tl.program_id(axis=0) returns local PE id within cube, tl.program_id(axis=1) returns cube id. Enables cube-aware sharding in benchmark kernels. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>