tilegym-cutile-python

Define a function decorated with @ct.kernel that loads tiles with ct.load using tile indices, performs tile operations, and stores results with ct.store. Launch it with ct.launch using a grid computed via ct.cdiv, then validate output against a PyTorch reference.

For modules with three or more operations, the skill uses an orchestration workflow: an Analyzer decomposes the forward pass into kernel specs, Kernel Agents generate each kernel in parallel, and a Composer merges them into one validated file. Every compute op must run through ct.launch.

cuTile requires every tile dimension in ct.load and ct.store to be a power of 2. Round sizes up with 2**((size-1).bit_length()) and pass both the original and padded sizes as ct.Constant-annotated kernel parameters.

ct.load and ct.store use tile indices, not element indices. Use ct.load(A, index=(bid_m, k), shape=(BLOCK_M, BLOCK_K)) rather than multiplying the block id by the block size, which is a common source of wrong results.

No. The skill enforces a pure cuTile forward path where all compute goes through @ct.kernel and ct.launch. PyTorch is only permitted for tensor allocation, reshaping, concatenation, and weight storage in __init__.

Use atol=1e-3 and rtol=1e-3 for float32 outputs, and atol=1e-2 and rtol=1e-2 for float16 or bfloat16. Accumulate in float32 and cast back to the output dtype to avoid precision failures.