[Intel GPU] Avoid atomic add for XPU device in satter_add by deterministic mode #137966
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/137966
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 1b26d16 with merge base 565a794 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
EikanWang
changed the title
|
May I know what's the impact? Does it lead to any case failure? |
EikanWang
added
topic: not user facing
|
Please seek CI approval before scheduling CIFlow labels |
|
Please add test cases. |
PenghuiCheng
changed the title
Yes, the UT covered in https://github.com/intel/torch-xpu-ops/blob/main/test/xpu/test_scatter_gather_ops_xpu.py. |
PenghuiCheng
force-pushed
the
penghuic/scatter_add_deterministic
branch
2 times, most recently
from
3d7a311 to
fc26d80
Compare
Signed-off-by: Cheng Penghui <[email protected]>
PenghuiCheng
force-pushed
the
penghuic/scatter_add_deterministic
branch
from
fc26d80 to
1b26d16
Compare
EikanWang
changed the title
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
…istic mode (pytorch#137966) The "scatter_add" op with the deterministic mode in XPU device is not implemented, it will report that "scatter_add_kernel" does not have a deterministic implementation in UT. Just like the implementation of CUDA, we need to check _deterministic_algorithms in scatter_add op for the XPU device. The UT is in: https://github.com/intel/torch-xpu-ops/blob/main/test/xpu/test_scatter_gather_ops_xpu.py. We reused [PyTorch UT code]( https://github.com/pytorch/pytorch/blob/96b30dcb25c80513769dae2a8688aec080b00117/test/test_scatter_gather_ops.py#L233). Now the UT case is [skipped in torch-xpu-ops test](https://github.com/intel/torch-xpu-ops/blob/4fa7921f1e9a0bf300d25da9b8758524f2751092/test/xpu/skip_list_common.py#L731). Will open it when this PR is merged. Pull Request resolved: pytorch#137966 Approved by: https://github.com/EikanWang, https://github.com/guangyey, https://github.com/ezyang
6 participants
The "scatter_add" op with the deterministic mode in XPU device is not implemented, it will report that "scatter_add_kernel" does not have a deterministic implementation in UT.
Just like the implementation of CUDA, we need to check _deterministic_algorithms in scatter_add op for the XPU device.
The UT is in: https://github.com/intel/torch-xpu-ops/blob/main/test/xpu/test_scatter_gather_ops_xpu.py. We reused PyTorch UT code.
Now the UT case is skipped in torch-xpu-ops test. Will open it when this PR is merged.