What problem does it solve? When adding or modifying a Triton kernel in the Flash Linear Attention library, it is hard to know which test combinations are missing and which low-level implementation pitfalls (grid width, integer overflow, deprecated APIs) can silently break correctness on NVIDIA, AMD, or Intel hardware. ## Core Features & Use Cases - Coverage Matrix Checklist: Enumerates test axes such as varlen layout, forward/backward, gate modes, GVA, head dimensions, and backend verifiers so you can spot missing combinations. - Triton Safety Rules: Enforces int64 address arithmetic, narrow program-ID handling, and avoidance of deprecated tl.make_block_ptr. - Platform-Aware Test Style: Directs use of fla.utils device helpers and IS_NVIDIA/IS_AMD/IS_INTEL flags instead of hard-coded torch.cuda checks. - Use Case: Before submitting an MR that changes the KDA kernel's varlen offsets, use this Skill to list the coverage matrix, add the missing varlen backward tests in tests/ops/test_kda.py, and run pytest to verify. ## Quick Start Ask the assistant to review the test coverage for the kernel you just modified in fla/ops and suggest which tests to add or run before opening the merge request.
