o h8P @sxddlZddlZddlZddlmZddlmmmmZ ddl mmmZ ej jZ ddlmZmZmZmZmZmZmZddlmZddlmZddlmZmZddlmZddlm Z dd l!m"Z"m#Z#Gd d d ej$Z%d ed ededee&ee"fdee%e%ff ddZ'd ed edee&ee"fdeeeej(e)feej(e*fffddZ+d ed edefddZ,d ed ede*fddZ-d edee*fddZ.d ed ede&fddZ/de#de&de#fdd Z0de#ddfd!d"Z1d#d$Z2e2d%ej(d&ej(dej(fd'd(Z3e2d%ej(d&ej(dej(fd)d*Z4e2d%ej(d&ej(dej(fd+d,Z5d ede6fd-d.Z7d ed ed/e*defd0d1Z8dS)2N)TupleCallableDictSetListOptionalUnion) GraphModule)Node) ObserverBaseFakeQuantizeBase)getattr_from_fqn)_is_activation_post_process)NSNodeTargetType NSResultsTypec@s4eZdZeZeZeZeZeZ dS)NodeInputOrOutputTypeN) __name__ __module__ __qualname__enumautoFP32INT8FP16UNKNOWN FP32_OR_INT8rrJ/var/www/html/ai/venv/lib/python3.10/site-packages/torch/ao/ns/fx/utils.pyrs  rnodegm logger_clsnode_type_to_io_type_mapreturncs|d}|d}|d}|d}|d}|d} |d} |d} |jd krk|j|vr0tjtjfS|j|vr;tjtjfS|j|vrFtjtjfS|j|vret||d } t| tsXJt | |||\} }||fStj tj fS|jd kr|jd kswJt|jt sJt ||jt fd d | D}t|ttfs|rt||d } t| tsJt | |||\} }||fSt fdd |D}t fdd | D}|rtjtjfS|rtjtjfStj tj fS|jdkrc|jdkrt||d }t|tsJt ||||\} }|tjfS|jdkrO  z7get_node_first_input_and_output_type..c3r.r/r0r2r5rrr7ar8c3r.r/r0r2r5rrr7dr8 call_method dequantizetorz handling needs to be added)optargetrrrrget_normalized_nth_inputr1r $get_node_first_input_and_output_typerstrr anyr r torchfloat16)rr r!r"FUNS_IO_TYPE_FP32FUNS_IO_TYPE_FP16FUNS_IO_TYPE_INT8FUNS_IO_TYPE_FP32_OR_INT8MODS_IO_TYPE_FP32MODS_IO_TYPE_INT8MODS_IO_TYPE_FP32_OR_INT8METHS_IO_TYPE_FP32_OR_INT8 first_arg_prev_node_input_typeprev_node_output_type"is_known_fp32_or_int8_input_moduleis_known_fp32_input_moduleis_known_int8_input_module prev_nodecur_node_dtype_targetrr5rr?&s                               r?cs@t||d}t|ts dS|d}dd}|jdkr=|jtjkr'|||ddS|jtjtj tj tj fvr;|||dd SdS|jd krt|jt sJJt ||jttjtjtjtjtjtjtjtjtjtjtjtjtjtjtjtjtj tj!tj"tj#tj$tjtj%tj&frj'j(fSt)fd d |D}|rt*|||SdS) z{ Returns the qparams (scale, zero_point) of the first input to `node`, if they can be inferred from the graph. rNr*cSslt|||}t|||}t|trt|jtsJt|tr$t|jts&Jt||j}t||j}||fSr/)r>r1r r=r@r )rr scale_arg_idx zp_arg_idx scale_nodezp_node scale_objzp_objrrr _get_scale_zp_from_function_argss    z@get_node_input_qparams.._get_scale_zp_from_function_argsr,rr-c3r.r/r0r2 module_objrrr7r8z)get_node_input_qparams..)+r>r1r r<r=rBquantize_per_tensortoqaddadd_relumulmul_relur@r nnqLinearConv1dConv2dnniq ConvReLU2dConv3d BatchNorm2d BatchNorm3dConvTranspose1dConvTranspose2dELU GroupNormInstanceNorm1dInstanceNorm2dInstanceNorm3d LayerNorm Hardswish LeakyReLUReLU6BNReLU2dBNReLU3d ConvReLU1d ConvReLU3d LinearReLUscale zero_pointrAget_node_input_qparams)rr r"rRrJrZrOrr]rrsb       rcCs|jdkrQt||j}t|rQt|jdksJt|jdts"J|jd}t|jts/Jt||j}t|rQt|jdksBJt|jdtsLJ|jd}|S)a If node is not an observer, returns it. If node is an observer, navigates up the graph and returns the first parent which is not an observer. For example, graph: (node_non_obs), node = node_non_obs : returns node_non_obs graph: (node_non_obs -> obs0), node = obs0 : returns node_non_obs graph: (node_non_obs -> obs0 -> fq0), node = fq0 : returns node_non_obs r-rr) r<r r=rlenargsr1r r@rr node_objrrrreturn_first_non_observer_nodes    rcCs*|jdkrt||j}t|tjrdSdS)aO Assumes that all non-param args occur first. Returns the number of non-param args expected for a node. For example, for F.linear(x, weight, bias) Returns 1, because x is a non-param arg and weight and bias are params. For lstm_mod(x, hid) Returns 2, because both x and hid are non-param args. r-r[r)r<r r=r1nnLSTMrrrrget_number_of_non_param_argss   rcCst|jdkr gS|jdkrA|jtjtjjjtjfvs(|jtj tjjj tj fvrAg}t dD]}t |j|t kr>| |q.|SdgS)a- Returns the indices of args of the node which we should attach loggers to, if input logging is enabled. For example, * for (x + y), returns [0, 1] * for (1 + y), returns [1] * for (x + 1), returns [0] * for (linear(x, w, b)) returns [0] * by default, returns [0] rr,r[)rrr<r=rBraops quantizedoperatorrcrangetyper append)rresultirrr get_arg_indices_of_inputs_to_log+s    rcCsRd}|jdvrt|j}|S|jdkr't|jtsJt||j}t|}|S)z Returns a string representation of the type of the function or module pointed to by this node, or '' for other node types. )r,r9r-)r<rBtypenamer=r1r@r )rr r4 target_modrrrget_target_type_strFs     rresults model_namec Cszi}|D]4\}}d}|D]}|D]\}}||kr+t|s$J|dd}qqq|dur6|||<q|||<q|S)a  Rekeys the layer name of a results dictionary to use node names from `model_name`. For example, transforms {'base_op_1_0': {'node_output': {'model_a': [{'ref_node_name': 'linear1', ...}]}}} into {'linear1': {'node_output': {'model_a': [{'ref_node_name': 'linear1', ...}]}}} Note: we cannot use these node names directly because they are not guaranteed to be consistent across models. This is why we extract the results first and rekey afterwards. Nr ref_node_name)itemsvaluesr) rr new_resultsold_layer_nameresult_type_to_resultsnew_layer_namemodel_name_to_resultscur_model_namelist_of_resultsrrr'rekey_logger_info_on_node_name_of_modelUs    rc Csd}|D]$}|D]}|D]\}}t|dkr(|dddur(|}nq|rb|D]2}|D])}||}|D]\}}||krJqAtt|D]}||d}|||d<qPqAq7q1dSdS)ay If `fqn` entries are filled in for one of the models in `results`, copies them over to any models which do not have them filled out. A common use case benefitting from this is comparing a model prepared by quantization to a quantized model. In this case, the model prepared by quantization would have `fqn` entries, and the quantized model would not. Nrfqn)rrrr) rmodel_name_with_fqnsrrr model_resultsref_model_resultsrrrrrmaybe_add_missing_fqns|s4      rcsfddS)Nc s|^}}}t|trt|tst|tr8t|tr8g}t||D]\}}||g|R}||i|q |St|tjrRt|tjrR|jrK|}|jrR|}|j tj ks^|j tj kr`dS||g|R}|i|Sr/) r1tuplelistziprrBTensor is_quantizedr:dtypefloat) rkwargsa0a1a_otherrel0el1new_argsfinnerrrrs( zGmaybe_dequantize_first_two_tensor_args_and_handle_tuples..innerr)rrrr8maybe_dequantize_first_two_tensor_args_and_handle_tuplessrxycCs*t|}t||}dt||S)z Computes the SQNR between `x` and `y`. Args: x: Tensor or tuple of tensors y: Tensor or tuple of tensors Return: float or tuple of floats )rBnormlog10)rrPsPnrrr compute_sqnrs rcCs"t||d|dS)z Computes the normalized L2 error between `x` and `y`. Args: x: Tensor or tuple of tensors y: Tensor or tuple of tensors Return: float or tuple of floats r[)rBsqrtsumrrrrrcompute_normalized_l2_errors" rcCs(|dd}|dd}tjj||S)z Computes the cosine similarity between `x` and `y`. Args: x: Tensor or tuple of tensors y: Tensor or tuple of tensors Return: float or tuple of floats r)reshaperBr functionalcosine_similarityrrrrcompute_cosine_similaritys  rcCs4|jdkr|jtjtjtjtjtjtjfvrdSdS)Nr,FT)r<r=rBrarcrcatstack)rrrrop_type_supports_shadowings "ridxcCs"z[|j|dd}|dur0|\}}t|t||ksJ|t|kr'||WSt||WSt|jt|j|ks>J|t|jkrK|j|WS|t|j}t|j|WStyt|jt|j|kspJ|t|jkr~|j|YS|t|j}t|j|YSw)zu Given a node, gets the n'th input to that node, normalizing args and kwargs to the best of its ability. T)normalize_to_only_use_kwargsN)normalized_argumentsrrrrr RuntimeError)rr rnorm_args_and_kwargs norm_args norm_kwargs kwargs_idxrrrr>s,    r>)9rrrBtorch.nnrtorch.ao.nn.intrinsic.quantizedao intrinsicrritorch.ao.nn.quantizedrerr`typingrrrrrrrtorch.fxr torch.fx.graphr torch.ao.quantizationr r torch.ao.quantization.utilsr torch.ao.quantization.observerrns_typesrrEnumrr@r?rrintrrrrrrrrrrrboolrr>rrrrs $       |" R   '"