| name | tilelang |
| description | examples/tilelang 目录的完整生成 skill:TileLang JIT 内核测试(CUDA)。 |
CLAUDE.md — examples/tilelang
TileLang JIT-compiled kernel tests for CUDA. Exercises the TileLang integration in pyqcu/tools/_einsum.py and pyqcu/tools/_matul.py.
Files
| File | Purpose |
|---|
conftest.py | TileLang test entry |
Usage
python examples/tilelang/conftest.py
Requires TileLang to be installed (optional dependency, silently degrades if unavailable).
Known upstream defects & hardware limits (2026-08-24, bug36)
- v0.1.7.post3 ships four same-named emitter classes (one per architecture fallback path)
missing
_legalize_to_buffer_region — the installed copy monkey-patches all four.
Locate the real throw site via full-stack traceback; patching one copy is not enough.
Installed version vs upstream tag diff is IDENTICAL → native upstream defect.
- SM70 (V100) TensorCore has no fp32 input instruction (
mma_sm70.h static_assert):
slow fp32-gemm is a hardware limit, not fixable by hot-patch; long-term fix = upgrade TileLang.
- Measured anchor after hot-fix: 4096³ fp16 matmul 38.7 TFLOPS ≈ 94% of cuBLAS.