Return (definition.json, workload.jsonl, config.json?, solution.json?) inside a problem directory.

Write stdout/stderr from a failed CLI invocation to a log file.

Convert an SM version string (e.g. 'sm_90', 'sm_100a') to a gencode flag.

Number of contiguous storage elements spanned by *tensor*. For a standard contiguous tensor this equals ``numel()``. For broadcast/e

Write solution source files to the staging directory.

Wrap an external ``solution.py`` file as a Solution dict.

Construct a Solution dict that wraps the definition's reference code.

Build a solution for a problem directory. If *solution_name* is given, looks for that file inside *problem_dir*. ``.json`` files are loaded d

Check for non-finite values and all-zeros output. Returns a ``Correctness`` describing the failure when either tensor contains inf/nan values

Evaluate a SOL-ExecBench solution on GPU. \b Two ways to specify the problem: 1) Positional: sol-execbench <problem_dir> --solution sol

Return a sorted list of problem directories under *benchmark_dir*. Each problem directory must contain definition.json and workload.jsonl. If

Get concrete variable axis values from input shapes. Parameters ---------- input_shapes : Iterable[Optional[tuple[int, ...]]]

Get L2 cache size in bytes for the given CUDA device. Args: device: CUDA device (int, torch.device, or None for current device).

Get concrete output shapes given variable axis values. Parameters ---------- var_values : Optional[dict[str, int]], default=N

Check if this is a workload-only trace. Returns ------- bool True if this is a workload trace without solution/ev

Parse JSONL Trace dicts from eval_driver stdout.

Print a table summarizing all problem results.

Safely evaluate a simple arithmetic expression with variables. Allowed: numbers, variable names, + - * / // % **, parentheses, unary +/-

Invoke ``sol-execbench`` and return parsed trace dicts (or None on error).

Compile a .cu file and link against -lcuda -lcudart.

Compile a .cpp file and link against -lcudnn.

RMSNorm forward pass using cuDNN Frontend graph API.

Vectorized bf16 elementwise add using CuTe DSL with 128-bit copies.

Compile a .cu file with CUTLASS headers.

--- Pybind11 Module Definition --- Exposes the 'run' function to Python, making it callable as a C++ extension.

Ensure clock env vars are unset so preset logic is exercised.

Get all variable names from a mathematical expression.

Forward pass for RMS normalization with improved numerical stability.

Validate entry_point format. Raises ------ ValueError If entry_point doesn't follow the required format.

Validate that input names are not axes. Raises ------ ValueError If any input name is an axis.

Validate that input and output names are unique and do not overlap. Raises ------ ValueError If the input or outp

Validate languages support matrix. Raises ------ ValueError If the languages are not valid.

Validate that reference contains valid Python code with a 'run' function. Raises ------ ValueError If the referen

Validate that ``run()`` parameter names match the ``inputs`` keys in order. Raises ------ ValueError If the param

Validate source path for security. Raises ------ ValueError If the path contains security issues (absolute paths

Validate source file paths for uniqueness and entry file existence. Raises ------ ValueError If duplicate source

Validate correctness and performance fields based on status. Ensures that correctness and performance metrics are present or absent b

Validate that tensor shapes reference defined axes. Ensures that all axis names used in input and output tensor shapes are properly d

Verify that the custom inputs entrypoint is a valid identifier defined in reference code.

Append a list of Pydantic BaseModel objects to a JSONL file. Each object is serialized as a separate JSON object and appended to the end of t

Get all constant axes and their values. Returns ------- dict[str, int] Dictionary mapping constant axis names to

Get all expression axis names. Returns ------- dict[str, AxisExpr] List of all expression axis names defined in t