Register a C++ FFI class and install structural dunder methods. Combines :func:`~tvm_ffi.register_object` with structural comparison, hashing

Return path prefix for test_call_global object.

! * \brief Get the cause chain of the error object. * \return The cause chain of the error object. */

Check header status of file without modifying it.

Remove all elements from the dict.

! \brief Release reference to all the elements */

! * \brief clear all entries */

! \brief Release reference to all the elements */

Collect all files from the given paths and parse them into FileInfo objects.

Collect all files that need header checking from git.

Collect global functions from TVM FFI's global registry.

Collect registered object type keys from TVM FFI's global registry.

! * \brief Constructor. * \param include Whether the field should participate in comparison. */

Construct a TVM FFI device with given device type and index. Parameters ---------- device_type: str or int The device type or nam

Return path prefix for test_error object.

fibonacci: recursive fibonacci */

Check if name is allowed by the current policy. Paramaters ---------- name : str Input name Returns ------- allowed

! \return find the key and returns the associated iterator */

Find libtvm_ffi. Returns ------- path The full path to the located library.

Find header files for C compilation.

Generate HTML for the documentation footer.

Create a dtype from a DLPack data type tuple. Parameters ---------- dltype_data_type The DLPack data type tuple `

Construct an `ObjectInfo` from a `TypeInfo` instance.

Generate a ``__ffi_init__`` stub when it's not already in TypeMethod. For types whose ``__ffi_init__`` is auto-generated by ``RegisterFFIInit

Generate an ``__init__`` stub from init-eligible field metadata.

Define CUDA kernels and compile them to a CUBIN file. The kernels are named `add_one` and `mul_two` and compute y[i] = x[i] + 1 and y[i] = x[i] *

Compile the empty kernel to CUBIN. Returns ------- bytes Compiled CUBIN bytes

Define a Triton kernel in-process and compile it to a CUBIN file. The kernel is named `square_kernel` and computes y[i] = x[i] * x[i]. Retur

Generate an `__all__` variable for the given names.

Retrieves an item at the given index.

Get the name of the CPU.

Like structural_equal(), but returns the AccessPath pair of the first detected mismatch. Parameters ---------- lhs The left opera

Get a function from the module by name. # Arguments `name` - The name of the function to retrieve # Returns `Result<Function>` - A `Function` instan

Get list of files tracked by git.

Get the include paths for building with torch.

Helper to extract the first float value from a Tensor

Hidden visibility: caller uses ADRP+ADD instead of GOT */

Hidden visibility: caller uses ADRP+ADD instead of GOT

! * \brief Insert an element into the given position * \param position An iterator pointing to the insertion point * \param val The element t

! * \brief True iff this slab has ever hosted a FinalizedAlloc and * currently has zero live ones. * * Used by `SlabPoolMemoryMana

Get the kind of the error # Returns The kind of the error

Load CUBIN kernel through TVM-FFI. Parameters ---------- cubin_bytes : bytes Compiled CUBIN bytes Returns ------- Mo

Load the torch c dlpack extension based on torch version.

Main entry point for the embed_cubin utility.

Build the torch c dlpack extension.

Generate the FFI library for examples.

Build, load, and run inline CPU/CUDA functions.

Compile and launch CUDA kernel through NVRTC -> TVM-FFI.

Main benchmark entry point.

Load and launch Triton kernel through TVM-FFI.

Test the lib_embedded.so library with embedded CUBIN.

Test the lib_embedded.so library with embedded CUBIN.

Test the lib_embedded.so library with embedded CUBIN.

Test the lib_dynamic.so library with dynamic CUBIN loading.

Run the quick start example.

Create an access path to the item of the current map. Parameters ---------- key The key of the item to access

Create an access path that indicate a map item is missing. Parameters ---------- key The key of the item to acces

Mark a ``@py_class`` method for FFI TypeMethod registration. Decorate any staticmethod or plain instance method on a ``@py_class`` body to ha

! * \brief Record that a new FinalizedAlloc was published from this slab. * Called by `InFlightAlloc::finalize` just before returning the

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

! * \brief Read element from map. * \param key The key * \return the corresonding element. */

Load the torch cuda stream protocol for older versions of torch.

Compat wrapper for Path.walk (3.12+) with a fallback for < 3.12.