Vh0ddlmZddlZddlZddlZddlZddlZddlZddlZddl m Z m Z m Z ddl Z ddlZ ddlmZddlmZmZe rddlmZej,GddZej,Gd d Zej2d Zdd Z ddZddZGdd ej<ZGddej<Z Gddej<Z!y)) annotationsN)AnyCallable TYPE_CHECKING)unset_fake_temporarily) diagnosticsonnxfunction_dispatcher) fake_tensorcFeZdZUded<ded<ded<d dZed dZy) PackageInfostr package_name str | Noneversion commit_hashc |djtdd|j|j|jfS)N.pkg)joinfilterrrr)selfs M/home/dcms/DCMS/lib/python3.12/site-packages/torch/onnx/_internal/fx/_pass.pyto_onnx_domain_stringz!PackageInfo.to_onnx_domain_strings6xx 4%!2!2DLL$BRBRS T  ct|tr |j}|jdd}t |}t |dd}d}||||S)Nrr __version__) isinstancetype __module__split __import__getattr)clspython_class_namerpackagerrs rfrom_python_classzPackageInfo.from_python_class"sZ ' . 1 < < (..s3A6 \*'=$7 <+66rN)returnr )r$z type | strr'r )__name__r __qualname____annotations__r classmethodr&rrr r s-  77rr ceZdZUded<y)GraphModuleOnnxMetar package_infoN)r(rr)r*r,rrr.r..srr.c#Ktjjdfd }|tj_ dtj_y#tj_wxYww)aUContext patching `difflib.SequenceMatcher` for fx readable graph. Under this context, the `autojunk` argument of `difflib.SequenceMatcher` will always be considered as `False`. This is to prevent `difflib.SequenceMatcher` recognizing stacktrace messages in fx readable graph as junk, as these messages tend to be long (>200) and repeat multiple times, which falls under the junk filter criteria. `difflib.SequenceMatcher` is used underneath by all sorts of diffing functions in `difflib`, including `difflib.unified_diff`, `difflib.ndiff`, `difflib.context_diff`. Unfortunately, there is no way to pass `autojunk` argument to these functions, and they all default to `True`. This context patching will affect all of them. `Reference: Automatic junk heuristic `_ Nc ||||dy)NF)autojunkr,)risjunkabr2 original_inits r patched_initz:_patch_difflib_sequence_matcher_init..patched_initEsdFAq59r)Nr8T)difflibSequenceMatcher__init__)r7r6s @r$_patch_difflib_sequence_matcher_initr<3sS ++44M:(4G$9 +8(=(s6A.AA.A++A.c|jd}|jd}t5djtj||t j }dddsy|S#1swYxYw)aReturn a string containing the unified diff of two strings. This function calls a patched version of `difflib.unified_diff` with `autojunk` set to `False` for `difflib.SequenceMatcher` class. More details can be found in `_patch_difflib_sequence_matcher_init` function. Args: a: The first string. b: The second string. Returns: The unified diff of the two strings. If there is no diff, return "". Example:: >>> a = '''class GraphModule(torch.nn.Module): ... def forward(self, input_ids : torch.Tensor, attention_mask : torch.Tensor): ... # File: /modeling.py:770, code: input_ids = input_ids.view(-1, input_shape[-1]) ... view = input_ids.view(-1, 3); input_ids = None ... ''' >>> b = '''class (torch.nn.Module): ... def forward(self, input_ids: i64[1, 3], attention_mask: i64[1, 3]): ... # File: /modeling.py:770, code: input_ids = input_ids.view(-1, input_shape[-1]) ... view: i64[1, 3] = torch.ops.aten.view.default(input_ids, [-1, 3]); input_ids = None ... ''' >>> print(_unified_diff(a, b)) --- +++ @@ -1,4 +1,4 @@ -class GraphModule(torch.nn.Module): - def forward(self, input_ids : torch.Tensor, attention_mask : torch.Tensor): +class (torch.nn.Module): + def forward(self, input_ids: i64[1, 3], attention_mask: i64[1, 3]): # File: /modeling.py:770, code: input_ids = input_ids.view(-1, input_shape[-1]) - view = input_ids.view(-1, 3); input_ids = None + view: i64[1, 3] = torch.ops.aten.view.default(input_ids, [-1, 3]); input_ids = None T)keependsr8)nNz ) splitlinesr<rr9 unified_diffsysmaxsize)r4r5a_listb_listdiffs r _unified_diffrGOsrN\\4\ (F \\4\ (F - /Lwww++FFckkJKL  K LLs 6A22A; Transformc6d|jjdS)NzRunning z pass. ) __class__r()runrargskwargss r*_transform_diagnose_call_message_formatterrNs dnn--.g 66rctj}tj|5 |j  ddd|j S#t $r YdddywxYw#1swY|j SxYw)a"Return the Graph nodes in tabular format. Equivalent to stdout of `graph.print_tabular()`. If `tabulate` is not installed, return `None`. Args: graph: The Graph to print. Returns: The Graph printed in a tabular format. None if `tabulate` is not installed. N)ioStringIO contextlibredirect_stdout print_tabular ImportErrorgetvalue)graphfs rmaybe_fx_graph_tabularrYsx A  # #A &     ! ::<    ::<s(A,A A)A,(A))A,,BceZdZUdZded< ded< ded< ddZdd Z dd Zejdd Z e je jje dd Zy)rHaBase class for FX graph transformations to be used by FX-ONNX exporter. Similar to `FX Interpreter `_, specializations of this class execute the FX graph Node-by-Node. Methods in the `Transform` class can be overridden to customize the behavior of the model. This pattern can be useful for many things, including writing code transformations as well as analysis passes. The following methods can be overridden:: _run() +-- run_node() +-- placeholder() +-- get_attr() +-- call_function() +-- call_method() +-- call_module() +-- output() One important aspect to note is that if the transformation modifies the model input and/or output signature, (e.g. additional inputs/outputs are added to the model), :class:`InputAdaptStep` and/or :class:`OutputAdaptStep` are needed to reconcile :attr:`ONNXProgram.model_proto`. That is, the model signature and the model representation must match. As an additional feature, this class provides builtin support for transformation recording using the diagnostics. The granularity of overriding is up to the user. And it affects the granularity of the diagnostics information. For example, if `_run()` is overridden, the diagnostics information will only contain graph level transformation. Instead, if `call_function()` is overridden, the diagnostics information will additionally contain the node level information of `call_function()`. TODO(bowbao): Add more overridable methods in call hierarchy TODO(bowbao): Create an example once more overridable methods are added. diagnostics.DiagnosticContextdiagnostic_contexttorch.fx.GraphModulemodule!fake_tensor.FakeTensorMode | None fake_modecJ||_||_|j|_y)zInitialize the transform. Args: diagnostic_context: The diagnostic context for recording diagnostics. module: The module to be transformed. N)r\r^_detect_fake_moder`)rr\r^s rr;zTransform.__init__s##5 //1rc*|jjjDcgc]}|jj d}}t 5t jjj|cdddScc}w#1swYyxYw)z}Detect fake mode from the graph. Scan through all nodes in graph and their meta['val'] to detect fake mode. valN) r^rWnodesmetagetrtorch_dynamoutilsdetect_fake_mode)rnode fake_tensorss rrbzTransform._detect_fake_modesw :>9J9J9P9PQ e,Q Q # % F==&&77 E F FR F Fs"B)B  Bc4|Stfd|DS)Nc3xK|]1}t|tjrj|n|3ywN)rrhTensor from_tensor).0tr`s r z/Transform._maybe_fakefy_args..s3 OP 1ell(CI ! !! $ J s7:)tuple)rr`rLs ` r_maybe_fakefy_argszTransform._maybe_fakefy_argss,  K TX   rcyrpr,)rrLrMs r_runzTransform._runs=@r)diagnostic_message_formatterc |jjtjj}|j dt j}|jj|r<|jjd}t|jj}nd}d}|j|i|}|jj|r|jd}t|j} |j|d5|j!|dtj"t$||ddd|j|d 5|| |j!|d n,|j!|dtj"t$|| ddd|S|S#1swYkxYw#1swY|SxYw) a`Run the transform on `self.module`. Note that this method may or may not mutate `self.module`, and the returned `GraphModule` could be either `self.module` or a new `GraphModule`. Args: *args: Positional arguments for `self.module` to run. **kwargs: Keyword arguments for `self.module` to run. )rulezFor detailed logging of graph modifications by this pass, either set `DiagnosticOptions.verbosity_level` to `logging.DEBUG` or use the environment variable `TORCH_LOGS='onnx_diagnostics'`.F) print_outputr8z Graph diff:z ``` %s ```Nz Tabular diff:zBTabular diff is not available because `tabulate` is not installed.)r\inflight_diagnosticrrulesfx_passinfologgingDEBUGlogger isEnabledForr^print_readablerYrWry log_sectionlog LazyStringrG) rrLrM diagnosticgraph_diff_log_levelold_readable_graph old_tabularr^new_readable_graph new_tabulars rrKz Transform.runs,,@@""**A   /  '}}    ) )*> ?!%!;!;!;!O 01B1BCK"$ KD+F+    ) )*> ?!'!6!6E!6!J 0>K''($rcyrpr,)rdiagnostic_levels ranalyzezAnalysis.analyzeBsTWrN)r\r[r^r]r z.onnxfunction_dispatcher.OnnxFunctionDispatcher)rzdiagnostics.infra.Levelr'r)r(rr)r;rrrr,rrrr7s:?9?%?"P ? WWrr)r4r r5r r'r ) rKrrrHrLrrMrr'r )rWztorch.fx.Graphr'r)" __future__rrrR dataclassesr9rPrrBtypingrrrrhtorch.fxtorch._subclasses.fake_tensorrtorch.onnx._internal.fxrr torch._subclassesr dataclassr r.contextmanagerr<rGrNrYABCrHrrr,rrrs"   // @H- 777,  9960f7 7 7 7 7  7&RRjSWW Xsww Xr