DŽgNddlZddlZddlZddlZddlZddlZddlZddlZddlZddl Z ddl m Z m Z ddlm Z mZmZmZmZmZmZmZmZmZddlZddlmcmZddlmZddlmZddl m!Z!ddlm"Z"m#Z#m$Z$m%Z%erddl&m'Z'gd Z(d d l)m*Z*m+Z+m,Z,m-Z-d d l.m/Z/m0Z0m1Z1d d l2m3Z3m4Z4d dl5m6Z6m5Z5m7Z7eejpjrgeefZ: d)dddddejvjxdee dfdeee=e fdeeee=e fee ee fde>dee=dfde/fdZ? d)dddddejvjxdee dfdeee=e fdeeee=e fee ee fde>dee=dfde/fdZ@dddde/dee=ejejfdeee=e fd eee=eCfddf d!ZDddd"dee=ejejfdeee=e fd#eee=eCfde/fd$ZEdd%d&ee d'ee=ddfd(ZFy)*N)autoEnum) AnyCallableDictIteratorListOptionalTupleType TYPE_CHECKINGUnion) compatibility) PassResult) PassManager) FlattenFuncFromDumpableContextFnToDumpableContextFn UnflattenFunc)StrictMinMaxConstraint) ConstraintDimExportBackwardSignatureExportGraphSignatureExportedProgramModuleCallEntryModuleCallSignaturedimsexportexport_for_trainingloadregister_dataclasssave unflattenFlatArgsAdapterUnflattenedModule)rrrShapesCollection)rrr)rr)r%r$r&T)dynamic_shapesstrictpreserve_module_call_signaturemodargs.kwargsr*r+r,returncddlm}t|tjj st dt|dt|tjjr t d|||||||S)aB :func:`export_for_training` takes any nn.Module along with example inputs, and produces a traced graph representing only the Tensor computation of the function in an Ahead-of-Time (AOT) fashion, which can subsequently be executed with different inputs or serialized. The traced graph (1) produces normalized operators in the all ATen operator set (as well as any user-specified custom operators), (2) has eliminated all Python control flow and data structures (with certain exceptions), and (3) records the set of shape constraints needed to show that this normalization and control-flow elimination is sound for future inputs. This API is intended for PT2 quantization training use cases and will soon be the default IR of torch.export.export in the near future. **Soundness Guarantee** See :func:`export()` docstring for more details. Args: mod: We will trace the forward method of this module. args: Example positional inputs. kwargs: Optional example keyword inputs. dynamic_shapes: An optional argument where the type should either be: 1) a dict from argument names of ``f`` to their dynamic shape specifications, 2) a tuple that specifies dynamic shape specifications for each input in original order. If you are specifying dynamism on keyword args, you will need to pass them in the order that is defined in the original function signature. The dynamic shape of a tensor argument can be specified as either (1) a dict from dynamic dimension indices to :func:`Dim` types, where it is not required to include static dimension indices in this dict, but when they are, they should be mapped to None; or (2) a tuple / list of :func:`Dim` types or None, where the :func:`Dim` types correspond to dynamic dimensions, and static dimensions are denoted by None. Arguments that are dicts or tuples / lists of tensors are recursively specified by using mappings or sequences of contained specifications. strict: When enabled (default), the export function will trace the program through TorchDynamo which will ensure the soundness of the resulting graph. Otherwise, the exported program will not validate the implicit assumptions baked into the graph and may cause behavior divergence between the original model and the exported one. This is useful when users need to workaround bugs in the tracer, or simply want incrementally enable safety in their models. Note that this does not affect the resulting IR spec to be different and the model will be serialized in the same way regardless of what value is passed here. WARNING: This option is experimental and use this at your own risk. Returns: An :class:`ExportedProgram` containing the traced callable. **Acceptable input/output types** Acceptable types of inputs (for ``args`` and ``kwargs``) and outputs include: - Primitive types, i.e. ``torch.Tensor``, ``int``, ``float``, ``bool`` and ``str``. - Dataclasses, but they must be registered by calling :func:`register_dataclass` first. - (Nested) Data structures comprising of ``dict``, ``list``, ``tuple``, ``namedtuple`` and ``OrderedDict`` containing all above types. r')_export_for_training;Expected `mod` to be an instance of `torch.nn.Module`, got .Exporting a ScriptModule is not supported. Maybe try converting your ScriptModule to an ExportedProgram using `TS2EPConverter(mod, args, kwargs).convert()` instead.)r+r,) _tracer2 isinstancetorchnnModule ValueErrortypejit ScriptModule)r-r.r/r*r+r,r2s ]/home/mcse/projects/flask_80/flask-venv/lib/python3.12/site-packages/torch/export/__init__.pyr r IsJ- c588?? +I$s)TU V  #uyy--. K    'E  c ddlm}t|tjj st dt|dt|tjjr t d|||||||dS)a :func:`export` takes an arbitrary Python callable (an nn.Module, a function or a method) along with example inputs, and produces a traced graph representing only the Tensor computation of the function in an Ahead-of-Time (AOT) fashion, which can subsequently be executed with different inputs or serialized. The traced graph (1) produces normalized operators in the functional ATen operator set (as well as any user-specified custom operators), (2) has eliminated all Python control flow and data structures (with certain exceptions), and (3) records the set of shape constraints needed to show that this normalization and control-flow elimination is sound for future inputs. **Soundness Guarantee** While tracing, :func:`export()` takes note of shape-related assumptions made by the user program and the underlying PyTorch operator kernels. The output :class:`ExportedProgram` is considered valid only when these assumptions hold true. Tracing makes assumptions on the shapes (not values) of input tensors. Such assumptions must be validated at graph capture time for :func:`export` to succeed. Specifically: - Assumptions on static shapes of input tensors are automatically validated without additional effort. - Assumptions on dynamic shape of input tensors require explicit specification by using the :func:`Dim` API to construct dynamic dimensions and by associating them with example inputs through the ``dynamic_shapes`` argument. If any assumption can not be validated, a fatal error will be raised. When that happens, the error message will include suggested fixes to the specification that are needed to validate the assumptions. For example :func:`export` might suggest the following fix to the definition of a dynamic dimension ``dim0_x``, say appearing in the shape associated with input ``x``, that was previously defined as ``Dim("dim0_x")``:: dim = Dim("dim0_x", max=5) This example means the generated code requires dimension 0 of input ``x`` to be less than or equal to 5 to be valid. You can inspect the suggested fixes to dynamic dimension definitions and then copy them verbatim into your code without needing to change the ``dynamic_shapes`` argument to your :func:`export` call. Args: mod: We will trace the forward method of this module. args: Example positional inputs. kwargs: Optional example keyword inputs. dynamic_shapes: An optional argument where the type should either be: 1) a dict from argument names of ``f`` to their dynamic shape specifications, 2) a tuple that specifies dynamic shape specifications for each input in original order. If you are specifying dynamism on keyword args, you will need to pass them in the order that is defined in the original function signature. The dynamic shape of a tensor argument can be specified as either (1) a dict from dynamic dimension indices to :func:`Dim` types, where it is not required to include static dimension indices in this dict, but when they are, they should be mapped to None; or (2) a tuple / list of :func:`Dim` types or None, where the :func:`Dim` types correspond to dynamic dimensions, and static dimensions are denoted by None. Arguments that are dicts or tuples / lists of tensors are recursively specified by using mappings or sequences of contained specifications. strict: When enabled (default), the export function will trace the program through TorchDynamo which will ensure the soundness of the resulting graph. Otherwise, the exported program will not validate the implicit assumptions baked into the graph and may cause behavior divergence between the original model and the exported one. This is useful when users need to workaround bugs in the tracer, or simply want incrementally enable safety in their models. Note that this does not affect the resulting IR spec to be different and the model will be serialized in the same way regardless of what value is passed here. WARNING: This option is experimental and use this at your own risk. Returns: An :class:`ExportedProgram` containing the traced callable. **Acceptable input/output types** Acceptable types of inputs (for ``args`` and ``kwargs``) and outputs include: - Primitive types, i.e. ``torch.Tensor``, ``int``, ``float``, ``bool`` and ``str``. - Dataclasses, but they must be registered by calling :func:`register_dataclass` first. - (Nested) Data structures comprising of ``dict``, ``list``, ``tuple``, ``namedtuple`` and ``OrderedDict`` containing all above types. r')_exportr3r4r5T)r+r, pre_dispatch) r6rBr7r8r9r:r;r<r=r>)r-r.r/r*r+r,rBs r?rrs| c588?? +I$s)TU V  #uyy--. K    'E r@) extra_files opset_versionepfrDrEc Lt|ts"tdt|jdddlm}ddlm}m }|||}t|ttjfrtj|}tj|d5}t|j t"sJ|j%d|j |j%d|j&|j%d |j(|j%d |j*|j%d d j-t/t||r>|j1D]+\} } | j3d } |j%d| | -dddy#1swYyxYw)a" .. warning:: Under active development, saved files may not be usable in newer versions of PyTorch. Saves an :class:`ExportedProgram` to a file-like object. It can then be loaded using the Python API :func:`torch.export.load `. Args: ep (ExportedProgram): The exported program to save. f (Union[str, os.PathLike, io.BytesIO): A file-like object (has to implement write and flush) or a string containing a file name. extra_files (Optional[Dict[str, Any]]): Map from filename to contents which will be stored as part of f. opset_version (Optional[Dict[str, int]]): A map of opset names to the version of this opset Example:: import torch import io class MyModule(torch.nn.Module): def forward(self, x): return x + 10 ep = torch.export.export(MyModule(), (torch.randn(5),)) # Save to file torch.export.save(ep, 'exported_program.pt2') # Save to io.BytesIO buffer buffer = io.BytesIO() torch.export.save(ep, buffer) # Save with extra files extra_files = {'foo.txt': b'bar'.decode('utf-8')} torch.export.save(ep, 'exported_program.pt2', extra_files=extra_files) zBThe 'ep' parameter must be an instance of 'ExportedProgram', got 'z ' instead.rSCHEMA_VERSION) serializeSerializedArtifactw serialized_exported_program.jsonserialized_state_dict.ptserialized_constants.ptserialized_example_inputs.ptversionr4utf-8z extra_files/N)r7r TypeErrorr<__name__torch._export.serde.schemarJtorch._export.serde.serializerKrLstrosPathLikefspathzipfileZipFileexported_programbyteswritestr state_dict constantsexample_inputsjoinmapitemsencode) rFrGrDrErJrKrLartifactzipfextra_file_namecontentencoded_contents r?r#r#sdh b/ *PQUVXQYQbQbPccm n  :K#,R#?H!c2;;'( IIaL C QD(33U;;; 8(:S:ST 0(2E2EF /1C1CD 4h6M6MN i#c>*B!CD ,7,=,=,? Q(").."9 _,=>P QQQQs C`. Args: ep (ExportedProgram): The exported program to save. f (Union[str, os.PathLike, io.BytesIO): A file-like object (has to implement write and flush) or a string containing a file name. extra_files (Optional[Dict[str, Any]]): The extra filenames given in this map would be loaded and their content would be stored in the provided map. expected_opset_version (Optional[Dict[str, int]]): A map of opset names to expected opset versions Returns: An :class:`ExportedProgram` object Example:: import torch import io # Load ExportedProgram from file ep = torch.export.load('exported_program.pt2') # Load ExportedProgram from io.BytesIO object with open('exported_program.pt2', 'rb') as f: buffer = io.BytesIO(f.read()) buffer.seek(0) ep = torch.export.load(buffer) # Load with extra files. extra_files = {'foo.txt': ''} # values will be replaced with data ep = torch.export.load('exported_program.pt2', extra_files=extra_files) print(extra_files['foo.txt']) print(ep(torch.randn(5))) rrRr4rrIzSerialized version z+ does not match our current schema version ) deserializerLNrNzserialized_state_dict.jsonz"This version of file is deprecatedzserialized_constants.jsonrOrPrQrD/r'rS)r7rXrYrZr[r\r]readdecodesplitrVrJlen RuntimeErrorrWrprLinfolistfilenamewarningswarn startswith)rGrDrmrirRrJrprLserialized_exported_programserialized_state_dictserialized_constantsserialized_example_inputs file_info file_contentrxrhrFs r?r!r!ksPf!c2;;'( IIaL#K C 8D))I&--/55c:=7|s>2222 1:^A./ /%gY/""0!14  R8<#150459! EI99Y%7%78L!!%GG.:+##'CC BC(4%##'BB BC'3$##'AA(4%##'@@'3$##'EE,8)##..}=$--33C;A>(4(;(;G(D H%' E*+666$000#///(444'9 ' ! % ( #9 :q888sE$H7AHH%serialized_type_nameclsrc"ddlm}|||S)aV Registers a dataclass as a valid input/output type for :func:`torch.export.export`. Args: cls: the dataclass type to register serialized_type_name: The serialized name for the dataclass. This is required if you want to serialize the pytree TreeSpec containing this dataclass. Example:: @dataclass class InputDataClass: feature: torch.Tensor bias: int class OutputDataClass: res: torch.Tensor torch.export.register_dataclass(InputDataClass) torch.export.register_dataclass(OutputDataClass) def fn(o: InputDataClass) -> torch.Tensor: res = res=o.feature + o.bias return OutputDataClass(res=res) ep = torch.export.export(fn, (InputDataClass(torch.ones(2, 2), 1), )) print(ep) r)!register_dataclass_as_pytree_noder)torch._export.utilsr)rrrs r?r"r"sHF , "6 r@)N)Gbuiltinscopy dataclassesinspectiorYsystypingryr\enumrrrrrrr r r r r rr8torch.utils._pytreeutils_pytreepytreetorch.fx._compatibilityrtorch.fx.passes.infra.pass_baser"torch.fx.passes.infra.pass_managerrrrrr%torch.fx.experimental.symbolic_shapesr__all__r*rrrr(r^rrrgraph_signaturerrr$r%r&fx GraphModulePassTyper9r:rXboolr rrZBytesIOintr#r!r"r)r@r?rs      $$16:M (DCSSJDD UXX))*HZ,@@ A (,X NR68X X S/X T#s(^ $X U4S>5:tCy#HIJ X  X%*#s(OXX|(,r NR68r r S/r T#s(^ $r U4S>5:tCy#HIJ r  r%*#s(Orrr-1.2 OQOQ S"++rzz )*OQ$sCx.) OQ DcN+ OQ  OQj-17; p S"++rzz )*p$sCx.)p%T#s(^4 p  pl+/( c(#3-( (r@