Consecutive calls shift data_ptr by _POOL_ALIGNMENT bytes.

Blob root contains multiple leading components of the relative path.

unlock_clocks is best-effort — exceptions are swallowed.

Returned tensors have the same dtype as the original input.

End-to-end evaluation: load sample -> package -> compile (if C++) -> execute -> assert PASSED.

_emit() must use allow_nan=False so NaN/Inf crashes loudly (no torch required).

Works with no inputs and no outputs.

SOL_EXECBENCH_GPU_CLK_MHZ and SOL_EXECBENCH_DRAM_CLK_MHZ override preset values.

Env var overrides allow locking even for unknown devices.

build_driver() must produce a syntactically valid Python script.

Full relative path exists directly under blob root.

Error exactly at the tolerance boundary.

End-to-end evaluation: load example -> package -> compile (if C++) -> execute -> assert PASSED.

Expanded (stride-0) tensors store only physical elements, not logical numel.

Falls through to second blob root when first has no match.

float16 inputs are upcast to float32 internally.

FP8 dtypes must skip heuristics (torch.randn etc. don't support FP8).

A single spurious inf in an otherwise correct tensor fails.

Inf in output produces inf abs_error and fails.

Inf in reference (but not output) is also incorrect.

Kernel that outputs Inf must produce strictly valid JSON traces.

Inf in output returns Correctness with has_inf=True and finite errors.

Returned views contain the same data as the original input.

Source data is re-copied each call, so in-place mutation of a view does not corrupt subsequent calls.

4096x4096 matmul takes measurably longer than 128x128.

4096x4096 matmul takes measurably longer than 128x128.

Same as above but with relaxed ratio — passes.

One near-zero element with error, rest are fine.

Kernel that returns a torch.Tensor subclass → REWARD_HACK.

Kernel that calls load_inline at import time → subprocess crash.

Kernel that calls load_inline inside run() → RUNTIME_ERROR.

lock_clocks=False + SOL_EXECBENCH_CLOCKS_LOCKED=1 → should proceed.

lock_clocks=False + SOL_EXECBENCH_CLOCKS_LOCKED=0 → should proceed.

lock_clocks=True + SOL_EXECBENCH_CLOCKS_LOCKED=1 → should proceed.

lock_clocks=True + SOL_EXECBENCH_CLOCKS_LOCKED=0 → should reject.

Known device → calls -lgc and -lmc with correct frequencies.

Matching infinities (both +inf) are still incorrect.

Matching negative infinities are still incorrect.

Both tensors have -inf at the same position — accepted.

Default behavior: matching -inf fails.

Trimmed min/max ratio stays bounded across matmul sizes. Small matmuls (64-512) are launch-overhead dominated and show wider variance

Cap uses strict greater-than (not strictly greater).

Error above cap fails even with good matched ratio.

Without cap, large outlier errors pass if matched ratio is met.

Error below cap passes normally.

+inf vs -inf at the same position fails.