U 9%e(D @sddlZddlZddlZddlmZmZddlmZddlm m Z ddl m Z ddlmZmZmZmZmZmZmZddlmZddd d gZd d Zd dZeeeeeeddddZeddddZeeeeeeffdddZeeedfedddZ edddd Z!edddZ"dd Z#d,ej$je%eeed"d#d$Z&d-ej$jeeeeeee'e(e%ej)fe(fd%d&d'Z*edd(d Z+ed)d*d+Z,dS).N) GraphModuleNode)replace_pattern_with_filters)fuse_conv_bn_weights)AnyCallableDictOptionalTupleListUnion)LeafSpecfold_bn_weights_into_conv_nodeget_aten_graph_modulemove_exported_model_to_eval"remove_tensor_overload_for_qdq_opscCs&|dkr dS|jdkstt||jS)Nget_attr)opAssertionErrorgetattrtarget)nodemr_/var/www/html/Darija-Ai-API/env/lib/python3.8/site-packages/torch/ao/quantization/pt2e/utils.py_get_tensor_constant_from_nodesrcCsdg}t|D]R\}}|j|kr0|||jq |jsR|t|krR|||q ||jq |S)N) enumeratenameappendZ kwarg_onlylen default_value) orig_argsZ orig_kwargsZ args_schemaZall_argsiZschemarrr_get_all_argumentss r#) conv_nodeconv_weight_nodeconv_bias_nodebn_noderreturnc Cs t||}t||}|jtjjjjk }|jtjjjjkrZt|j dt ksPt |j d}|jj j }t|j |j|} t| d|} t| d|} t| d|} t| d|} |jtjjjjkrd}n$|jtjjjjkrd}n td|j| |}t||| | || | |d\}}t|j }t|dkrB|jtjjjjkrB|d|j}t|tsXt t||||dk rv|j}n4|d }|j||j|}W5QRX||d<t|||t||_ |jD]>}|jd ks|jt j!ks|j dd krq|"|qdS) NzBN node target is unexpected ) transposeZ_bias call_functionr)#rrtorchopsatenconv2ddefault convolutiontypeargsboolrZ_schema argumentsr#kwargs$_native_batch_norm_legit_no_trainingZ_native_batch_norm_legit ValueErrorrlistrr isinstancestrsetattrgraphZinserting_beforertupleZusersroperatorgetitemZreplace_all_uses_with)r$r%r&r'rZconv_wZconv_br/Zbn_args_schemaZbn_argsZbn_wZbn_bZbn_rmZbn_rvZ eps_arg_indexZbn_epsZ fused_weightZ fused_biasZ conv_argsZweight_attr_nameZbias_attr_nameZ get_bias_nodeuserrrrr'sJ      "      *)rr(cCs|jjD]}|jdks|jtjjjjkr*q|}|j d}|jdks|jtjjj jkrh|jtjjj jkrhq|}|j d}t |j dkr|j dnd}t |||||q|j|dS)Nr0rr*r+)rBnodesrrr1r2r3r<r5r8r4r6rreliminate_dead_code recompile)rnr'r$r%r&rrr_fuse_conv_bn_xs"     rK)modelr(cCshi}|jjD]V}|jdd}dtdf}|rXt|d}|ddd|df}|||j<q |S)Nnn_module_stackr.r*) rBrGmetagetr7r>valuessplitr)rLZnode_name_to_scoperJrMZ current_scopeZbtrrr_get_node_name_to_scopes   rU.)patternexample_inputsr(cKs.ddlm}||||}|j||S)zF Convert the pattern to an FX graph with decomposed aten ops. r)capture_pre_autograd_graph)Z torch._exportrXrBrHrI)rVrWr;rXZ aten_patternrrrrs  ) match_patternr(cCstjjjjtjjjtjjjjtjjjtjjjjtjjjtjjjjtjjjtjjjjtjjjtjjjjtjjjtjjjjtjjjtjjj jtjjj tjj j j tjj j i }|j jD]&}|jdkrq|j|kr||j|_qdS)z Remove .tensor overload for quantize/dequantize ops so that we can use the match_pattern that we get from torchdynamo export to match the output of convert_pt2e r0N)r1r2Zquantized_decomposedZquantize_per_tensorr5Zdequantize_per_tensorZtensorZtensor2Zquantize_per_channelZdequantize_per_channelr3clampZTensorrBrGrr)rYZ_MAPrJrrrrs0            rcCs`|j|dd}dd}tdf}t||}t||}t|||gdd|dS) a% Replace the aten training dropout pattern with a noop, intended for eval. For models with dropout torch ops (nn.Dropout, F.dropout), calling model.eval() effectively turns these dropout ops into noops. For exported models, however, this is not done automatically, since the aten dropout patterns previously generated for training remain in the graph. Here we rewrite these dropout patterns with noops to avoid incorrectly applying further dropout during eval. See https://github.com/pytorch/pytorch/issues/103681. cSstj|dddS)N?TpZtrainingFZdropoutxrrr dropout_trainsz0_replace_dropout_for_eval..dropout_traincSstj|dddS)Nr\Fr]r_rarrr dropout_evalsz/_replace_dropout_for_eval..dropout_evalr*T)Z match_filtersZignore_literalsN)rBrHrIr1Zrandnrr)rrcrdrWrYZreplacement_patternrrr_replace_dropout_for_evals    recCs2t|ttfrdSt|ttfr.ttt|SdS)NTF)r?intfloatrCr>allmap _is_literal)argrrrrjs rjF)gm merge_dupexclude_literalsc Csd}d}i}|dkrg}|jjD]}|jdkr<|}|d7}q |j|g}|jD]}t|r||kr|r||kr|||q|jdt|} || |j j dj t |d7}|r| ||<qT||qTt |}W5QRX||_q |S)aReplace the literals in the graph with placeholder nodes that's created on the fly while we traverse the graph, so that the literal arguments in the graph can be matched and replaced To use this, the pattern and replacement graph should have the exact same number of literal args and they should be used in the exact same order in the pattern and replacement graph. If the literal arguments are not used in the same order in pattern and replacement graph, please use `_replace_literals_with_existing_placeholders` instead Args: `gm`: input GraphModule that we'll transform `merge_dup`: boolean flag to indicate that if the same literal appears multiple times in the graph, whether they should correspond to the same placeholder or not `exclude_literals`: a list of literals that will not be replaced with placeholders Example: # 1. Original Graph def pattern(self, x): return x + 3 def replacement(self, x): return x - 3 example_inputs = (torch.randn(1, 3, 3, 3),) pattern_gm = get_aten_graph_module(pattern, example_inputs) replacement_gm = get_aten_graph_module(pattern, example_inptus) # 2. Before calling replace literals we'll see the following graph: def pattern(self, x): return x + 3 def replacement(self, x): return x - 3 pattern_gm = _replace_literals_with_new_placeholders(pattern_gm) replacement_gm = _replace_literals_with_new_placeholders(replacement_gm) # 3. After replacing literals with new placeholder nodes def pattern(self, x, new_ph): return x + new_ph def pattern(self, x, new_ph): return x - new_ph Nr placeholderr*rk) rBrGrZinserting_afterr8rjrror@Z_in_specZchildren_specsr rC) rlrmrnZlast_phZcntZ literal_to_phrnew_argsrkph_noderrr'_replace_literals_with_new_placeholderss44       rr)rlrnliteral_to_ph_idxc Cs|dkr g}|dkri}dd|jjD}|jjD]l}|jdkrBq2g}|jD]B}t|r||kr||kr||}||}||qL||qLt|}||_q2|S)a Replace the literals in the graph with **existing** placeholder nodes, so that the literal arguments in the graph can be matched and replaced To use this, all literal args in the graph should be unique and each of them should correspond to exactly one placeholder node # 1. Original Graph def pattern(self, x_i8, scale, zero_point, quant_min, quant_max): return torch.dequantize_per_tensor(x_i8, scale, zero_point, quant_min, quant_max) def replacement(x_i8, scale, zero_point, quant_min, quant_max): x_i8 = torch.clamp(x_i8, quant_min, quant_max) return ((x_i8.to(torch.float32) - zero_point) * scale).to(dtype=torch.float32) example_inputs = ( torch.randn(1, 3, 3, 3), 1.0, 0, -128, 127, ) pattern_gm = get_aten_graph_module(pattern, example_inputs) replacement_gm = get_aten_graph_module(pattern, example_inptus) # 2. Before calling replace literals we'll see the following graph: def pattern(self, x_i8, scale, zero_point, quant_min, quant_max): # scale/zero_point/quant_min/quant_max are burnt in since they are scalar values return torch.dequantize_per_tensor(x_i8, 1.0, 0, -128, 127) def replacement(x_i8, scale, zero_point, quant_min, quant_max): # scale/zero_point/quant_min/quant_max are burnt in since they are scalar values x_i8 = torch.clamp(x_i8, -128, 127) return ((x_i8.to(torch.float32) - 0) * 1.0).to(dtype=torch.float32) # Note that literal args appear in different order in pattern and replacement graph, so # we can't use _replace_literals_with_new_placeholders literal_to_ph_idx = {1.0: 1, 0: 2, -128: 3, 127: 4} pattern_gm = _replace_literals_with_existing_placeholders(pattern_gm, literal_to_ph_idx) replacement_gm = _replace_literals_with_existing_placeholders(replacement_gm, literal_to_ph_idx) # 3. After replacing literals with existing placeholder nodes def pattern(self, x_i8, scale, zero_point, quant_min, quant_max): # scale/zero_point/quant_min/quant_max are burnt in since they are scalar values return torch.dequantize_per_tensor(x_i8, scale, zero_point, quant_min, quant_max) def replacement(x_i8, scale, zero_point, quant_min, quant_max): # scale/zero_point/quant_min/quant_max are burnt in since they are scalar values x_i8 = torch.clamp(x_i8, quant_min, quant_max) return ((x_i8.to(torch.float32) - zero_point) * scale).to(dtype=torch.float32) NcSsg|]}|jdkr|qS)ro)r).0rrrr ~s z@_replace_literals_with_existing_placeholders..r0)rBrGrr8rjrrC) rlrnrsZphsrrprkZph_idxrqrrr,_replace_literals_with_existing_placeholders?s$9     rvcCs t||S)z Move an exported GraphModule to eval mode. This is equivalent to model.eval() but only for certain special ops like dropout. QAT users should call this before performing inference on the model. )rer[rrrrs)rLcCs@dtddd}dtddd}t|||_t|||_|S) z Disallow calling `model.train()` or `model.eval()` on the given GraphModule. This is useful for exported models, where these methods don't actually behave as expected. T)modecSs tddS)Nz%Calling train() is not supported yet.NotImplementedErrorselfrwrrr_trainsz$_disallow_eval_train.._traincSs tddS)Nz$Calling eval() is not supported yet.rxrzrrr_evalsz#_disallow_eval_train.._eval)T)T)r9types MethodTypetraineval)rLr|r}rrr_disallow_eval_trains r)FN)NN)-rDr~r1Ztorch.fxrrZtorch.fx.subgraph_rewriterrZtorch.nn.functionalnnZ functionalr`Ztorch.nn.utils.fusionrtypingrrrr r r r Ztorch.utils._pytreer __all__rr#rrKr@r7rUrrrerjZfxr9rrrgrfZdtypervrrrrrrs^  $   Q   #  V  R