PyTorch MergeBot
9a883007a2
Revert "Implement cuda graphs implementation of torch.cond and torch.while_loop ( #140979 )"
...
This reverts commit c7515da7b0 .
Reverted https://github.com/pytorch/pytorch/pull/140979 on behalf of https://github.com/huydhn due to This change has been reported to break internal code ([comment](https://github.com/pytorch/pytorch/pull/140979#issuecomment-2657361940 ))
2025-02-13 18:04:26 +00:00
Daniel Galvez
c7515da7b0
Implement cuda graphs implementation of torch.cond and torch.while_loop ( #140979 )
...
This is a new PR for #130386 , which got stale and was closed. Since I force-pushed to that branch in order to rebase it on top of main, the PR can no longer be reopened, according to https://github.com/isaacs/github/issues/361
I fixed the possibly-not-warmed-up problem described here: https://github.com/pytorch/pytorch/pull/130386/files#r1690856534
Since starting this, torch.cond and torch.while_loop now apparently have support for backward passes. I will look into what it might take to support that.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/140979
Approved by: https://github.com/eqy , https://github.com/eellison
2025-02-11 18:16:15 +00:00
Boyuan Feng
7e80758efc
[CUDAGraph][Docs] add cuda to torch.randn ( #144793 )
...
Previous doc example created `torch.randn` tensor on cpu so CUDAGraph was skipped.
Fixes #144386
Pull Request resolved: https://github.com/pytorch/pytorch/pull/144793
Approved by: https://github.com/eellison
2025-01-15 18:02:10 +00:00
Boyuan Feng
43998711a7
[CUDAGraph] add more docs for cudagraph trees ( #127963 )
...
This PR adds more documentation for CUDAGraph Trees, including
- Iteration Support
- Input Mutation Support
- Dynamic Shape Support
- NCCL Support
- Reasons for Skipping CUDAGraph
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127963
Approved by: https://github.com/eellison
2024-06-18 02:07:07 +00:00
Guo Yejun
5323b2daa5
[docs] add mode="reduce-overhead" into torch.compile to enable cuda g… ( #116529 )
...
…raph
Pull Request resolved: https://github.com/pytorch/pytorch/pull/116529
Approved by: https://github.com/eellison
2024-01-05 22:54:20 +00:00
Guo Yejun
4aa2c51a09
[doc] fix typo on graph 3 that is recorded ( #114666 )
...
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114666
Approved by: https://github.com/eellison
2023-11-28 20:40:13 +00:00
Guo Yejun
4a35ec3c0e
[docs] correct the code for cudagraph trees integration ( #114583 )
...
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114583
Approved by: https://github.com/eellison
2023-11-28 20:28:52 +00:00
eellison
7fe51e3e9b
Add cudagraph_mark_step_begin in torch.compiler, reference in error message ( #111722 )
...
Pull Request resolved: https://github.com/pytorch/pytorch/pull/111722
Approved by: https://github.com/ezyang , https://github.com/msaroufim
2023-10-25 21:53:21 +00:00
Svetlana Karslioglu
4d3ea5df65
Restructure torch.compile docs ( #105376 )
...
Current torch.compile docs have become a bit of a mess with the docs expanded in the left nav. This PR moves them under the torch.compiler menu item in the left nav. A bunch of rewrites were made in collaboration with @msaroufim to address formatting issues, latest updates that moved some of the APIs to the public torch.compiler namespace were addressed as well. The documentation is broken down in three categories that address three main audiences: PyTorch users, Pytorch Developers and PyTorch backend vendors. While, the user-facing documentation was significantly rewritten, dev docs and vendor docs kept mostly untouched. This can be addressed in the follow up PRs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/105376
Approved by: https://github.com/msaroufim
2023-07-28 20:58:57 +00:00