[DeviceMesh] Move global state into class method #164510
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/164510
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit bd15c9d with merge base ffc9559 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
pytorch-bot
bot
added
the
oncall: distributed
cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci [ghstack-poisoned]
cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci [ghstack-poisoned]
This is PR trying to move bookkeeping state maps from MeshEnv to DeviceMesh class members. The reason is that in general global variables are thread local and cause potential issue. We will also need to do DTensor CPU overhead benchmark for this change. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci [ghstack-poisoned]
lw
approved these changes
Oct 3, 2025
Contributor
lw
left a comment
lw
left a comment
There was a problem hiding this comment.
This is great! I like this a lot! Thanks!
torch/distributed/device_mesh.py
Outdated
Comment on lines
423
to
434
| @property | ||
| def root_mesh(self) -> "DeviceMesh": | ||
| # If a mesh does not have a root mesh stored, it is a root mesh itself. | ||
| # A root mesh is not created through slicing. | ||
| return self._root_mesh if self._root_mesh else self | ||
|
|
||
| @root_mesh.setter | ||
| def root_mesh(self, mesh: Optional["DeviceMesh"]) -> None: | ||
| # you can add validation logic here if needed | ||
| if mesh is not None and not isinstance(mesh, DeviceMesh): | ||
| raise TypeError(f"Expected DeviceMesh or None, got {type(mesh)}") | ||
| self._root_mesh = mesh |
Contributor
There was a problem hiding this comment.
Could we make these attributes private? I'm still not convinced that the distinction between root and child mesh is really useful. At the end of the day, I believe, all we care about is which meshes belong to the same "universe" (i.e., which ones have the same shared state?).
While we discuss this and decide, I'd prefer if we avoid adding public API attributes which will limit us in the future.
Contributor
There was a problem hiding this comment.
Agree this. I also have concerns on other public methods. Please see the flatten one.
Contributor
Author
There was a problem hiding this comment.
yes making it private and will have another PR for the idea for universe.
| self.device_type, | ||
| self.mesh_dim_names, | ||
| self._thread_id, | ||
| self._root_mesh, |
Contributor
There was a problem hiding this comment.
Here and in the equality operator below, could we just use the id(...) of the root mesh?
Contributor
Author
There was a problem hiding this comment.
id(...) is the fast path, if the CPU overhead is not too big, I think we can still keep the old way. Once we have the universe thing, we can do id based comparison.
Contributor
Author
There was a problem hiding this comment.
3-5% of CPU overhead regression was observed which should be fine, because this enables us for cleaner device mesh implementation.
fegin
requested changes
Oct 3, 2025
Contributor
fegin
left a comment
fegin
left a comment
There was a problem hiding this comment.
These methods should be private.
torch/distributed/device_mesh.py
Outdated
Comment on lines
423
to
434
| @property | ||
| def root_mesh(self) -> "DeviceMesh": | ||
| # If a mesh does not have a root mesh stored, it is a root mesh itself. | ||
| # A root mesh is not created through slicing. | ||
| return self._root_mesh if self._root_mesh else self | ||
|
|
||
| @root_mesh.setter | ||
| def root_mesh(self, mesh: Optional["DeviceMesh"]) -> None: | ||
| # you can add validation logic here if needed | ||
| if mesh is not None and not isinstance(mesh, DeviceMesh): | ||
| raise TypeError(f"Expected DeviceMesh or None, got {type(mesh)}") | ||
| self._root_mesh = mesh |
Contributor
There was a problem hiding this comment.
Agree this. I also have concerns on other public methods. Please see the flatten one.
This is PR trying to move bookkeeping state maps from MeshEnv to DeviceMesh class members. The reason is that in general global variables are thread local and cause potential issue. We will also need to do DTensor CPU overhead benchmark for this change. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci [ghstack-poisoned]
This is PR trying to move bookkeeping state maps from MeshEnv to DeviceMesh class members. The reason is that in general global variables are thread local and cause potential issue. We will also need to do DTensor CPU overhead benchmark for this change. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci [ghstack-poisoned]
This is PR trying to move bookkeeping state maps from MeshEnv to DeviceMesh class members. The reason is that in general global variables are thread local and cause potential issue. We will also need to do DTensor CPU overhead benchmark for this change. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci [ghstack-poisoned]
This is PR trying to move bookkeeping state maps from MeshEnv to DeviceMesh class members. The reason is that in general global variables are thread local and cause potential issue. We will also need to do DTensor CPU overhead benchmark for this change. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci [ghstack-poisoned]
This is PR trying to move bookkeeping state maps from MeshEnv to DeviceMesh class members. The reason is that in general global variables are thread local and cause potential issue. We will also need to do DTensor CPU overhead benchmark for this change. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci [ghstack-poisoned]
fduwjj
added a commit
that referenced
this pull request
Oct 20, 2025
…oot mesh" We moved the method to get root mesh into class in #164510. This is to further clean code up. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci [ghstack-poisoned]
pytorchmergebot
pushed a commit
that referenced
this pull request
Oct 21, 2025
…165787) We moved the method to get root mesh into class in #164510. This is to further clean code up. Differential Revision: [D85090191](https://our.internmc.facebook.com/intern/diff/D85090191) Pull Request resolved: #165787 Approved by: https://github.com/fegin
Chao1Han
pushed a commit
to Chao1Han/pytorch
that referenced
this pull request
Oct 21, 2025
This is PR trying to move bookkeeping state maps from MeshEnv to DeviceMesh class members. The reason is that in general global variables are thread local and cause potential issue. We will also need to do DTensor CPU overhead benchmark for this change. 3-5% CPU overhead in DTensor has been observed: before: <img width="1147" height="535" alt="image" src="https://github.com/user-attachments/assets/9e4ac018-ec0a-46a4-8f2c-64b4dbec465c" /> After: <img width="1114" height="576" alt="image" src="https://github.com/user-attachments/assets/eaf83660-652b-4c6b-8591-f6049ccdd14c" /> running the benchmark mentioned here: pytorch#159169 Pull Request resolved: pytorch#164510 Approved by: https://github.com/lw, https://github.com/fegin
Chao1Han
pushed a commit
to Chao1Han/pytorch
that referenced
this pull request
Oct 21, 2025
pytorch#161224) Pull Request resolved: pytorch#161224 Approved by: https://github.com/lw, https://github.com/fegin ghstack dependencies: pytorch#164510
Chao1Han
pushed a commit
to Chao1Han/pytorch
that referenced
this pull request
Oct 21, 2025
…ytorch#165787) We moved the method to get root mesh into class in pytorch#164510. This is to further clean code up. Differential Revision: [D85090191](https://our.internmc.facebook.com/intern/diff/D85090191) Pull Request resolved: pytorch#165787 Approved by: https://github.com/fegin
zhudada0120
pushed a commit
to zhudada0120/pytorch
that referenced
this pull request
Oct 22, 2025
…ytorch#165787) We moved the method to get root mesh into class in pytorch#164510. This is to further clean code up. Differential Revision: [D85090191](https://our.internmc.facebook.com/intern/diff/D85090191) Pull Request resolved: pytorch#165787 Approved by: https://github.com/fegin
pytorchmergebot
pushed a commit
that referenced
this pull request
Oct 23, 2025
…hat we don't need to compare root mesh (#166003) Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. Pull Request resolved: #166003 Approved by: https://github.com/Skylion007, https://github.com/fegin
fduwjj
added a commit
that referenced
this pull request
Oct 25, 2025
…hat we don't need to compare root mesh (#166003) Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. Pull Request resolved: #166003 Approved by: https://github.com/Skylion007, https://github.com/fegin Internal: << DO NOT EDIT BELOW THIS LINE >> **GitHub Author**: fduwjj <fduwjj@gmail.com> (Meta Employee) **GitHub Repo**: [pytorch/pytorch](https://github.com/pytorch/pytorch) **GitHub Pull Request**: [#166003](#166003) Initially generated by: https://www.internalfb.com/intern/sandcastle/job/9007201528851998/ This was imported as part of a Diff Train. Please review this as soon as possible. Since it is a direct copy of a commit on GitHub, there shouldn't be much to do. Below line forces Sandcastle to run only specified contbuilds. @build_only[github-export-checks,executorch,pytorch_benchmark,pytorch_benchmark_fb,pytorch_quantization,pytorch_distributed,pytorch_distributed_gpu,pytorch_dynamo,pytorch_inductor,pytorch_inductor_fb,pytorch_functorch,pytorch_fx2trt,pytorch_diff_train_tests_ads,glow_fb_pytorch_tests,training_platform,training_platform_compatibility,training_toolkit_applications,training_toolkit_examples,training_toolkit_model_optimization,dper3_pytorch,xplat_caffe2,pytorch_dev,android-pytorch-instrumentation-tests,smart__pytorch__github_first_try_merge,frl-target-determinator,f6-buck,training_platform_for_github,sigmoid_cpu,sigmoid_gpu,aiplatform_modelprocessing_for_github,accelerators_workloads_models_slimdsnn,ae_aotinductor_benchmark_test,aps_,apf,aps_deterministic_ne_tests,dper_lib_silvertorch,torchrec,torchrec_fb,deeplearning_aot_inductor,aiplatform_modelstore] #skipfbcodelongtail #disable_code_coverage @pytorch-oss-diff-train diff-train-source-id: 8625ffb Differential Revision: [D85394822](https://our.internmc.facebook.com/intern/diff/D85394822/) ghstack-source-id: 318594805
fduwjj
added a commit
that referenced
this pull request
Oct 25, 2025
…ace _flatten_mesh_list so that we don't need to compare root mesh" Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci Differential Revision: [D85394822](https://our.internmc.facebook.com/intern/diff/D85394822) [ghstack-poisoned]
fduwjj
added a commit
that referenced
this pull request
Oct 25, 2025
…h_list so that we don't need to compare root mesh" Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci Differential Revision: [D85394822](https://our.internmc.facebook.com/intern/diff/D85394822) [ghstack-poisoned]
fduwjj
added a commit
that referenced
this pull request
Oct 25, 2025
…ace _flatten_mesh_list so that we don't need to compare root mesh" Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci Differential Revision: [D85394822](https://our.internmc.facebook.com/intern/diff/D85394822) [ghstack-poisoned]
fduwjj
added a commit
that referenced
this pull request
Oct 25, 2025
…h_list so that we don't need to compare root mesh" Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci Differential Revision: [D85394822](https://our.internmc.facebook.com/intern/diff/D85394822) [ghstack-poisoned]
fduwjj
added a commit
that referenced
this pull request
Oct 25, 2025
…hat we don't need to compare root mesh (#166003) Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. Pull Request resolved: #166003 Approved by: https://github.com/Skylion007, https://github.com/fegin Internal: << DO NOT EDIT BELOW THIS LINE >> **GitHub Author**: fduwjj <fduwjj@gmail.com> (Meta Employee) **GitHub Repo**: [pytorch/pytorch](https://github.com/pytorch/pytorch) **GitHub Pull Request**: [#166003](#166003) Initially generated by: https://www.internalfb.com/intern/sandcastle/job/9007201528851998/ This was imported as part of a Diff Train. Please review this as soon as possible. Since it is a direct copy of a commit on GitHub, there shouldn't be much to do. Below line forces Sandcastle to run only specified contbuilds. @build_only[github-export-checks,executorch,pytorch_benchmark,pytorch_benchmark_fb,pytorch_quantization,pytorch_distributed,pytorch_distributed_gpu,pytorch_dynamo,pytorch_inductor,pytorch_inductor_fb,pytorch_functorch,pytorch_fx2trt,pytorch_diff_train_tests_ads,glow_fb_pytorch_tests,training_platform,training_platform_compatibility,training_toolkit_applications,training_toolkit_examples,training_toolkit_model_optimization,dper3_pytorch,xplat_caffe2,pytorch_dev,android-pytorch-instrumentation-tests,smart__pytorch__github_first_try_merge,frl-target-determinator,f6-buck,training_platform_for_github,sigmoid_cpu,sigmoid_gpu,aiplatform_modelprocessing_for_github,accelerators_workloads_models_slimdsnn,ae_aotinductor_benchmark_test,aps_,apf,aps_deterministic_ne_tests,dper_lib_silvertorch,torchrec,torchrec_fb,deeplearning_aot_inductor,aiplatform_modelstore] #skipfbcodelongtail #disable_code_coverage @pytorch-oss-diff-train diff-train-source-id: 8625ffb ghstack-source-id: 318681631 Differential Revision: [D85394822](https://our.internmc.facebook.com/intern/diff/D85394822/)
fduwjj
added a commit
that referenced
this pull request
Oct 25, 2025
…ace _flatten_mesh_list so that we don't need to compare root mesh" Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci Differential Revision: [D85394822](https://our.internmc.facebook.com/intern/diff/D85394822) [ghstack-poisoned]
fduwjj
added a commit
that referenced
this pull request
Oct 25, 2025
…hat we don't need to compare root mesh (#166003) Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. Pull Request resolved: #166003 Approved by: https://github.com/Skylion007, https://github.com/fegin Internal: << DO NOT EDIT BELOW THIS LINE >> **GitHub Author**: fduwjj <fduwjj@gmail.com> (Meta Employee) **GitHub Repo**: [pytorch/pytorch](https://github.com/pytorch/pytorch) **GitHub Pull Request**: [#166003](#166003) Initially generated by: https://www.internalfb.com/intern/sandcastle/job/9007201528851998/ This was imported as part of a Diff Train. Please review this as soon as possible. Since it is a direct copy of a commit on GitHub, there shouldn't be much to do. Below line forces Sandcastle to run only specified contbuilds. @build_only[github-export-checks,executorch,pytorch_benchmark,pytorch_benchmark_fb,pytorch_quantization,pytorch_distributed,pytorch_distributed_gpu,pytorch_dynamo,pytorch_inductor,pytorch_inductor_fb,pytorch_functorch,pytorch_fx2trt,pytorch_diff_train_tests_ads,glow_fb_pytorch_tests,training_platform,training_platform_compatibility,training_toolkit_applications,training_toolkit_examples,training_toolkit_model_optimization,dper3_pytorch,xplat_caffe2,pytorch_dev,android-pytorch-instrumentation-tests,smart__pytorch__github_first_try_merge,frl-target-determinator,f6-buck,training_platform_for_github,sigmoid_cpu,sigmoid_gpu,aiplatform_modelprocessing_for_github,accelerators_workloads_models_slimdsnn,ae_aotinductor_benchmark_test,aps_,apf,aps_deterministic_ne_tests,dper_lib_silvertorch,torchrec,torchrec_fb,deeplearning_aot_inductor,aiplatform_modelstore] #skipfbcodelongtail #disable_code_coverage @pytorch-oss-diff-train diff-train-source-id: 8625ffb ghstack-source-id: 79a788c Differential Revision: [D85394822](https://our.internmc.facebook.com/intern/diff/D85394822/)
fduwjj
added a commit
that referenced
this pull request
Oct 25, 2025
…h_list so that we don't need to compare root mesh" Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci Differential Revision: [D85394822](https://our.internmc.facebook.com/intern/diff/D85394822) [ghstack-poisoned]
fduwjj
added a commit
that referenced
this pull request
Oct 26, 2025
…ace _flatten_mesh_list so that we don't need to compare root mesh" Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci Differential Revision: [D85394822](https://our.internmc.facebook.com/intern/diff/D85394822) [ghstack-poisoned]
fduwjj
added a commit
that referenced
this pull request
Oct 26, 2025
…h_list so that we don't need to compare root mesh" Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci Differential Revision: [D85394822](https://our.internmc.facebook.com/intern/diff/D85394822) [ghstack-poisoned]
fduwjj
added a commit
that referenced
this pull request
Oct 26, 2025
…hat we don't need to compare root mesh (#166003) Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. Pull Request resolved: #166003 Approved by: https://github.com/Skylion007, https://github.com/fegin Internal: << DO NOT EDIT BELOW THIS LINE >> **GitHub Author**: fduwjj <fduwjj@gmail.com> (Meta Employee) **GitHub Repo**: [pytorch/pytorch](https://github.com/pytorch/pytorch) **GitHub Pull Request**: [#166003](#166003) Initially generated by: https://www.internalfb.com/intern/sandcastle/job/9007201528851998/ This was imported as part of a Diff Train. Please review this as soon as possible. Since it is a direct copy of a commit on GitHub, there shouldn't be much to do. Below line forces Sandcastle to run only specified contbuilds. @build_only[github-export-checks,executorch,pytorch_benchmark,pytorch_benchmark_fb,pytorch_quantization,pytorch_distributed,pytorch_distributed_gpu,pytorch_dynamo,pytorch_inductor,pytorch_inductor_fb,pytorch_functorch,pytorch_fx2trt,pytorch_diff_train_tests_ads,glow_fb_pytorch_tests,training_platform,training_platform_compatibility,training_toolkit_applications,training_toolkit_examples,training_toolkit_model_optimization,dper3_pytorch,xplat_caffe2,pytorch_dev,android-pytorch-instrumentation-tests,smart__pytorch__github_first_try_merge,frl-target-determinator,f6-buck,training_platform_for_github,sigmoid_cpu,sigmoid_gpu,aiplatform_modelprocessing_for_github,accelerators_workloads_models_slimdsnn,ae_aotinductor_benchmark_test,aps_,apf,aps_deterministic_ne_tests,dper_lib_silvertorch,torchrec,torchrec_fb,deeplearning_aot_inductor,aiplatform_modelstore] #skipfbcodelongtail #disable_code_coverage @pytorch-oss-diff-train diff-train-source-id: 8625ffb ghstack-source-id: 318735710 Differential Revision: [D85394822](https://our.internmc.facebook.com/intern/diff/D85394822/)
pytorch-bot bot
pushed a commit
that referenced
this pull request
Oct 26, 2025
…hat we don't need to compare root mesh (#166003) Summary: Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci imported-using-ghimport Test Plan: Imported from OSS Differential Revision: D85526705 Pulled By: fduwjj
pytorchmergebot
pushed a commit
that referenced
this pull request
Oct 27, 2025
…hat we don't need to compare root mesh (#166003) (#166264) Summary: Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci imported-using-ghimport Test Plan: Imported from OSS Differential Revision: D85526705 Pulled By: fduwjj Pull Request resolved: #166264 Approved by: https://github.com/XilunWu
yiming0416
pushed a commit
that referenced
this pull request
Oct 27, 2025
…hat we don't need to compare root mesh (#166003) (#166264) Summary: Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci imported-using-ghimport Test Plan: Imported from OSS Differential Revision: D85526705 Pulled By: fduwjj Pull Request resolved: #166264 Approved by: https://github.com/XilunWu
yiming0416
pushed a commit
that referenced
this pull request
Oct 27, 2025
…hat we don't need to compare root mesh (#166003) (#166264) Summary: Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci imported-using-ghimport Test Plan: Imported from OSS Differential Revision: D85526705 Pulled By: fduwjj Pull Request resolved: #166264 Approved by: https://github.com/XilunWu
yiming0416
pushed a commit
that referenced
this pull request
Oct 27, 2025
…hat we don't need to compare root mesh (#166003) (#166264) Summary: Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci imported-using-ghimport Test Plan: Imported from OSS Differential Revision: D85526705 Pulled By: fduwjj Pull Request resolved: #166264 Approved by: https://github.com/XilunWu
yiming0416
pushed a commit
that referenced
this pull request
Oct 27, 2025
…hat we don't need to compare root mesh (#166003) (#166264) Summary: Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci imported-using-ghimport Test Plan: Imported from OSS Differential Revision: D85526705 Pulled By: fduwjj Pull Request resolved: #166264 Approved by: https://github.com/XilunWu
tianrengao
pushed a commit
that referenced
this pull request
Oct 30, 2025
…hat we don't need to compare root mesh (#166003) (#166264) Summary: Since we are already share a flattened tensor `_rank_map` across all meshes from a same root mesh, we can just use a flattened list of it to replace the comparison of root_mesh and flattened_mesh_list (because with same _rank_map and layout, the mesh tensor is guaranteed to be the same). This way we can also give back the CPU overhead added in #164510 and further simply the code. We do have a more ambitious universe-based change here: #165680 but it needs more discussions and would lead to BC breaking. We might eventually merge that PR but probably not now and this is a change which is not BC breaking and will help concatenate and 2D integration with concatenate. cc H-Huang awgu wanchaol fegin wz337 wconstab d4l3k pragupta msaroufim dcci imported-using-ghimport Test Plan: Imported from OSS Differential Revision: D85526705 Pulled By: fduwjj Pull Request resolved: #166264 Approved by: https://github.com/XilunWu
6 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
This is PR trying to move bookkeeping state maps from MeshEnv to DeviceMesh class members. The reason is that in general global variables are thread local and cause potential issue.
We will also need to do DTensor CPU overhead benchmark for this change.
3-5% CPU overhead in DTensor has been observed:
before:
After:
running the benchmark mentioned here: #159169
cc @H-Huang @awgu @wanchaol @fegin @wz337 @wconstab @d4l3k @pragupta @msaroufim @dcci