()
| 311 | |
| 312 | |
| 313 | def _get_module(): |
| 314 | global _module |
| 315 | if _module is None: |
| 316 | _module = compile_cuda(CUDA_SRC, "layernorm_cuda") |
| 317 | return _module |
| 318 | |
| 319 | |
| 320 | def kernel_fn(x: torch.Tensor, weight: torch.Tensor, bias: torch.Tensor) -> torch.Tensor: |
no test coverage detected