Get concrete variable axis values directly from input values. Convenience method that combines extract_shapes and get_var_axes_values.

Extract the file path from the entry point specification. The entry point format is '{file_path}::{function_name}', and this method r

Get the entry source file specified in the build spec. Returns ------- Optional[SourceFile] The SourceFile object

Extract the function/symbol name from the entry point specification. The entry point format is '{file_path}::{function_name}', and this metho

Return the memoized deterministic hash of the solution content. This hash is computed from all fields that affect the solution's behavior:

Load a Pydantic BaseModel object from a JSON file. Parameters ---------- model_cls : Type[BaseModel] The Pydantic BaseModel

Load a list of Pydantic BaseModel objects from a JSONL file. Each line in the JSONL file should contain a valid JSON object that can be deser

Skip tests based on hardware availability. Also skips timing_serial tests unless explicitly selected with -m.

Register custom markers.

Redirect stdout/stderr to log file. Returns original stdout and stderr file descriptors for printing to terminal.

Wrapper for Triton kernel (destination-passing style).

Post-normalization residual connection: output = residual + RMSNorm(sublayer_output) RMSNorm computation: 1. Compute variance: mean

Fused attention output projection with residual addition. This performs: 1. Linear projection: attn_output @ o_proj_weight.T 2.

* @brief Python-bindable 'run' function for RMSNorm. * * This function serves as the entry point from Python. It performs extensive * validation on

Apply rotary position embeddings to query or key tensor. The rotation is applied using the formula: - Split features into pairs (x1,

Fused attention output projection with residual addition. This performs: 1. Linear projection: attn_output @ o_proj_weight.T 2.

Fused gate and up projection with GELU-tanh activation. Computes: gelu_tanh(x @ gate_proj.T) * (x @ up_proj.T) Args: x:

Save a Pydantic BaseModel object to a JSON file. Parameters ---------- object : BaseModel The Pydantic BaseModel instance to

Save a list of Pydantic BaseModel objects to a JSONL file. Each object is serialized as a separate JSON object on its own line. Paramete

Compute anchored score S(T_k). S(T_k) = 1 / (1 + (T_k - T_SOL) / (T_b - T_SOL)) T_b can be set as any fast implementation of the refere

Entirely inf output fails when reference is finite.

All elements are matching -inf — treated as fully correct.

Entirely NaN output fails.

All-zeros output (nonzero ref) returns finite error values, no flags.

For near-zero references, atol term dominates tolerance.

setup() is called once per timed iteration (not during measurement).

Different return_mode values produce correct summaries.

Warmup and rep counts are respected.

bfloat16 inputs are upcast to float32 internally.

Both output and reference are zero — no error.

CLI e2e: run sol-execbench on a GQA paged-decode problem with safetensors inputs.

When clocks are not locked, log starts with 'Clocks locked: no'.

When clocks are locked, log starts with 'Clocks locked: yes'.

Tensor arguments are cloned (different storage).

compute_error_stats with NaN output returns finite abs/rel error values.

Doubling matmul dimension should roughly 8x the FLOPs and increase time.

Doubling matmul dimension should roughly 8x the FLOPs and increase time.

Verify extracted source files match the content in definition.json and solution JSON.

All Correctness instances from compute_error_stats have JSON-safe floats.

C++ files may use getCurrentCUDAStream (the default stream).

CUPTI median should be <= CUDA events median. For short kernels CUPTI is strictly less due to the ~4-6us event overhead. For long ke

tolerance_mhz=10 makes a 30 MHz deviation fail.