Fill the tensor with a value. If not materialized, the fill is deferred until sync().

Zero the tensor.

Dispatch an function to a library :param lib_name: the name of the library :param lib_path: the path of the library :

Get the underlying tensor (must be materialized).

Create a 2D TMA descriptor (convenience function). Parameters ---------- tensor : torch.Tensor Source tensor on CUDA device.

Create a shallow copy of the node to avoid modifying the original

Process a list of nodes/values (recursively visit nodes)

Get the dtype (available before materialization).

sleep for duration_ms in CUDA kernel

fail_if_too_small: if `x` or `y` is not much larger than `atol * 5`, raise error.

Compile and build a kernel that can be launched. Parameters ---------- passes : list List of pas

Write a line to main buffer (for mixin compatibility).

Write a line with proper indentation.

Get size like torch.Tensor.size().

Override visit to cache results.

for float types, return uniform distribution [-1, 1] for int types, return uniform distribution in [int_type_min, int_type_max]

Public method to infer type of a single AST node. Delegates to the appropriate visit_* method.

Create a (potentially lazy) tensor. Args: name: Name for debugging/tracking shape: Tensor shape

Dump JSON with parallel processing of large parallel_field field

Evaluate an AST annotation node to LLType.

Convert LLType to C++ type string (for mixin compatibility).

Checks if 'nvidia-smi' is available on the system's PATH.

Free a nvshmem tensor, handling both LazyTensor and regular torch.Tensor.

Convert Python native values (int/float/bool) to Literal nodes

Synchronize the CUDA device.

Compile kernel to CUDA code. Parameters ---------- passes : list List of passes to apply codegen_

Pretty-print a results table to console.

Combine headers, structs, and main code into final C++ output.

Create a new AST node of the specified type, preserving special attributes from source. Args: node_type: The AST node class to i

Build a LittleKernel kernel with standard passes.

Helper function to codegen with proper type inference.

Get the LLType of a value (for mixin compatibility).

Retrieve comprehensive GPU information. Returns: dict with keys: name, sm_count, sm_arch, clock_mhz, l2_size_bytes, memory_bytes,

Get or create results directory.

Check if the type is a pointer

Pretty-print GPU information.

Sync, free symmetric memory, sync. Use _mori_ptr when present (mori tensor), else data_ptr().

Get struct stub information for a Python class name.