+ QV-jB^RIt^RIt^RIt^RIt^RIHtHt^RIHt^RI H t ^RI H t H t ] !4'd4^RIt^RIHtRt]P$P'4'd^RItRtMRt] P*!]4tR t]P2!R 4tR R ltR tRRRlltRRRlltRRltRt ]P2!R4t!Rt"Rt#RRlt$R RRllt%] !R!R7]R"RRll44t&R#)#N)contextmanagerredirect_stdout)StringIO)logging)is_torch_availablerequires) save_fileFTc\'d%\PP4'gR#\PP 4^8H#)z7Return True if rank=0 or we aren't running distributed.T)_torch_distributed_availabletorch distributedis_initializedget_ranks/Volumes/fast/ai/experiments/ui-tars-smoke/.venv/lib/python3.14/site-packages/transformers/model_debugging_utils.py _is_rank_zeror,s9 ( (U->->-M-M-O-O    % % '1 ,,rzobject at 0x[0-9A-Fa-f]+c0V^8dQhR\R\/#)x_strreturnstr)formats"r __annotate__r6sCC3C3Crc.\PRV4#)z Replace memory addresses in an object's repr with a stable placeholder so that beautiful JSON diffs won't be ruined by ephemeral addresses. zobject at 0xXXXXXXXX)MEMORY_ADDRESS_REGEXsub)rs&r_sanitize_repr_for_diffr6s # #$:E BBrcV\4'dR\VP4 2#R#)z@Return a stable string representation for a DTensor-like object.zDTensor (rank0) -> zDTensor(non-rank0))rrepr _local_tensor)xs&r _dtensor_reprr$>s#$T!//%:$;<< rcXV^8dQhR\R,R\R\R,/#r debug_pathNuse_repr path_to_valuerbool)rs"rrrEs...Tz.48.PSVZPZ.rc\P!RR7V'd \V4pMV'dVPR4'g VR, pV'd \P P W4MTp\RVP4P4P4/V4RV 2pM\RV: RV: R24hR \VP4R \VP4R V/pVP\P\P \P"09dVP%R \'\VP)444R \'\VP+444R\'\VP-444R\'\VP/444/4V#)aa Converts Tensors and DTensors to a JSON-serializable dictionary representation. Args: value: Any Python object, often including torch Tensors, lists, dicts, etc. debug_path (`str`, *optional*, defaults to `None`): Directory to dump debug JSON and SafeTensors files. use_repr (bool, *optional*, defaults to `True`): Whether to save a `repr()`-ized version of the tensor as the `value` property in the asscoiated FULL_TENSORS.json file, or to store the full tensors in separate SafeTensors file and store the relative path to that file in the `value` property in the dictionary. path_to_value (`str`, *optional*, defaults to `None`): The file name for the SafeTensors file holding the full tensor value if `use_repr=False`. Returns: A nested Python structure (list, dict, or sanitized string) that is safe to json.dump. T)sci_modez .safetensorsdataz./z use_repr=z and path_to_value=z cannot both be falsy.shapedtypevaluemeanstdminmax)r set_printoptions _repr_to_listendswithospathjoinr contiguousdetachcpu ValueErrorr!r/r0float16float32bfloat16updaterr2r3r4r5)r1r'r(r) value_outfilepathouts&&&& r_serialize_tensor_like_iorGEsa$ D)!%( %%n55 ^ +M>H277<< :m65++-446::<=xH( IH;&:M+;;QRSS ekk"ekk" C  {{u}}emmU^^DD /UZZ\0BC.tEIIK/@A.tEIIK/@A.tEIIK/@A    JrcXV^8dQhR\R,R\R\R,/#r&r*)rs"rrrvs,(0(0S4Z(0$(0^adh^h(0rc \V\\34'd0\V4UUu.uFwrE\ WQW# RV 2R7NK upp#\V\ 4'd6VP 4UUu/uFwreV\ WQW# RV 2R7bK upp#\VR4'd\VPWVR7#\V\P4'd\WW#R7#\\V44#uuppiuuppi)aR Recursively build a JSON-serializable Python structure from `value`. Tensors and DTensors become either sanitized repr strings, or are saved to disk as SafeTensors files and their relative paths are recorded in the returned Python structure. Lists/tuples/dicts are recursed into. All memory addresses are replaced with a stable placeholder. Args: value: Any Python object, often including torch Tensors, lists, dicts, etc. debug_path (`str`, *optional*, defaults to `None`): Directory to dump debug JSON and SafeTensors files. use_repr (bool, *optional*, defaults to `True`): Whether to save a `repr()`-ized version of the tensors as the `value` property in the asscoiated FULL_TENSORS.json file, or to store full tensors in separate SafeTensors files and store the relative path to that file in the `value` property. path_to_value (`str`, *optional*, defaults to `None`): The file name for the SafeTensors file holding the full tensor value if `use_repr=False`. Returns: A nested Python structure (list, dict, or sanitized string) that is safe to json.dump. _r'r(r)r") isinstancelisttuple enumerate _serialize_iodictitemshasattrrGr"r Tensorrr!)r1r'r(r)ivks&&&& rrPrPvs (%$''"%( ( !XWffghigjUk l(  %  % }QZiijkljmXno o%  uo&&(   JYf  %&&(PXvv "4; //'  s D5Dc8V^8dQhR\P/#)rr1)r rT)rs"rrrs555rc T\P!R^xR7\4;_uu_4p\V4;_uu_4\ V4VP 4pRRR4RRR4\ X4P4# +'giL1;i +'giL<;i)z Converts a tensor into a sanitized multi-line string representation. Args: value (`torch.Tensor`): The tensor to represent. Returns: `list[str]`: List of string lines representing the tensor. T)r- linewidthN)r r6rrprintgetvaluer splitlines)r1bufraws& rr7r7sd DC8 sOC00 e lln1 #3 ' 2 2 4410s"BB B B B B' cVPR4'd0VPRR4VR,Fp\V4K R#R#)childrenoutputsN)getpopprune_outputs_if_children)nodechilds& rreres@ xx  D!*%%E %e ,&rz (.*)\.(\d+)$cBa\PVPRR44pV'dVPR4'gR#VP^4o\;QJd&V3RlVR,4F 'gK R# R#!V3RlVR,44#)z Checks whether a node represents a layer block with submodules. Args: node (`dict`): A node from the call tree. Returns: `bool`: Whether the node is a layer block. module_pathraFc3V<"TFpRS R2VPRR49xK R#5i).rirjNrc).0rgnumbers& r !is_layer_block..s+[JZ6(!} - <E ,, [[^F 3[$zJZ[33[3[3[$zJZ[ [[rcVPR4'gR#\VR,4UUu.uFwr\V4'gKW3NK ppp\V4^8dJV^RUUu.uFwrVNK ppp\VR,4UUu.uFwrW9gK VNK uppVR&VR,Fp\ V4K R#uuppiuuppiuuppi)z Recursively removes intermediate layers from the tree to improve readability. Keeps at least the first and last layers if many consecutive layers are present. Args: node (`dict`): The root or subnode to prune recursively. raN)rcrOrvlenprune_intermediate_layers)rfrUrg layer_blocksrJ to_removes& rrzrzs 88J  /8j9I/Jd/J81n]bNcJQJ/JLd <1#/"#56#541Q#5 62;D.strip_values..cleansX#t$$&A!H&C&&D$K 'rinputsrbraNrm)rfrgr strip_valuess& @rr+log_model_debug_trace..strip_valuessM  dhhx$% dhhy"%&XXj"-E  .r)r9makedirsr:r;_debugger_module_dump_name Exceptionr?loggerinfore _call_treeopenjsondumploadsdumps) r'modelbasee full_path summary_pathf tree_copyrs && @rlog_model_debug_tracers? X KK T 277<< ,L,L},\]D//-? KK)$u56++I/)Le../ i   %""Aa0   4::e&6&678I lC A )q) ! E XA*QOPVW W X  . ! s0AE ;$F=F F+E;;F F  F' c<V^8dQhR\R\R\/#)rr'do_prune_layersr(r*)rs"rrrs-|(|(|(|( |(rcaaaaa a SPPo RS RRRRR./Sn.SnS SnS'd\ P !SRR7VVV3R lpSP4FwrgVR 8XdKV!VS R V 24K SPo \P!S 4V VVVV V3R l4pVSn R# \dp\RS R 24ThRp?ii;i) a Attaches a debugging wrapper to every module in the model. This records structured inputs and outputs during the forward pass into a call tree. Args: model (`PreTrainedModel`, `nn.Module`): Model to wrap. debug_path (`str`): Optional directory to dump debug JSON files. do_prune_layers (`bool`, *optional*, defaults to `True`): Whether to prune intermediate layers. use_repr (bool, *optional*, defaults to `True`): Whether to save a `repr()`-ized version of the tensors as the `value` property in the associated FULL_TENSORS.json file, or to store full tensors in separate SafeTensors files and store the relative path to that file in the `value` property. rirNrbraTrrrlc|<aaaSPo\P!S4VVVVVV3Rl4pVSnR#)c <\4'dlRVRV/pVUu/uF$p\W#,4^8gKW2V,bK& ppRSR\VSS S R2R7RRR./pS PP V4\ P !4;_uu_4S !V/VBpRRR4\4'd\R S P444^8dRXR&M\XSS S R 2R7XR&S PP4pVR,'gVPR4S P'd*S PR ,R,P V4X#uupi +'giL;i) argskwargsrir_inputsrKrbNrac3&"TFp^xK R#5i)Nr)rnrJs& rrpX_attach_debugger_logic..wrap_forward..wrapped_forward..Fs:"9Qq"9s_outputsrx) rryrP_debugger_model_call_stackappendr no_gradsumnamed_childrenrd) inpskws dict_inputsrWrfrFfinishedr'rrmodule orig_forwardr(s *, rwrapped_forwardE_attach_debugger_logic..wrap_forward..wrapped_forward1si%tXs; :Ea+Q[^I\_`I`0qa.0+ a!9m##-!))2 7&;  t 0077="D0C0!:&"7"7"9::Q>&*DO&3#-!))2 8&< 'DO!;;??A ++LL,33344R8DKKHUJEb!sE2 E2 E77 F N)forward functoolswraps)rrrrr'rr(sff @r wrap_forward,_attach_debugger_logic..wrap_forward.s6~~  &% %  '% N)rrjc <\4'd=RS R2R\RVRV/SS S R2R7RRR ./pS PPV4S !V/VBp\4'EdS P'Ed\VSS S R 2R7XR&S PP 4pVR,S P R&VR,S P R&VR ,S P R &\ S P P44Uu.uF:pS P V,'dKS P P VR4NK< pS'd\S P 4\SS R 7V#uupi) riz (top-level)rrrrrKrbNrar)r'r) rrPrrrdrrMkeysrzr) rrtop_noderFrrW class_namer'rrreal_top_forwardr(s *, rtop_wrapped_forward3_attach_debugger_logic..top_wrapped_forwarddsu ??*\:-T8S1)%%/L"8  4B H  , , 3 3H =,, ??u???"/%!!+ H5 #HY  77;;=H)1();E  X &*29*=E  Y '+3J+?E  Z (489I9I9N9N9P4Q m4QqY^YiYijkYlYl *U   ! !!T *4Q m)%*:*:; !Zu E  ns E>0 E>) __class____name__rrrr9rrr? named_modulesrrr) rr'rr(rrname submodulerrrs ffff @@r_attach_debugger_logicrs&))J&z8T9dT^`bcE')E$'1E$ X KK T 2+)\!..0 2: Y:,av 671 }}__%&##'#J(EM XA*QOPVW W XsC C,C''C,)backendscJV^8dQhR\R,R\R\/#)rr'Nrr(r*)rs"rrrs,8585d 8585 85rc#8"VP4UUu/uFwrEWUPbK pppVPW`&\WW#4VxVP4F wrxWnK R#uuppi TP4F wrxYnK i;i5i)a # Model addition debugger - context manager for model adders This context manager is a power user tool intended for model adders. It tracks all forward calls within a model forward and logs a slice of each input and output on a nested JSON file. If `use_repr=True` (the default), the JSON file will record a `repr()`-ized version of the tensors as a list of strings. If `use_repr=False`, the full tensors will be stored in separate SafeTensors files and the JSON file will provide a relative path to that file. To note, this context manager enforces `torch.no_grad()`. ## Usage add the context manager to a model to debug ```python import torch from PIL import Image from transformers import LlavaProcessor, LlavaForConditionalGeneration, model_addition_debugger_context torch.random.manual_seed(673) # load pretrained model and processor model_id = "llava-hf/llava-1.5-7b-hf" processor = LlavaProcessor.from_pretrained(model_id) model = LlavaForConditionalGeneration.from_pretrained(model_id) # create random image input random_image = Image.fromarray(torch.randint(0, 256, (224, 224, 3), dtype=torch.uint8).numpy()) # prompt prompt = "Describe this image." # process inputs inputs = processor(text=prompt, images=random_image, return_tensors="pt") # call forward method (not .generate!) with model_addition_debugger_context(model, debug_path="Your_debug_path", do_prune_layers=False): output = model.forward(**inputs) ``` N)rrrrR) rr'rr(rJm orig_forwardsmodule_instanceforward_methods &&&& rmodel_addition_debugger_contextrsf/4.A.A.CD.CdaQ \.CMD ==M5oH5 /rs  6<+#(  %%''''+$#(    H %-zz"=>C .b(0V5"-**_-\")((*V|(~ :8585r