Ti5SrSSKJr SS/rSSKrSSKrSSKrSSKr \R"\ 5r S Sjr Sr"SS\ RR 5r"S S\ RR 5rg) z>Pass for removing duplicated initializer tensors from a graph.) annotationsDeduplicateInitializersPass!DeduplicateHashedInitializersPassNc^UR5(dUR5(a![RSUR5 gUR nUc![RSUR5 gUR U:a"[RSURU5 gg)z=Check if the initializer should be skipped for deduplication.zJSkipped deduplication of initializer '%s' as it is a graph input or outputTzqSkipped deduplication of initializer '%s' as it has no constant value. The model may contain invalid initializerszDSkipped initializer '%s' as it exceeds the size limit of %d elementsF)is_graph_inputis_graph_outputloggerwarningname const_valuesizedebug) initializer size_limit const_vals n/var/www/html/ai-image-ml/venv/lib/python3.13/site-packages/onnx_ir/passes/common/initializer_deduplication.py_should_skip_initializerrs!!##{'B'B'D'D X    ''I     ~~ " R      cURR5(a2[R"UR 55R 5$UR 5$)aStringTensor does not support tobytes. Use 'string_data' instead. However, 'string_data' yields a list of bytes which cannot be hashed, i.e., cannot be used to index into a dict. To generate keys for identifying tensors in initializer deduplication the following converts the list of bytes to an array of fixed-length strings which can be flattened into a bytes-string. This, together with the tensor shape, is sufficient for identifying tensors for deduplication, but it differs from the representation used for serializing tensors (that is string_data) by adding padding bytes so that each string occupies the same number of consecutive bytes in the flattened .tobytes representation. )dtype is_stringnparray string_datatobytes)vals r_tobytesr2sB yyxx)*2244 ;;=rc>^\rSrSrSrSSU4SjjjrSSjrSrU=r$) rDaORemove duplicated initializer tensors from the main graph and all subgraphs. This pass detects initializers with identical shape, dtype, and content, and replaces all duplicate references with a canonical one. Initializers are deduplicated within each graph. To deduplicate initializers in the model globally (across graphs), use :class:`~onnx_ir.passes.common.LiftSubgraphInitializersToMainGraphPass` to lift the initializers to the main graph first before running pass. .. versionadded:: 0.1.3 .. versionchanged:: 0.1.7 This pass now deduplicates initializers in subgraphs as well. c.>[TU]5 XlgNsuper__init__rselfr __class__s rr$$DeduplicateInitializersPass.__init__Ss $rc|SnUR5GH n0n[URR55Hn[ XPR 5(aMUR nUceUR[UR5[U54nXt;axSnXGnURU5 URceURRUR5 [RSURUR5 MXTU'M GM [R R#XS9$)NFT8Replaced initializer '%s' with existing initializer '%s'modelmodified)graphstuple initializersvaluesrrr rshaperreplace_all_uses_withr popr infoirpasses PassResult) r&r,r-graphr0rrkeyinitializer_to_keeps rcall DeduplicateInitializersPass.callWs\\^EWYL$U%7%7%>%>%@A +KII'33  ,,, ioo(>@ST&#H*6*;'556IJ&++777&&**;+;+;<KKR#((+00 )4%) B$0yy##%#CCrr)irintr,zir.Modelreturnzir.passes.PassResult __name__ __module__ __qualname____firstlineno____doc__r$r<__static_attributes__ __classcell__r's@rrrDs %%DDrc>^\rSrSrSrSSU4SjjjrSSjrSrU=r$) ruaRemove duplicated initializer tensors (using a hashed method) from the graph. This pass detects initializers with identical shape, dtype, and hashed content, and replaces all duplicate references with a canonical one. This pass should have a lower peak memory usage than :class:`DeduplicateInitializersPass` as it does not store the full tensor data in memory, but instead uses a hash of the tensor data. .. versionadded:: 0.1.7 c.>[TU]5 Xlgr!r"r%s rr$*DeduplicateHashedInitializersPass.__init__s $rcSnUR5GHn0n[URR55GHin[ XPR 5(aM UR nUce[R"5nUR5nURU5 UR5n [URR55n URX4n X;a[XKR 5[U5:wa[R!SXKU5 MSnXKn UR#U 5 UR$ceURR'UR$5 [R)SUR$U R$5 GMeXTU 'GMl GM [*R,R/XS9$)NFzVInitializer deduplication failed: hashes match but values differ with values %s and %sTr*r+)r.r/r0r1rrr hashlibsha512numpyupdate hexdigestr2rrr r r3r r4r5r6r7r8) r&r,r-r9r0rrhashed tensor_data tensor_digest tensor_dimsr:r;s rr<&DeduplicateHashedInitializersPass.calls\\^EUWL$U%7%7%>%>%@A +KII'33  ,,,!)'oo/  k* & 0 0 2 #IOO$9$9$;<   C& 1 = =>(9BUUS(-'  !#H*6*;'556IJ&++777&&**;+;+;<KKR#((+00 )4%K B$Tyy##%#CCrr>)lr?rArCrKs@rrrus %% -D-Dr)rzir.Valuerr@rBbool)rH __future__r__all__rQloggingrSronnx_irr6 getLoggerrDr rrr7 InPlacePassrrrrrcsrE" (*M N   8 $<$.D"))"7"7.Db>D (=(=>Dr