,K ' ' ' ' ' ' ' ' ' ' ' ' ' ' '";///0"'...C]!S!STT0 "',,4mDD E E ')A,,K ' ' ' ' ' ' ' ' ' ' ' ' ' ' '";/// 5s$8BB BEEEmodule nn.Modulerrrcxt||||D]}t|||dS)z5 Set a value to a module and all submodules. N)rchildrenset_attribute_for_modules)rrr submodules r5rrksN FC__&&99 !)S%888899r7ct||rt|||D]}t||dS)z: Delete a value from a module and all submodules. N)rwdelattrrdel_attribute_from_modules)rrrs r5rrts\ vs__&&33 "9c222233r7c<tfd}|S)z Decorator to wrap model method, to call output.to_tuple() if return_dict=False passed as a kwarg or return_dict=False is set in the config. Note: output.to_tuple() convert output to tuple skipping all `None` values. ct|dr |jjnd}|d|}||}|g|Ri|}|s)t |t s|}|S)NrT return_dict)rwrrrrlrr )rrrrreturn_dict_passedr!rs r5rz!can_return_tuple..wrappers18x1H1HRdk--d #ZZ {CC  ),Kd,T,,,V,, ':fe#<#< '__&&F r7rrrs` r5can_return_tuplers5 4[[[ Nr7)imagevideoaudiomodalitycbdtfdtDfd}|S)u Decorator for `get__features` methods that: - strips the modality prefix from incoming kwargs whose stripped name isn't an existing parameter (e.g. `image_cu_seqlens` → `cu_seqlens`, forwarded via `**kwargs`); - drops kwargs prefixed with another known modality (e.g. `video_*` passed to an image method), so an outer `forward()` can blindly forward `**kwargs` to each modality method without leaking the wrong tensors into the wrong encoder; - leaves everything else untouched (including kwargs that match a named parameter). Used so multimodal models can accept arbitrary precomputed tensors (`image_cu_seqlens`, `video_position_ids`, …) without enumerating each one in every signature. NOTE: Apply this decorator **only once per modality**, on the innermost base model's `get__features` (i.e. on `Model.get_image_features`, not on the outer `ForConditionalGeneration.get_image_features` wrapper). Stacking it at multiple layers causes premature prefix-stripping: the outer layer rewrites `image_foo` → `foo` based on its own (narrower) signature, hiding kwargs that the inner method declares as named parameters. Outer wrappers should just forward `**kwargs` through. TODO: these modality-prefixed kwargs (`image_cu_seqlens`, `video_position_ids`, …) are currently power-feature-only — they have no visible declaration in any public signature, so users have to discover them from helper functions or docs. We should find a way to surface them properly (e.g. in `TransformersKwargs`, in a dedicated `MultimodalKwargs` typed dict, or returned grouped from the processor as `BatchFeature.images_data={...}`) so the supported set is discoverable in one place. _c3.K|]}|k|dVdS)rNrW)rXmrs r5rz-accepts_precomputed_kwargs..s-OOqha777OOr7cttjjt fd}|S)Nci}|D]R\}}|r|r|vr|||<M|||<S|i|Sr)rrQr) rr translatedrrexisting_paramsrother_prefixesprefixs r5rz>accepts_precomputed_kwargs..decorator..wrappersJ  & &1<<//<<''&A_,D,D9:Jq~~f5566$%JqMM4,,, ,r7)rrKrLrNr)rrrrrs` @r5rz-accepts_precomputed_kwargs..decorators_g/55@AA t - - - - - - -  -r7)r_KNOWN_MODALITIES)rrrrs` @@r5accepts_precomputed_kwargsrs[6^^^FOOOO,=OOOOON" r7c<tfd}|S)z Decorator using config field (if they exist) as default value for some args and kwargs. Precedence is always given to the args/kwargs that are explicitly passed. cgd}|D]1}d}| jjvr" jj|dz }|$t||kr|| ||}n4|| ||}nt |j|d}||dkr7t |ddr%|jr|rt dd}n$|dkrdd g}||vrtd |d |d |8t||kr%t|}|||<t|},|||<3|d t |jd d}|4t|jd } | r |jj} ||j_||d < |ddr_ddlm} | ||dd|d5 |g|Ri|} dddn #1swxYwYn  |g|Ri|} || r | |j_n%|j`n#|| r | |j_n|j`wxYw| S)N) use_cachevision_feature_layervision_feature_select_strategyvision_aspect_ratiorrgradient_checkpointingFzX`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`.rrfullz%`Unexpected select feature strategy: z. Please select from r#rdebug_ior)model_addition_debugger_context debug_io_dir model_debug prune_layers)__code__ co_varnamesindexrrrxrtraininglogger warning_oncerGr~rrwrmodel_debugging_utilsr)rrrargs_with_config_defaultsarg_name arg_index arg_valuevalid_strategiesris_causal_in_configis_causal_original_valuerr!rs r5rz+merge_with_config_defaults..wrappersW% % % ! 2 1 1HI4=444 M5;;HEEI $TY)>)>4 ?C^ O H%%1"8, #DK4@@ ${**t%=uEE*$-*\e*++v%* !AAA(16':$ (888(wIwwdtwww(SYY-B-B::D&/DO ;;DD'0F8$JJ{GDKd,S,STT  ")$+{"C"C " A+/;+@($-DK !"+F;  .zz*e,, 5SSSSSS44&**^]CCVZZP^E_E_99"T$888888F999999999999999 d4T444V44$&.,DDK)) - $&.,DDK)) ----- s1AH6, H9 H6H  H6 H  H66Irrs` r5merge_with_config_defaultsrs;  4[[FFFF[FP Nr7cTtdt|S)NzZThe `check_model_inputs` decorator is deprecated in favor of `merge_with_config_defaults`.)rrr)rs r5check_model_inputsr!s% tuuu %d + ++r7rc|S)zs Identity decorator that prevents the modular converter from propagating its decorators to specific files. rWrs r5no_inherit_decoratorr&s  Jr7cZeZdZdZiZdZdZdZdZdZ dZ e dd Z ddZ dS)GeneralInterfacez Dict-like object keeping track of a class-wide mapping, as well as a local one. Allows to have library-wide modifications through the class mapping, as well as local modifications in a single file with the local mapping. ci|_dSr_local_mappingrs r5rzGeneralInterface.__init__7s r7cH||jvr |j|S|j|Sr)r_global_mappingrrs r5r zGeneralInterface.__getitem__:s, $% % %&s+ +#C((r7c>|j||idSr)rr)rrrs r5rzGeneralInterface.__setitem__@s# ""C<00000r7c|j|=dSrrrs r5rzGeneralInterface.__delitem__Ds   $ $ $r7c<ti|j|jSr)rrrrs r5__iter__zGeneralInterface.__iter__Gs!Ct+Ct/BCDDDr7ct|j|jzSr)rrrrrs r5__len__zGeneralInterface.__len__Ks34',,..1D1I1I1K1KKLLLr7rrOrrc>|j||idSr)rr)rrrs r5registerzGeneralInterface.registerNs# ""C<00000r7r r&cDt|Sr)r~rrs r5 valid_keyszGeneralInterface.valid_keysRsDIIKK   r7N)rrOrr)r r&)r1r0rrrrr rrr r r/r rrWr7r5rr-sO!!!))) 111%%%EEEMMM111[1!!!!!!r7rg?g>@c"fd}|S)a Decorator that retries a function call with exponential backoff. Args: max_retries (`int`, *optional*, defaults to 5): Maximum number of retry attempts. initial_delay (`float`, *optional*, defaults to 1.0): Initial delay in seconds before the first retry. max_delay (`float`, *optional*, defaults to 30.0): Maximum delay in seconds between retries. jitter (`bool`, *optional*, defaults to `True`): Whether to add random jitter to the delay. exceptions (`tuple`, *optional*, defaults to `(Exception,)`): Tuple of exception types to catch and retry on. cHtfd}|S)Nc}td dzD]} |i|cS#$r}| krt| } r|tjddz}tdjd|d d|d|d d tj|t|d z }Yd}~d}~wwxYwdS) Nrg?g333333?[z ] attempt /z failed: z Retrying in z.1fzs...r) rangeminrandomuniformrinfor1timesleep) rrdelayattemptexc sleep_for exceptionsr initial_delayjitter max_delay max_retriess r5rz)retry..decorator..wrapperns;!E K!O44 6 664000000! 6 6 6+-- #E9 5 5I>!V^C%=%== KK;DM;;W;;{;;UX;;'0:;;;Jy))) 955EEEEEE 6 6 6s$B;BB66B;r)rrr!r"r#r$r%s` r5rzretry..decoratormsM t 6 6 6 6 6 6 6 6 6  6*r7rW)r%r"r$r#r!rs````` r5retryr&Vs<.2 r7)rrr r!)r r8)r rJ)r re)NTN)rrOryrrrerr)NNN)rrJrrr re)rrJr r)rrO)r!rr r"r)rgr$r%r&rr'r rrd)r]rr^rOr_rO)rr)rrr re)rrOr re)rrrrOrr)rrrrO)rrO)rrr r)hr __future__rrKrrrrrr collectionsrrcollections.abcrrr contextlibr r r dataclassesr r enumr functoolsrrtypingrrrrrrMr+r import_utilsrrrr'rr get_loggerr1rr&rrrr6rrIrTrdrjr_r^rurzrrrr`rrrrrrr.r/rOr+r1r9r>rQr[rbrgrjrprsrxr~rrrrrrrrrrrrrrrr&rWr7r5r1s#""""" -------->>>>>>>>>>EEEEEEEEEE,,,,,,,,$$$$$$$$999999999999QQQQQQQQQQLLL GCLL  H % %14666644444 5 5 5 5    999&"%%%% ' ' ' ' ' ' ' '&&&&",!%!% :### ::::ko$7$7$7$7$7N ; ; ; ;D0I3I3I3I3I3+I3I3I3X6666-155555      3    l--------$    A A A 9 9 9 9 9 N N N N L L L L L L L P P P P P P c c c e e eLLLLL^'$'$'$'$'$%'$'$'$'$T++++ 89999 3 3 3 3.0////dOOOj,,, &!&!&!&!&!~&!&!&!T | 000000r7