Discover available C++ compiler variants (for __dso_handle tests).

Return extra compiler flags for the current platform.

Classes whose annotations become this type's own fields.

Find the CUDA install path.

Find the library directory for tvm-ffi.

Find LLVM clang/clang++ via LLVM_PREFIX or PATH. Returns (clang, clang++) paths, or None if not found.

Retry all pending classes. Called after each successful phase 2.

Format init-eligible fields as a parameter string with defaults and kw_only.

Get a ctypes handle to libc with correct mmap/munmap signatures.

Generate a unique hash for the given sources and functions.

Install ``__ffi_init__`` as a method that delegates to ``__init_handle_by_constructor__``. Custom ``__init__`` methods call ``self.__ffi_init__(*

Construct a subclass instance with auto-assigned ``_value``/``_name``.

Return True if *annotation* is ``ClassVar`` or ``ClassVar[...]``.

Build ``__init__`` that delegates to ``__ffi_init__``. Both ``@c_class`` and ``@py_class`` use the same constructor-call path (``self.__init_

Load two objects under VA pressure and return (distance, overflowed). Returns ``(distance_bytes, False)`` when both objects load successfully,

Inject and validate the user-visible ``value`` field for payload enums.

Register the object type with the FFI core.

Undo :func:`_register_type_without_fields` after a phase-2 failure. The C-level type index is permanently consumed (cannot be reclaimed), but

Normalize plain Python sequences for structural container equality.

Containment with a Python-level structural fallback for nested sequences.

Validate that argument names are valid Python identifiers and unique. Parameters ---------- names List of argument names to valid

Add one to the input integer.

! \brief Get the i-th element of the shape. */

! * \brief Immutably read i-th element from the shape tuple. * \param idx The index * \return the i-th element. */

Get the backtrace of the error # Returns The backtrace of the error

! \return begin iterator */

Benchmark a function by calling it multiple times. Parameters ---------- name : str Name of the benchmark func : callable

Build the module in the given build directory using ninja.

Build test objects for all available compiler variants. Objects are placed under *out_dir* (default: ``{tempdir}/tvm_ffi_orcjit_tests``). Al

! \brief Identify sections eligible for the overflow (separate-mmap) * path. * * Name-based candidate selection followed by edge validatio

Return path prefix for test_funcs_conflict object.

Convert a Python object into TVM FFI values. This helper mirrors the automatic argument conversion that happens when calling FFI functions. I

Helper to create a Tensor with a specific float value and shape

Return path prefix for test_ctor_dtor object.

! * \brief Get the data type of the Tensor. * \return The data type of the Tensor. */

Find header files for C compilation.

Generate method definitions for this object.

Generate the initial FFI API stub code for a given module.

Generate import statements for the types used in the stub. It processes: import-section

Generate the `__init__.py` file for the `tvm_ffi` package.

Generate a class definition for an object type. It processes: object/${type_key}

Get an attribute by name from a derive input # Arguments `derive_input` - The derive input to get the attribute from `name` - The name of the attribu

Get the tvm-rt crate name \return The tvm-rt crate name

Implement a pythonic __getitem__ helper. Parameters ---------- obj The original object elem_getter A simple function

Add module to the import list of current one. Parameters ---------- module The other module.

! \brief True iff the reservation succeeded. */

Return path prefix for test_link_order_base object.

Get list of global functions registered. Returns ------- names List of global functions names.

Create a faster get_current_cuda_stream for torch through cpp extension.

Construct an `ObjectInfo` from an object type key.

Print benchmark result in a formatted way. Parameters ---------- name : str Name of the benchmark time_per_call : float

mutation in std::vector, implements copy-on-write ! * \brief push a new item to the back of the list * \param item The item to be pushed. */

Register a Python-defined FFI class with dataclass-style semantics. Can be used as: .. code-block:: python @py_class # bare decora

Reload the code blocks from disk while preserving original `lines`.

Must run in a separate function to ensure deletion happens before mod unloads. When a module returns an object, the object deleter address is

Vector dtype roundtrips correctly.

Measures overhead of running dlpack via auto convert by directly take test wrapper as inputs. This effectively measure DLPack exchange in tvm ffi.

Measures overhead of running dlpack via auto convert by directly take test wrapper as inputs. This effectively measure DLPack exchange in tvm ffi.

Create a new dtype with the given number of lanes. Parameters ---------- lanes The number of lanes for the result

! \brief Default constructor */

[Any_AnyView.GetInt_Float.begin]