iESSKrSSKrSSKrSSKrSSKJr SSKJrJr SSK J r J r SSK J r SSKJr SSKrSSKJr SSKJr SS KJr SS KJrJrJr SS KJr "S S \ 5r\R<"\"S5(+S5"SS\ 55rSr "SS\RBRD5r#"SS\RBRD5r$Sr%Sr&SSjr'SSSSSSSS.Sjjr(g) N)get_executorch_backend_config)#_convert_to_reference_decomposed_fx prepare_fx)NodeSpecQuantizationTestCase)TestCase)clone_preserve_strides_offset) FROM_NODE_KEY)_extract_node_source_debug_infobfs_trace_with_node_process) convert_pt2e prepare_pt2eprepare_qat_pt2e)torch_version_at_leastc \rSrSrSr\R RR\R RRR\R RR\R RRR\R RR\R RRR\R RR\R RRR\R RRR\R RRR\R RRR\R RRR0rSSjrSrg)PT2EQuantizationTestCase&z= Base QuantizationTestCase for PT2 with some helper methods. Nc B[RR5 UR5n [R "U 5n [ S[[U5555n[RRU UU(aUOSSS9R5n U (a [X5n O [X5n U (a [SU 5 U "U6 [U 5n U (a [SU 5 U "U6n[nUR!5VVs0sHunnUR#U5U_M nnnUc/nUVs/sHnUR#U5PM nnUR%U UUS9 U(GaUn['5n[R "U 5n[)UUUUS9nU"U6 [+UUS9n[RRUUU(aUOSS9R5n0n[,R.R!5H%unnUU;dMUUUUR#U5'M' U b5U R!5VVs0sHunnUR#U5U_M nnnUR%UUS 9 U"U6nUR1UU5 U $s snnfs snfs snnf) Nc3t# UH.nUS:Xa!S[RRS50OSv M0 g7f)rdimN)torchexportDim).0is Y/var/www/html/ai-image-ml/venv/lib/python3.13/site-packages/torchao/testing/pt2e/utils.py ;PT2EQuantizationTestCase._test_quantizer..Is6 /-.FQ   ' ( </s68T)dynamic_shapesstrictzprepared model:zquantized model)expected_node_occurrenceexpected_node_list)backend_config)r )r")r_dynamoresetevalcopydeepcopytuplerangelenrmodulerrprintrritems call_functioncheckGraphModuleNodesrrrr_MAP_TO_FX_TRACED_OPS assertEqual)selfmodelexample_inputs quantizerr"r#check_against_fx_quantfx_qconfig_mappingexport_with_dynamic_shapeis_qat is_debug_modetraining_ir_node_occurrencem_eagermr pt2_quant_outputnskvnode_occurrencen node_listqconfig_mappingr$m_copym_fxfx_quant_outputs r_test_quantizer(PT2EQuantizationTestCase._test_quantizer4s  **, MM' " 3~./   LL   -F>D  &(   .AQ*A  #Q ' > O  #Q 'n- /G/M/M/O /Otq!B  Q  "/O    %!# 2DE2DQR%%a(2D E "" I #  "0O: ?K F2#sJ=JJ)NFNFFFN)__name__ __module__ __qualname____firstlineno____doc__ropsquantized_decomposedquantize_per_tensordefaultdequantize_per_tensorquantize_per_channeldequantize_per_channeltensorr2rK__static_attributes__rMrrr&sD  &&::EII\>\>r>r>z>z &&;;UYY=[=[=p=p=x=x &&==uyy?]?]?t?t?|?| &&::AA599CaCaCuCuC|C| &&<<CCUYYEcEcEyEyFAFA  $"'$(Sr\rz2.7.0zRequires torch 2.7+cX\rSrSrSrS SjrS\\\44Sjr S\\\44Sjr Sr S r g) PT2ENumericDebuggerTestCasez Base test case class for PT2E numeric debugger tests containing common utility functions for numeric debugging functionality. returnNc(^U4Sjn[X5 g)Nc.>URS:XdURS:Xag[UR;dUR[cgUR[SRS:XagTR [URSUS35 g)N placeholderoutputz!ExportedProgram.module().unlift()Node z doesn't have from_node info)opr meta pass_nameassertIn)noder4s r!_assert_node_has_from_node_sourcemPT2ENumericDebuggerTestCase._assert_each_node_has_from_node_source.._assert_node_has_from_node_sourcesww-'477h+>DII-=1I1Q -(,6667 MM v9: r\r )r4r5rls` r&_assert_each_node_has_from_node_sourceBPT2ENumericDebuggerTestCase._assert_each_node_has_from_node_sources 2 $EMr\c.^0mU4Sjn[X5 T$)Ncl>[U5=nb%URUR4T[U5'ggN)r namegraph_idstr)rkroot_node_sourcefrom_node_source_maps r#_extract_from_node_source_from_nodebPT2ENumericDebuggerTestCase._extract_from_node_source.._extract_from_node_source_from_nodes=$CD$II V$))$--3$SY/Wr\r )r4r5rxrws @r_extract_from_node_source5PT2ENumericDebuggerTestCase._extract_from_node_sources/1  $EO##r\c.^0mU4Sjn[X5 T$)Nc>[UR;ajUR[bU[URRS55n[ U5nUT;aUTU'gTUU:XdSUSUSU35eggg)Nnn_module_stackrfz has different from_node info z.than previous node sharing the same decomp op )r rhrugetr )rkprev_decomp_opfrom_node_source&prev_decomp_op_to_from_node_source_maps r7_extract_from_node_source_with_prev_decomp_op_from_nodeڊPT2ENumericDebuggerTestCase._extract_from_node_source_with_prev_decomp_op.._extract_from_node_source_with_prev_decomp_op_from_nodes )dii .F.R!$TYY]]3D%E!F#B4#H !)OO(;>J ?~N+, v%CDTCUHHXZ ,/S)r\r )r4r5rrs @r-_extract_from_node_source_with_prev_decomp_opIPT2ENumericDebuggerTestCase._extract_from_node_source_with_prev_decomp_ops#AC. $ $  65r\cURURUR:H=(a URUR:H5 grr) assertTruersrt)r4 node_source_1 node_source_2s rassertNodeSourcesEqual2PT2ENumericDebuggerTestCase.assertNodeSourcesEquals<    -"4"4 4 A&&-*@*@@ r\rM)r`N) rNrOrPrQrRrndictruanyrzrrr[rMr\rr^r^s; N8 $$sCx. $6d3PS8n64 r\r^czSSKJs Js Js Jn SSKJn U"5nUR URXS95 U$)z^ Create a default X86InductorQuantizer configured for the given mode (QAT and dynamic quant). rN)X86InductorQuantizer)r; is_dynamic):torchao.quantization.pt2e.quantizer.x86_inductor_quantizer quantizationpt2er7x86_inductor_quantizerr set_global,get_default_x86_inductor_quantization_config)r;rxiqrr7s rget_default_x86_quantizerrsF ML%&I  88 9  r\c2^\rSrSrSrU4SjrSrSrU=r$) FP8QDQLinearzU Float8 quantize-dequantize (QDQ, a.k.a., fake quant) linear layer used for testing. c(>[TU]5 [RUl[R "X!45R UR5UlSUlSUl SUl U(a[R "U45Ul ggNg@) super__init__r float8_e4m3fnqtyperandntoweight weight_scalescalebias)r4 in_features out_featureshas_bias __class__s rrFP8QDQLinear.__init__sm (( kk<"=>AA$**M    \O4DI r\c[RRRR UR R [R"UR/5[RS9n[RRRR U[R"UR/5URS9n[RRRR U[R"UR/5[RS9n[RRRXBUR 5nU$N)rZr output_dtype)rZr float8_dtype)rrStorchao'dequantize_affine_float8_non_decomposedrVrdatarZrfloat%quantize_affine_float8_non_decomposedrrnn functionallinearr)r4inputrq_inputdq_inputouts rforwardFP8QDQLinear.forwards""JJRR;;##,, 1 123S  ))##IIQQ,, |,R  99$$LLTT,, |,U  hh!!((499E r\)rrrrr rNrOrPrQrRrrr[ __classcell__rs@rrrs5r\rc@^\rSrSrSrSU4SjjrSrSrU=r$) FP8QDQConv2dizU Float8 quantize-dequantize (QDQ, a.k.a., fake quant) conv2d layer used for testing. c b>[T U]5 [RUl[R "X!U-/UQ75R UR5UlSUlSUl SUl U(a[R "U45Ul X@l XPl X`l Xplgr)rrrrrrrrrrrstridepaddingdilationgroups) r4 in_channels out_channels kernel_sizerrrrrrs rrFP8QDQConv2d.__init__s (( kk &0 ?; ? "TZZ.     \O4DI    r\c $[RRRR UR R [R"UR/5[RS9n[RRRR U[R"UR/5URS9n[RRRR U[R"UR/5[RS9n[RRRUUUR UR"UR$UR&UR(5$r)rrSrrrVrrrZrrrrrrrconv2drrrrr)r4rrrrs rrFP8QDQConv2d.forward5s!""JJRR;;##,, 1 123S  ))##IIQQ,, |,R  99$$LLTT,, |,U  xx""))   II KK LL MM KK  r\) rrrrrrrrr)rrrTrrs@rrrs( 4  r\rcbURn[RRRR U[R "U/5[R5n[RRRR U[R "U/5U5nU$)zM Quantize-dequantize (QDQ, a.k.a., fake quant) pattern for FP8 quantization. ) dtyperrSrrrVrZrr)rrrrrs rqdq_fp8rQs KKEiiEEMM  eW G yy  HHPP eW H Or\c SnU"U5nUR5GH6up4URRnUS;aM$URn[R "U5nU[R "[R5R - nXl[R"Xh- [R "[R5R[R "[R5R 5R[R5n XRl US;aX[URURS5n UR U lUR#5U lXRl OUS;a[%UR&UR(UR*UR,UR.UR0UR2S5n UR U lUR#5U lXRl X$R4n X$R6n[9XW 5 GM9 g)z\ In-place convert a model to fake-quantized FP8 model by replacing Linear and Conv2d layers cR^^^SSKJn U"SSS/5m0mUUU4SjmT"U5 T$)Nr) namedtuple ModInstInforsparentcZ>UR5HupT"XS9TU'T"U5 M g)N)rsr)named_children)rrsmodcreate_mod_info_recursion mod_inst_infoparent_child_mod_dicts rrLfp8_convert_..generate_model_info..create_mod_info_recursionns0#224 -:-T%c*)#.5r\) collectionsr)r5rrrrs @@@rgenerate_model_info)fp8_convert_..generate_model_infohs2*"=682DE " / "%($$r\)LinearConv2d)rF)rN) named_modulesrrNrrmaxfinforrclampminrrrrrritemrrrrrrrrrrssetattr) r5rrrsr mod_type_strparamxmaxrq_param patched_modrs r fp8_convert_rcs %06((* }}-- 3 3  yyekk%*=*=>BBB '++  ! KK++ , 0 0 KK++ , 0 0  "U !  "  : %&s8H8H%PK"xxK '3'8'8':K $&-   # Z '&       K #xxK '3'8'8':K $&-   #&+22$)..k*I+r\Fc,U(a[R"5O[R"5nU U(a(U(aeU(ae[ U5 UsSSS5 $[R R XSS9R 5nU(aUO [X#5nU(a [Xt5O [Xt5nU"U6 [RRRU5 [U5n U sSSS5 $!,(df  g=f)z Generate a reference-quantized model with quant/dequant (fake quant) pattern For INT8: using the PT2E quantization flow. For FP8: using `fp8_convert_`. NT)r!) contextlib nullcontextrno_gradrrr-rrrrrrmove_exported_model_to_evalr) rinputsr;rr7is_fp8 maybe_no_grad export_model prepare_model convert_models r_generate_ref_quantized_modelrs17J**,EMMOM  ! !> :   !<<..s4.HOOQL.vB !9!,:  6 "  % % A A- P(7M - s&D$BD DT)aotiatolrtol fullgraphinductor_configsr c U=(d 0n[RR5 [R"S5 UV s/sHn [ U 5PM n n Un Un U "U 0U D6n [R R R5 U(a[R"5 [RRRSS9 [RRUUUSSS9n[R RXS9n[R RU5nSSS5 SSS5 O2[R"5 [R"XS9nSSS5 W"U6n[RR!U UXES9 gs sn f!,(df  Ns=f!,(df  NI=f!,(df  NZ=f) z Utility function to check the results of torch compiled model by comparing the outputs of the original model and the compiled model. rT)use_new_tracer_experimental)r r!+prefer_deferred_runtime_asserts_over_guards)rN)r)rr)rr%r& manual_seedr _inductormetricsr_exportconfigpatchraoti_compile_and_packageaoti_load_packagecompiletesting assert_close)r5r6kwargsrrrrrr x ref_inputs ref_kwargs ref_modelcorrectep package_pathcompiled_modelactuals rcheck_torch_compiled_modelrsh \rF MM a>|LN I O ]]_"]]5FN ^ ,F MMwTE7L I H O _s7F+(F/AF1F/G F, (F// F= G)FFNFrr))rr(unittestr$torch.ao.quantization.backend_configr!torch.ao.quantization.quantize_fxrr+torch.testing._internal.common_quantizationrr$torch.testing._internal.common_utilsr&torch.testing._internal.inductor_utilsr rtorchao.quantization.pt2er +torchao.quantization.pt2e._numeric_debuggerr %torchao.quantization.pt2e.graph_utilsr 'torchao.quantization.pt2e.quantize_pt2errr torchao.utilsrrskipIfr^rrModulerrrrrrrMr\rr*s  :P3WM 1a3aH +G446KLO (O MO d("588??"J8 588??8 v$8+|   $!T /F  /Fr\