Set is_non_overlapping_and_dense_ flag in OpaqueTensorImpl constructor#49470
Set is_non_overlapping_and_dense_ flag in OpaqueTensorImpl constructor#49470SS-JIA wants to merge 5 commits intogh/SS-JIA/14/basefrom
Conversation
[ghstack-poisoned]
💊 CI failures summary and remediationsAs of commit 98c6bf7 (more details on the Dr. CI page):
🕵️ 2 new failures recognized by patternsThe following CI failures do not appear to be due to upstream breakages:
|
…Impl" [ghstack-poisoned]
|
Not sure who the right person is to review this patch. The backstory is that we ran into runtime crashes on the Vulkan backend recently which Stephen has narrowed to #48625. |
|
Can you please add tests that would exercise codepaths previously leading to crashes? We are working on making TensorImpl more efficient, and unless you are testing for the behaviors you rely upon, it could lead to breakages in the future. |
#48625 changes the default contiguous settings for TensorImpl. Therefore, refresh_contiguous() should be called in OpaqueTensorImpl to get the correct settings. [ghstack-poisoned]
#48625 changes the default contiguous settings for TensorImpl. Therefore, refresh_contiguous() should be called in OpaqueTensorImpl to get the correct settings. [ghstack-poisoned]
|
As suggested, changed to just set the flag in the constructor. @xta0 may want to take a look at this as well, since OpaqueTensorImpl is also used by the Metal backend. |
|
Yeah we are puzzled by why this was not caught by CI as well. Investigating. Thanks. |
… constructor" #48625 changes the default contiguous settings for `TensorImpl` causing the Vulkan backend to crash. Therefore, set `is_non_overlapping_and_dense_` back to false for `OpaqueTensorImpl`. [ghstack-poisoned]
pytorch#49470) Summary: Pull Request resolved: pytorch#49470 pytorch#48625 changes the default contiguous settings for `TensorImpl` causing the Vulkan backend to crash. Therefore, add argument that can set `is_non_overlapping_and_dense_` back to false for `OpaqueTensorImpl` constructor. Test Plan: Imported from OSS Reviewed By: AshkanAliabadi Differential Revision: D25592826 Pulled By: SS-JIA fbshipit-source-id: e5d9de9a733875cb00c0546a3bc3271e5c6e23a3
pytorch#49470) Summary: Pull Request resolved: pytorch#49470 pytorch#48625 changes the default contiguous settings for `TensorImpl` causing the Vulkan backend to crash. Therefore, add argument that can set `is_non_overlapping_and_dense_` back to false for `OpaqueTensorImpl` constructor. Test Plan: Imported from OSS Reviewed By: AshkanAliabadi Differential Revision: D25592826 Pulled By: SS-JIA fbshipit-source-id: e5d9de9a733875cb00c0546a3bc3271e5c6e23a3
Stack from ghstack:
#48625 changes the default contiguous settings for
TensorImplcausing the Vulkan backend to crash. Therefore, add argument that can setis_non_overlapping_and_dense_back to false forOpaqueTensorImplconstructor.Differential Revision: D25592826