![]() ![]() ![]() Fix copy from mps to cpu device when storage_offset is non-zero ( #109557).Fix the meta device implementation for nn.functional.scaled_dot_product_attention ( #110893).Fix handling of non-contiguous bias_mask in torch.nn.functional.scaled_dot_product_attention ( #112673).Continue if param not exist in sharded load in ( #109116).Add planner argument to .load_sharded_optimizer_state_dict ( #111393).Remove unnecessary slow code in .load_sharded_optimizer_state_dict ( #111687).Fix the behavior of torch.new_zeros in onnx due to TorchScript behavior change ( #111694).Ignore beartype=0.16.0 within the onnx package as it is incompatible ( #111861).Fix missing CUDA initialization when calling FFT operations ( #110326).Fix cuda driver API to load the appropriate.Fix segfault in foreach_* operations when input list length does not match ( #112349).Remove spurious warning in comparison ops ( #112170). ![]() This release is meant to fix the following issues (regressions / silent correctness): ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |