(self, n_head: int, n_head_kv: int)
| 498 | return UnquantizedTensor(self.ndarray[r * n_part : r * n_part + r, ...]) |
| 499 | |
| 500 | def permute(self, n_head: int, n_head_kv: int) -> UnquantizedTensor: |
| 501 | return UnquantizedTensor(permute(self.ndarray, n_head, n_head_kv)) |
| 502 | |
| 503 | |
| 504 | def load_unquantized(lazy_tensor: LazyTensor, expected_dtype: Any = None, convert: bool = False) -> NDArray: |
nothing calls this directly
no test coverage detected