aoti_torch__reinterpret_tensor by larryliu0820 · Pull Request #14688 · pytorch/executorch

larryliu0820 · 2025-09-30T08:10:13Z

Stack from ghstack (oldest at bottom):

Summary:

Introduced aoti_torch__reinterpret_tensor, which creates a new tensor view that reinterprets the same underlying memory with custom shape and strides.

Reviewed By:

Differential Revision:

Summary: This diff introduce `aoti_torch_delete_tensor_object` for deleting tensors created during cuda backend inference. Reviewed By: [ghstack-poisoned]

Summary: This function introduce aoti_torch_create_tensor_from_blob_v2, a function that create tensor from data blob and custom stride and size. Worth to notice that unlike aoti_torch_empty_strided, the tensor created by aoti_torch_create_tensor_from_blob_v2 will not have the control of the memory blob. Therefore when we delete it, the memory will not be freed. Reviewed By: Differential Revision: [ghstack-poisoned]

Summary: Introduced aoti_torch__reinterpret_tensor, which creates a new tensor view that reinterprets the same underlying memory with custom shape and strides. Reviewed By: Differential Revision: [ghstack-poisoned]

pytorch-bot · 2025-09-30T08:10:18Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14688

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 1 New Failure, 27 Pending, 1 Unrelated Failure

As of commit b4adeda with merge base 44972ad ():

NEW FAILURE - The following job has failed:

Build Presets / zephyr (zephyr) / build (gh)

FLAKY - The following job failed but was likely due to flakiness present on trunk:

periodic / test-models-linux (cmake, vit, xnnpack-quantization-delegation, linux.2xlarge, 90) / linux-job (gh) (detected as infra flaky with no log or failing log classifier)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #14690 * #14689 * #14688 * #14687 * __->__ #14686 Summary: This diff introduce `aoti_torch_delete_tensor_object` for deleting tensors created during cuda backend inference. Reviewed By:

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #14690 * #14689 * #14688 * __->__ #14687 * #14686 Summary: This function introduce aoti_torch_create_tensor_from_blob_v2, a function that create tensor from data blob and custom stride and size. Worth to notice that unlike aoti_torch_empty_strided, the tensor created by aoti_torch_create_tensor_from_blob_v2 will not have the control of the memory blob. Therefore when we delete it, the memory will not be freed. Reviewed By: Differential Revision:

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #14690 * #14689 * #14688 * __->__ #14700 * #14686 Summary: This is a manual cherry pick of #14687 This function introduce aoti_torch_create_tensor_from_blob_v2, a function that create tensor from data blob and custom stride and size. Worth to notice that unlike aoti_torch_empty_strided, the tensor created by aoti_torch_create_tensor_from_blob_v2 will not have the control of the memory blob. Therefore when we delete it, the memory will not be freed. Reviewed By: Differential Revision:

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #14690 * __->__ #14689 * #14688 * #14687 * #14686 Summary: This diff introduce `aoti_torch_copy_`, the function for copying tensor inside cuda backend. Right now it only support copy between tensors with same dtype. Reviewed By: Differential Revision:

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #14690 * #14689 * #14688 * #14687 * #14686 Summary: this is a comprehensive update to use ET_CHECK macro to replace original if..else check for better follow et's law Reviewed By: Differential Revision:

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * pytorch#14690 * pytorch#14689 * pytorch#14688 * __->__ pytorch#14700 * pytorch#14686 Summary: This is a manual cherry pick of pytorch#14687 This function introduce aoti_torch_create_tensor_from_blob_v2, a function that create tensor from data blob and custom stride and size. Worth to notice that unlike aoti_torch_empty_strided, the tensor created by aoti_torch_create_tensor_from_blob_v2 will not have the control of the memory blob. Therefore when we delete it, the memory will not be freed. Reviewed By: Differential Revision:

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * pytorch#14690 * pytorch#14689 * __->__ pytorch#14688 * pytorch#14687 * pytorch#14686 Summary: Introduced aoti_torch__reinterpret_tensor, which creates a new tensor view that reinterprets the same underlying memory with custom shape and strides. Reviewed By: Differential Revision:

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * pytorch#14690 * __->__ pytorch#14689 * pytorch#14688 * pytorch#14687 * pytorch#14686 Summary: This diff introduce `aoti_torch_copy_`, the function for copying tensor inside cuda backend. Right now it only support copy between tensors with same dtype. Reviewed By: Differential Revision:

) Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ pytorch#14690 * pytorch#14689 * pytorch#14688 * pytorch#14687 * pytorch#14686 Summary: this is a comprehensive update to use ET_CHECK macro to replace original if..else check for better follow et's law Reviewed By: Differential Revision:

larryliu0820 · 2025-09-30T08:10:13Z