##h-ddlZddlZddlZddlZddlmZddlmZddlm Z ddl m Z e rddl Z ddl Z ddl mZddlmZdd l mZej$eZe j*j-Zd Zej2d Zd ed efdZdZdZdZdZ defdZ!e ddZ"e deddeefdZ#y)N)contextmanager)Optional)export)is_torch_available)nn)PreTrainedModel)loggingctrtjjsytjj dk(S)z7Return True if rank=0 or we aren't running distributed.Tr)_torch_distributed_availabletorch distributedis_initializedget_rank_/var/www/html/sandstorm/venv/lib/python3.12/site-packages/transformers/model_debugging_utils.py _is_rank_zeror,s3 (U->->-M-M-O    % % '1 ,,rzobject at 0x[0-9A-Fa-f]+x_strreturnc.tjd|S)z Replace memory addresses in an object's repr with a stable placeholder so that beautiful JSON diffs won't be ruined by ephemeral addresses. zobject at 0xXXXXXXXX)MEMORY_ADDRESS_REGEXsub)rs r_sanitize_repr_for_diffr6s # #$:E BBrcHtrdt|jSy)z@Return a stable string representation for a DTensor-like object.zDTensor (rank0) -> zDTensor(non-rank0))rrepr _local_tensor)xs r _dtensor_reprr>s!$T!//%:$;<< rct|ttfr|Dcgc] }t|c}St|tr-|j Dcic]\}}|t|c}}St |drRt|jjt|jjtt|dSt|tjr>t|jt|jtt|dStt|Scc}wcc}}w)a Recursively build a JSON-serializable Python structure from `value`. Tensors and DTensors become sanitized repr strings. 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. Returns: A nested Python structure (list, dict, or sanitized string) that is safe to json.dump. r)shapedtypevalue) isinstancelisttuple _serialize_iodictitemshasattrrrr!r"rr Tensor)r#vks rr'r'Es%$'*/0Q a 00%05 >1=##>>uo&%--334%--334,T%[9  %&ekk*T%++5FQhimnsitQuvv #4; //'1?s D7D<cv|jdr(|jdd|dD] }t|yy)Nchildrenoutputs)getpopprune_outputs_if_children)nodechilds rr3r3is? xx  D!*% -E %e , -rc|rF tj|dtjj||jdz}n|jdz}tjd|t|d5}t|jtj|j|d dddy#t $r}t d|d|d}~wwxYw#1swYyxYw) NF)exist_okz_debug_tree.jsonz"Unexpected or existing debug_path=z. zWriting model trace at w)indent)osmakedirspathjoin_debugger_module_dump_name Exception ValueErrorloggerinfoopenr3 _call_treejsondump) debug_pathmodel output_patheoutfiles rlog_model_debug_tracerMrs U KK U 3'',,z53S3SVh3hiK669KK  KK)+78 k3 77!%"2"23 %""GA677  UA*RPQsST T U 77s$AB=<8C= CCCC(rHc ddgd_g__fd}jD]\}}|dk(r ||d|j t j  fd}|_fd}j|ddg}|D]E} t| d} | st| tjtfs4| j|yy) N module_pathinputsr0r/crjtjfd}|_y)NcntrU||d}|Dcic]}t||dkDs|||}}t|dgd}jj |t j 5 |i|}dddtrtd jDdkDrdd<ntd<jj}|ds|jdjr!jddj |Scc}w#1swYxYw)NargskwargsrrOc3 K|]}dyw)rNr).0_s r zX_attach_debugger_logic..wrap_forward..wrapped_forward..s:Qq:s r0r/) rlenr'_debugger_model_call_stackappendr inference_modesumnamed_childrenr2) inpskws dict_inputsr-r4outfinished full_pathrImodule orig_forwards rwrapped_forwardzE_attach_debugger_logic..wrap_forward..wrapped_forwards2'+s; :EaQ[YZ^I\_`I`q+a.0a a#,+K8# "  0077=%%' 1"D0C0 1:&"7"7"9::Q>&*DO&3C&8DO ;;??A +LL,3344R8DKKHUJ1b 1 1sD&D&5 D++D4)forward functoolswraps)rhrgrjrirIs`` @r wrap_forwardz,_attach_debugger_logic..wrap_forwards0~~  &  ' :)r.c6tr1dt||ddgd}jj||i|}trʼnjrt|d<jj }|dj d<|dj d<|dj d<t j jDcgc].}j |rj j |d0c}|Scc}w)Nz (top-level)rTrOr0rQr/)rr'r]r^r2rEr%keys) rbrctop_nodererfr- class_namerIreal_top_forwards rtop_wrapped_forwardz3_attach_debugger_logic..top_wrapped_forwards ?",\:'(EF H  , , 3 3H =,, ?u??"/"4HY 77;;=H)1();E  X &*29*=E  Y '+3J+?E  Z (489I9I9N9N9P4Q mqY^YiYijkYlU   ! !!T * m  ns D3Dctr\jrPjj}|djd<|djd<|djd<trt yy)NrQr0r/)rHrI)rr]r2rErM)rYrQr0rfrHrIs r final_hookz*_attach_debugger_logic..final_hooksw ?u??77;;=H)1();E  X &*29*=E  Y '+3J+?E  Z ( ? !Zu E rlanguage_modelrI) rEr]r? named_modulesrkrlrmregister_forward_hookgetattrr$rModuler ) rIrtrHrnname submodulervrxpossible_model_calls model_callthis_model_callrus ``` @r_attach_debugger_logicrs'1Td`bcE')E$'1E$!)H!..08i 2: Y:,av 678 }}__%&'.(EMF  +-g6* !%T: z/BII;WX  1 1* =  r)r )backendscjtjjfd}|_S)a # Model addition debugger - a model adder tracer This decorator is a power user tool intended for model adders. It tracks all forward calls within a model forward and logs a slice of each input and output on a nested Json. To note, this decorator enforces `torch.inference_mode()`. ## Usage add decorator to your model class ```python from ...modeling_utils import model_addition_debugger @model_addition_debugger class MyModel(nn.Module) # Can inherit from PreTrainedModel too # ... nothing else changes ``` Then, in a separate script (example is for Llava) ```python import torch from PIL import Image import requests from transformers import LlavaProcessor, LlavaForConditionalGeneration torch.random.manual_seed(673) # load pretrained model and processor model_id = "llava-hf/llava-1.5-7b-hf" processor = LlavaProcessor.from_pretrained(model_id) model = LlavaForConditionalGeneration.from_pretrained(model_id, low_cpu_mem_usage=True) # create random image input random_image = Image.fromarray(torch.randint(0, 256, (224, 224, 3), dtype=torch.uint8).numpy()) # prompt prompt = "Describe this image." # process inputs inputs = processor(text=prompt, images=random_image, return_tensors="pt") # call forward method (not .generate!) with torch.no_grad(): output = model.forward(**inputs) ``` cJ|g|i|t|jyN)r__name__)selfrUrVcls orig_inits r wrapped_initz-model_addition_debugger..wrapped_inits#$(((tS\\2r)__init__rlrm)rrrs` @rmodel_addition_debuggerrs<\ I__S\\"3#3 CL Jrc#bKt||jj| |y#wxYww)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. To note, this context manager enforces `torch.inference_mode()`. ## Usage add the context manager to a model to debug ```python import torch from PIL import Image import requests from transformers import LlavaProcessor, LlavaForConditionalGeneration 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, low_cpu_mem_usage=True) # 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): output = model.forward(**inputs) ``` N)r __class__r)rIrHs rmodel_addition_debugger_contextrs/P5%//":":JG   s"/*/,/r)$rlrFr;re contextlibrtypingrtransformers.utils.import_utilsrutilsrr torch.distributed.tensorrmodeling_utilsr r get_loggerrrBr is_availabler rcompilerstrrrr'r3rMrrrrrrrs  %2%#/   H % %00==?-"rzz"=>C3C3C !0H- 7`#`F55p* x}* * r