Permalink
Comparing cha 10000 nges
Choose two branches to see what’s changed or to start a new pull request.
If you need to, you can also or
learn more about diff comparisons.
Open a pull request
Create a new pull request by comparing changes across two branches. If you need to, you can also .
Learn more about diff comparisons here.
...
- 9 commits
- 17 files changed
- 9 contributors
Commits on Oct 18, 2024
-
update getting started xpu (#138090)
update get start xpu (#137479) 1. respect the comment from the community, downgrade the "Beta" to "Prototype" for the first xpu release with wheel 2. add wheels installation of torchaudio & torchvision for nightly on Windows Pull Request resolved: #137479 Approved by: https://github.com/atalman, https://github.com/malfet (cherry picked from commit 7ba706c) Co-authored-by: Zheng, Zhaoqiong <zhaoqiong.zheng@intel.com>
Commits on Oct 22, 2024
-
Don't try to load cufile (#138539)
Don't try to load cufile (#138501) Trying to loading it caused a big issue with 2.5.0 release - #138324 cufile is not actually used currently by default, see #133489 Pull Request resolved: #138501 Approved by: https://github.com/atalman, https://github.com/mikaylagawarecki, https://github.com/malfet (cherry picked from commit 012ff2a) Co-authored-by: Sergii Dymchenko <sdym@meta.com>
-
Add link to torch.compile the missing manual in troubleshooting (#137369
) Add link to torch.compile the missing manual in troubleshooting (#137301) Fixes #ISSUE_NUMBER Pull Request resolved: #137301 Approved by: https://github.com/svekars Co-authored-by: Svetlana Karslioglu <svekars@meta.com> (cherry picked from commit 22e19bd) Co-authored-by: Michael Lazos <mlazos@meta.com>
-
Update cpuinfo submodule (#138600)
Spiritual cherry-pick of #138351 that picks pytorch/cpuinfo#258 into the branch Fixes #138333 Test Plan: `python -c "import torch"` finishes without any output on the screen
-
-
[SDPA-CUDNN] Make CuDNN Attention Opt in (#138587)
[SDPA-CUDNN] Make CuDNN Attention Opt in (#138522) # Summary Currently we have a `cudnn_order` that says on H100 w/ new enough CuDNN backend (we ship a 9.1 version in OSS) try to run CuDNN attention first. We have already encountered a few bugs with the release of 2.5: 1. #138529 2. huggingface/diffusers#9704 3. #138354 In light of the above we are going to make the CuDNN backend Opt-in by default. This can be done easily with the context manager for choosing backends I.e.: ``` Python from torch.nn.attention import sdpa_kernel, SDPBackend with sdpa_kernel(SDPBackend.CUDNN_ATTENTION): out = F.scaled_dot_product_attention(q, k, v) ``` This PR puts the CuDNN backend as the lowest precedence in the backend list, meaning that the Math backend will always be chosen unless disabled (which is done via the context manager). Cc @atalman Pull Request resolved: #138522 Approved by: https://github.com/ngimel, https://github.com/eqy, https://github.com/malfet (cherry picked from commit 9a9a0ab) Co-authored-by: drisspg <drisspguessous@gmail.com>
-
[MPS] Fix sliced cast (#138535)
[MPS] Fix sliced cast (#138314) This fixes internal crash due to the invalid bufer size computation if sliced API is used Not sure what was the purpose of ```c++ IntArrayRef baseShape; if (src.is_view()) { baseShape = src._base().sizes(); } else { baseShape = getIMPSAllocator()->getBufferShape(src.storage().data()); } int flattenedShaped = 1; for (const auto i : c10::irange(baseShape.size())) { flattenedShaped *= baseShape[i]; } ``` As flattenShaped could be much easier computed as `[srcBuf lengh]/src.element_size()`, and even if `srcBuf` is padded it's a safe thing to do. When someone allocated buffer to hold say uint8 and that view-casted it to float16, attempt to compute `baseShape` returned sizes of original tensor in its data type, rather than size in new dtypes Fixes #137800 Pull Request resolved: #138314 Approved by: https://github.com/albanD, https://github.com/DenisVieriu97 (cherry picked from commit de16159) Co-authored-by: Nikita Shulga <nikita.shulga@gmail.com>
Commits on Oct 23, 2024
-
Disabling amp context when invoking compiler (#138659)
Disabling amp context when invoking compiler (#138624) Fix for #133974 Pull Request resolved: #138624 Approved by: https://github.com/bdhirsh, https://github.com/drisspg (cherry picked from commit 5942b29) Co-authored-by: eellison <elias.ellison@gmail.com>
Loading
This comparison is taking too long to generate.
Unfortunately it looks like we can’t render this comparison for you right now. It might be too big, or there might be something weird with your repository.
You can try running this command locally to see the comparison on your machine:
git diff v2.5.0...v2.5.1