torch.compiler API reference

Created On: Jun 02, 2023 | Last Updated On: Jun 22, 2025

For a quick overview of torch.compiler, see torch.compiler.

compile	See torch.compile() for details on the arguments for this function.
reset	This function clears all compilation caches and restores the system to its initial state.
allow_in_graph	Tells the compiler frontend (Dynamo) to skip symbolic introspection of the function and instead directly write it to the graph when encountered.
substitute_in_graph	Register a polyfill handler for a function, usually a C function from the C extension, to be used in place of the original function when inlining the original function in the graph.
assume_constant_result	This function is used to mark a function fn as having a constant result.
list_backends	Return valid strings that can be passed to torch.compile(..., backend="name").
disable	This function provides a decorator to disable compilation on a function.
set_stance	Set the current stance of the compiler.
set_enable_guard_collectives	Enables use of collectives during guard evaluation to synchronize behavior across ranks.
cudagraph_mark_step_begin	Indicates that a new iteration of inference or training is about to begin.
is_compiling	Indicates whether a graph is executed/traced as part of torch.compile() or torch.export().
is_dynamo_compiling	Indicates whether a graph is traced via TorchDynamo.
is_exporting	Indicated whether we're under exporting.
skip_guard_on_inbuilt_nn_modules_unsafe	A common function to skip guards on the inbuilt nn modules like torch.nn.Linear.
skip_guard_on_all_nn_modules_unsafe	A common function to skip guards on all nn modules, both user defined as well inbuilt nn modules (like torch.nn.Linear).
keep_tensor_guards_unsafe	A common function to keep tensor guards on all tensors.
skip_guard_on_globals_unsafe	A common function to skip guards on all globals.
nested_compile_region	Tells ``torch.compile`` that the marked set of operations forms a nested compile region (which is often repeated in the full model) whose code can be compiled once and safely reused.