pytorch / pytorch/executorch

[Vulkan] Fix 64-bit I/O in Vulkan Delegate

Open
#13,083 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

module: vulkan
Dominant language
Python
Stars
5k
Forks
1.2k
Avg merge
2d 10h
Merged PRs (30d)
581

Description

🐛 Describe the bug

Context

During export, the Vulkan delegate will downcast int64 and float64 tensors to int32 and int32 since 64-bit types are cumbersome to handle in compute shaders. However, this produces an issue the inputs and outputs of a model are 64-bit types; the ExecuTorch runtime will still be expecting to receive 64-bit data, but under the hood Vulkan will be treating the inputs and outputs as 32-bit.

Fixing

  • Add a mechanism to allow the Vulkan delegate to automatically convert data to/from the source/destination ETensor type during input and output.
Versions

N/A

cc @manuelcandales @cbilgin

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with the Vulkan delegate's export path and the ExecuTorch runtime's input and output handling. Trace how 64-bit tensors are represented during export, then verify that Vulkan inputs and outputs are converted to and from the source or destination ETensor types without changing the runtime's expected data.

Written by the indexing model from the issue text.

Assessment

Domain
computer-graphics
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.