accelerate

aba3b8c7 - Prefer `is_torch_tensor` over `hasattr` for torch.compile. (#2387)

Prefer `is_torch_tensor` over `hasattr` for torch.compile. (#2387) * Prefer `is_torch_tensor` over `hasattr` for `torch.compile`. `torch.compile` breaks when using `hasattr` but succeeds when using `isinstance(torch.Tensor)`. This commit short-circuits the `hasattr` call for `torch.Tensor`s if possible. Note: `is_npu_available` is also not torch.compila compatible due to (1) lru_cache and (2) importlib checks, so I've moved it into the try block, catching the AssertionError instead. * Fix torch.device("npu"). This is not available in non-npu pytorch. Note that torch.device automatically assigns an index when created as torch.device("npu"), so overwriting device with `"npu:0"` is only required if device is a string "npu". * Remove unittest.main execution. * Fix style broken by merge save. * Import operations functions directly. * fix style * Fix imports attempt 2. * Re-raise error if no NPU available.