+ jQRt^RIt^RIt^RIt^RIHt^RIt^RIHtH t H t H t H t H t HtHtHtR RRlltRRltRR ltR R R lltR#)z A collection of utilities for ensuring that training can always occur. Heavily influenced by the [toma](https://github.com/BlackHC/toma) library. N)Optional) is_cuda_availableis_hpu_availableis_mlu_availableis_mps_availableis_musa_availableis_neuron_availableis_npu_availableis_sdaa_availableis_xpu_availablec(V^8dQhR\RR/#)garbage_collectionreturnN)bool)formats"X/home/luca/screen-checking/.venv/lib/python3.14/site-packages/accelerate/utils/memory.py __annotate__r(s##4#D#chV'd\P!4\4'd!\PP 4R#\ 4'd!\PP 4R#\4'd!\PP 4R#\4'd!\PP 4R#\4'd!\PP 4R#\RR7'd!\PP 4R#\!4'd!\P"P 4R#\%4'dR#\'4'd!\P(P 4R#R#)z Clears the device cache by calling `torch.{backend}.empty_cache`. Can also run `gc.collect()`, but do note that this is a *considerable* slowdown and should be used sparingly. z2.0) min_versionN)gccollectr torchxpu empty_cachermlur sdaarmusar npurmpsrcudarrneuronrs&rclear_device_cacher$(s                    e , ,              " rc$V^8dQhR\/#)r r)list)rs"rrrFsrc\V\4'g \V4p\\V44FpRW&K \ RR7V#)a Releases memory from `objects` by setting them to `None` and calls `gc.collect()` and `torch.cuda.empty_cache()`. Returned objects should be reassigned to the same variables. Args: objects (`Iterable`): An iterable of objects Returns: A list of `None` objects to replace `objects` Example: ```python >>> import torch >>> from accelerate.utils import release_memory >>> a = torch.ones(1000, 1000).cuda() >>> b = torch.ones(1000, 1000).cuda() >>> a, b = release_memory(a, b) ``` NTr#) isinstancer&rangelenr$)objectsis* rrelease_memoryr-FsA, gt $ $w- 3w<  !$/ Nrc0V^8dQhR\R\/#)r exceptionr) Exceptionr)rs"rrrds drca.ROp\S\4'dU\SP4^8Xd;\;QJdV3RlV4F 'gK R# R#!V3RlV44#R#)z Checks if `exception` relates to CUDA, XPU, MPS or CPU out-of-memory, or CUDNN not supported Args: exception (`Exception`): An exception c3N<"TFqSP^,9xK R#5i)N)args).0errr/s& r +should_reduce_batch_size..tsC{)..++{s"%TF)z out of memory.zMPS backend out of memoryz(cuDNN error: CUDNN_STATUS_NOT_SUPPORTED.z*DefaultCPUAllocator: can't allocate memoryz1FATAL ERROR :: MODULE:PT_DEVMEM Allocation failed)r( RuntimeErrorr*r4any)r/ _statementssf rshould_reduce_batch_sizer<dsUK)\**s9>>/Ba/GsC{CssCsCsC{CCC rchV^8dQhR\\,R\R\\,/#)r functionstarting_batch_sizereduce_batch_size_fn)rcallableint)rs"rrrxs5IIx II#8,IrcvaaaSf\P!\VSR7#VoSfV3RloVVV3RlpV#)a A basic decorator that will try to execute `function`. If it fails from exceptions related to out-of-memory or CUDNN, the batch size is multiplied by 0.9 and passed to `function` `function` must take in a `batch_size` parameter as its first argument. Args: function (`callable`, *optional*): A function to wrap starting_batch_size (`int`, *optional*): The batch size to try and fit into memory reduce_batch_size_fn (`callable`, *optional*): A function to determine the new batch size after an out-of-memory error. If not provided, the batch size is multiplied by 0.9 on each failure. The function takes no arguments and should return the new (reduced) batch size as an `int`. Example: ```python >>> from accelerate.utils import find_executable_batch_size >>> @find_executable_batch_size(starting_batch_size=128) ... def train(batch_size, model, optimizer): ... ... >>> train(model, optimizer) ``` )r?r@c,<\SR,4oS#)g?)rB) batch_sizesrr@8find_executable_batch_size..reduce_batch_size_fnsZ#-.J rc h<\RR7\\P!S4PP 44p\ V4\ V4^,8dlRP\VR,VR,4UUu.uF wr4V RV 2NK upp4p\RSP RSP RV R24hS^8Xd \R 4hS!S.VO5/VB#uuppi \d1p\T4'd\RR7S !4oR p?K[hR p?ii;i) Tr#z, :NN=zBatch size was passed into `zS` as the first argument when called.Remove this as the decorator already does so: `(z)`z-No executable batch size found, reached zero.N)r$r&inspect signature parameterskeysr*joinzip TypeError__name__r9r0r<) r4kwargsparamsargvaluearg_strerEr>r@s *, r decorator-find_executable_batch_size..decorators)d3g''1<<AACD v;#d)a- (iiCr TXY[T\D] ^D]jcC5%!1D] ^_G.x/@/@.ABBBJBSBSATTUV]U^^`b Q"#RSS  r?r@rYrEsf&f @rr]r]xsGF  & 3!5  %J#  . r)F)NN)__doc__r[rrKtypingrrimportsrrrrrrr r r r$r-r<r]rrrcsE      #<<(IIr