U 9%e5@sddlZddlZddlZddlZddlmZmZmZmZddl Z ddl m m Z ddlmZmZddlmZddlmZddlmZddlmZe jjZe jjZe jjZeeZd d Zee d d d Z!ddZ"Gddde j#j$Z%e j j&'ddZ(e j#j)e j#j)ee j*j+ee j#j)ee fdddZ,Gddde j-Z.e j j&'ddZ/e j j&'ddZ0ddZ1dd Z2ed!d"Z3dS)#N)CallableListOptionalTuple) dynamo_timedlazy_format_graph_code) fx_graph_cse)freezing_passes)view_to_reshape)configc Cs|j}t|dsd|_|j}d|}t||s2q<|d7}q|d|_||8|d|di}|||j|j||W5QRX| ||t |||dS)N_frozen_param_countrZ _frozen_paramr get_attr) graphhasattrr inserting_beforeZ create_nodeZreplace_all_uses_withmetaupdate erase_nodeZregister_buffersetattr)gmnodeconstantgiqualnameZnew_input_noderrW/var/www/html/Darija-Ai-API/env/lib/python3.8/site-packages/torch/_inductor/freezing.pyreplace_node_with_constants         r)returnc Csdd|jjD}|dt|}g}dd|jD}tt||D]6\}\}} ||jksb||krn||qDt|| |qD| t t|t|| |S)z Replaces the parameters of a PyTorch GraphModule with constants wherever possible. Returns a list of indices representing the input parameters that were not converted to constants. cSsg|]}|jdkr|qS) placeholder)op).0rrrr :s z1replace_params_with_constants..NcSsg|]}|jdk r|jqSN)Zbase_idx)r"Zout_inforrrr#=s ) rnodeslenZ output_info enumeratezipZmutated_inp_indicesappendrextendrange recompile) rZ flat_params fw_metadataparamsZfake_inp_nodespreserved_arg_indicesZaliased_input_argsrZ real_inputrrrrreplace_params_with_constants5s r0cOsdS)NTr)argskwargsrrr return_trueNsr3cs`eZdZd eeejgefdfdd Zej j j dddZ fd d Z fd d ZZS)ConstantFolderFN)insertable_tensor_checkcs@t|i|_t|_t|_||_|dk r6|nt |_ dSr$) super__init__node_replacements collectionsCounter replaced_usesobject unknown_valueskip_constructorsr3r5)selfrr>r5 __class__rrr7Ss  zConstantFolder.__init__)rcCs|jtjjjjkrdSdS)NTF)targettorchopsZquantized_decomposedZdequantize_per_channeldefault)r?rrrr is_impuredszConstantFolder.is_impurec stjj}||\}}|jdkr,t|St||fd}|j |krN|j S|j dkrl|j|j j krl|j S|j r|j dkrtdd|Ds|j St|jtjjrtjj|jjkr|j St|}|j dkrt|tjr||s|S||r|j S||j|<t|j|jfd}|D],}t|tjjs:q"|j|d7<q"|j|gD]*} |j| t | j!kr^|j"| dq^|S)Noutputr call_functionrcss|]}t|tjVqdSr$) isinstancerCTensorr"errr sz*ConstantFolder.run_node..r )#rCrDatenZfetch_args_kwargs_from_envrBr6run_nodepytree tree_flattenr=r!Z_efficientzerotensorrEr>anyrIZ_opsZ OpOverloadTagZnondeterministic_seededtagsrJr5rFr8r1r2fxNoder;Zuser_to_last_usesgetr&userspop) r?rrNr1r2Zflattened_inputsoutZflattened_node_inpsnZ to_deleter@rrrOmsP        zConstantFolder.run_nodecs6i}|jjjD]}|jdkr|j||<qtj|dS)Nr )Z initial_env)modulerr%r!r=r6run)r?envr[r@rrr]s   zConstantFolder.run)FN)__name__ __module__ __qualname__rrrCrJboolr7rUrrVrFrOr] __classcell__rrr@rr4Rs =r4cCst|dd}||jD]\}}t|||qg}|jjD]2}|jdkr@t|j dkr@t ||j | |q@|D]}|j |qx|j|j|dS)NT)r>rr)r4r]r8itemsrrr%r!r&rXdelattrrBr)rZeliminate_dead_codelintr,)rcfrrZ erased_paramsrrr constant_folds      rh) dynamo_gmaot_autograd_gmexample_inputsrcst|tjjj}tjjj}|dk r4|dk s8tt|||}t |j }||_ | fdd|D}t ||t |tjrtt|tdtd|||fS)a5 Inlines parameters that are not mutated into constants and optimizes the graph through constant propagation and other techniques. If enabled, the function also discards the original parameters of the module for memory efficiency. Assumes that this function is run in dynamo tracing post aot_autograd. Args: dynamo_gm (torch.fx.GraphModule): The Dynamo constructed GraphModule. aot_autograd_gm (torch.fx.GraphModule): The aot_autograd constructed GraphModule to be frozen. example_inputs (List[torch.Tensor]): A list of example input tensors to be used in the freezing process. Returns: Tuple[torch.fx.GraphModule, List[int]]: A tuple containing the frozen GraphModule and a list of indices of the inputs that were preserved (not turned into constants). Ncsg|] }|qSrr)r"indrkrrr#szfreeze..z%sz FROZEN GRAPH)r rC_guardsTracingContextrWr- params_flatAssertionErrorr0rrr,r rhr Zfreezing_discard_parametersinvalidate_eager_modulesdiscard_traced_gm_paramslogdebugr)rirjrkr-rpr/Z cse_graphZaot_example_inputsrrmrfreezes(  rvcs@eZdZefddZeedddZed dd Z Z S) ErasedTensorcst||jddS)Nr)Zdevice)r6__new__to)clselemnameZ owning_modr@rrrxszErasedTensor.__new__)r|cCs||_t||_dSr$) erased_nameweakrefrefowning_mod_ref)r?r{r|modrrrr7szErasedTensor.__init__rNcCsXddt||fdD}t|dks,t|d}td|d|jd|dS)NcSsg|]}t|tr|qSr)rIrwrKrrrr#s z3ErasedTensor.__torch_dispatch__..rzTrying to Run Pytorch Eager Module After Dynamo Freezing. The original parameters have been discarded for memeory efficiency. Found in op z for erased parameter z of )rPrQr&rq RuntimeErrorr}r)rzfunctypesr1r2Zerased_tensorsrLrrr__torch_dispatch__szErasedTensor.__torch_dispatch__)rN) r_r`ra staticmethodrxrstrr7 classmethodrrcrrr@rrws rwc CstjjjjD]}t|tjj s(qt t |j dd|jddD]V\}}tjjt|||}W5QRXt|tjjr|dd|_t|||qHqdSNF)recurseT)rCrnrorWZmodule_contextZ nn_modulesvaluesrInnModulelist itertoolschainnamed_parameters named_buffers _dispatchpythonno_python_dispatcherrw Parameterrequires_grad_ _is_paramrr attr_nameZtensorZe_trrrrrs   rrc Cs|tt|jdd|jddD]V\}}tjjt |||}W5QRXt |tj j rj| dd|_t|||q dSr)rrrrrrCrrrrwrIrrrrrrrrrrs)s   rsc Cs|jj^}}|jd}|j|f|D]Z}t|jdtjr*tj |jdsTq*|jd}|j t j j ||f}|||q*W5QRX|j|dS)z Make sure the output node's layout does not change due to compiler optimizations by adding aten.as_strided nodes with the expected strides. Only used for inference so we can assume all graph outputs are model outputs. rvalN)rr%r1rrIrrCrJZ _prims_commonZis_non_overlapping_and_denserHprimsinductor_force_stride_orderrEstridereplace_input_withrfr,)r_Z output_nodeZout_listr[ftnew_noderrrenforce_output_layout8s&    rc stjjjjtjjjjtjjjjgfdd|jjD}|D]^}|j |H|j dj d}|j t jj|j d|f}||j d|W5QRXq<|j|dS)z Make sure the as_strided node's input's layout does not change due to compiler optimizations, because the as_strided strides info depends on input tensor stride info. csg|]}|jkr|qSr)rBr"r[Zas_strided_opsrrr#bs z3enforce_as_strided_input_layout..rrN)rCrDrNZ as_stridedrEZ as_strided_Zas_strided_scatterrr%rr1rrHrrrrrfr,)rZ strided_nodesr[rrrrrenforce_as_strided_input_layoutWs    rc Csdd|jjD}|D]z}|jd}t|jddks|jdjtjdrPq|j |.|j t j j |fdtji}|||W5QRXqt|t|dS) z Convert 4d convolution weight tensor to channels last format. This pass is performed before freezing so the added nodes can be constant folded by freezing. cSsg|]}|jtjjkr|qSr)rBrNZ convolutionrErrrrr#xsz9convert_conv_weights_to_channels_last..r r) memory_formatrN)rr%r1r&rsizeZ is_contiguousrCZ channels_lastrrHrNclonerErrr)rZconvsconvZ weight_noderrrr%convert_conv_weights_to_channels_lastps$ r)4r9rloggingr~typingrrrrrCZtorch.utils._pytreeutilsZ_pytreerPZtorch._dynamo.utilsrrZtorch._functorch.compile_utilsrZ+torch._inductor.fx_passes.freezing_patternsr Z#torch._inductor.fx_passes.post_gradr r rDrNr getLoggerr_rtrintr0r3rUZ Interpreterr4Z_python_dispatchZ_disable_current_modesrhZ GraphModuleZ _subclassesZ FakeTensorrvrJrwrrrsrrrrrrrsF     `    4