gLddlZddlmZddlmZmZmZmZmZm Z ddl m Z ddl m Z mZmZmZmZmZmZmZddedefd Zd Zd Zd Z ddee eeefdeeedeeedeeeefdeedeef dZdS)N) MethodType)AnyDictListOptionalTupleUnion) PartialState)calculate_maximum_sizes convert_bytescopy_tensor_to_devicesignorant_find_batch_sizeinfer_auto_device_mapis_pippy_availablepad_input_tensorssend_to_device num_processes max_memoryc| |dkrt||dS|t|\}}||dz|z t  d\}}t jt |dz d| fdt|D}t|||d }|S) zH Calculates the device map for `model` with an offset for PiPPy r F)no_split_module_classes clean_resultNr g?ci|]}|Sr).0imemorys P/var/www/html/ai-engine/env/lib/python3.11/site-packages/accelerate/inference.py z'generate_device_map..0s>>>Aa>>>)rrr)rr r splitmathceilfloatrange) modelrrr model_sizesharedvalueending device_maprs @rgenerate_device_mapr-s$UD[jopppp4U;; Fvay(M9v&& S)) v5<<((3.%%V%%>>>>})=)=>>> &  7 J r!cd}||D]}t|}|n|,|*|D]}t|}|n|SN)rvalues)argskwargsfound_batch_sizeargkwargs rfind_pippy_batch_sizer6:s   C7<< +, .6]]__  E7>> +, r!c ddlm}m m}t } fd|D}|||||} | |j|j} || |} | S)aB Attaches the split points to the model based on `self.device_map` and generates a `PipelineStage`. Requires passing in needed `args` and `kwargs` as the model needs on the CPU. Users can pass in custom `num_chunks` as an optional hyper-parameter. By default will use `AcceleratorState.num_processes` r) ScheduleGPipe SplitPointpipelinec i|] }|j Sr) BEGINNING)r split_pointr9s rr z"build_pipeline..VsTTT +z3TTTr!)mb_args mb_kwargs split_spec)device)torch.distributed.pipeliningr8r9r:r build_stagelocal_process_indexrA) r' split_pointsr1r2 num_chunksr8r:stater@pipestagescheduler9s @rbuild_pipelinerKIsQPPPPPPPPP NNETTTT|TTTJ 8     D   U6u|  L LE}UJ//H Or!cZt}d}|jdkr ||i|}nu|jrRt||}|t d||kr"t |||}t |||}||i|n|jr |}n ||rt|}|S)Nr z-Could not find batch size from args or kwargs)r ris_local_main_processr6 ValueErrorris_last_processr)forwardrF gather_outputr1r2rGoutputr3s r pippy_forwardrScs NNE F a$)&))  $ 0v>>  #LMM M:--(/?LL*63CZPP       0'// Mr!autorFrEr example_argsexample_kwargsrFrQcZ tstdt}t|d}t|d}|j|dkrgt ||}g}t dD]B |t fd| DC||_ t|||| |j |_ |j|_ |_||_  fd} t#| |} | | _| |_ |S) a Wraps `model` for pipeline parallel inference. Args: model (`torch.nn.Module`): A model we want to split for pipeline-parallel inference split_points (`str` or `List[str]`, defaults to 'auto'): How to generate the split points and chunk the model across each GPU. 'auto' will find the best balanced split given any model. Should be a list of layer names in the model to split by otherwise. no_split_module_classes (`List[str]`): A list of class names for layers we don't want to be split. example_args (tuple of model inputs): The expected inputs for the model that uses order-based inputs for a *single process*. Recommended to use this method if possible. example_kwargs (dict of model inputs) The expected inputs for the model that uses dictionary-based inputs for a *single process*. This is a *highly* limiting structure that requires the same keys be present at *all* inference calls. Not recommended unless the prior condition is true for all cases. num_chunks (`int`, defaults to the number of available GPUs): The number of different stages the Pipeline will have. By default it will assign one chunk per GPU, but this can be tuned and played with. In general one should have num_chunks >= num_gpus. gather_output (`bool`, defaults to `False`): If `True`, the output from the last GPU (which holds the true outputs) is sent across to all GPUs. zEUsing `torch.distributed.pipelining` requires PyTorch 2.4.0 or later.cpuNrT)rr c3.K|]\}}|k |VdSr/r)rkvrs r z prepare_pippy..s+$P$P41aaQ$P$Pr!c2tjg|Ri|Sr/)rSstep)r1r2rQrFrIs rrPzprepare_pippy..forwards'UZ]TTTTTVTTTr!)r ImportErrorr rrr-r&appendnextitemshf_split_pointsrKrP_original_forward__call___original_call pippy_stager __wrapped__) r'rErrUrVrFrQrGr,rP model_forwardrrIs `` @@r prepare_pippyrj|slB   cabbb NNE!,66L#NE::N( v( Tklll  q*%% R RA   $P$P$P$P:3C3C3E3E$P$P$P P P Q Q Q Q(E 5, nj Y YE#mE >EE(EUUUUUUU w..M'GEM Lr!)r NN)rTNrNNF)r#typesrtypingrrrrrr rGr utilsr r rrrrrrintdictr-r6rKrSstrboolrjrr!rrrs ::::::::::::::::                    cae6   465;37)+/3 $$)<<5d3i01<&d3i0<5:& < T#s(^, <  <D><<<<<<r!