[AOTI] set alignment for aot constant #124272
Closed
Conversation
This file contains 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/124272
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (1 Unrelated Failure)As of commit d4439de with merge base d0211e2 ( FLAKY - The following job failed but was likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
chunyuan-w
added a commit
that referenced
this pull request
Apr 17, 2024
ghstack-source-id: 51c6f943d72289906d0ece2085a38556be262d23 Pull Request resolved: #124272
GPU copies the constant blob to aligned memory ([RAII_cudaMalloc](https://github.com/pytorch/pytorch/blob/d0211e207c78fafac2edaf2e14954f668e898b4a/torch/csrc/inductor/aoti_runtime/model.h#L46 ), [64-alignment](https://github.com/pytorch/pytorch/blob/d0211e207c78fafac2edaf2e14954f668e898b4a/torch/csrc/inductor/aoti_runtime/model.h#L324)) while CPU doesn't have this copy procedure for constant blob, which may result in sub-optimal performance when we want to directly use the constant blob buffer in the computation (for example when these constant blobs are the weight tensor to the oneDNN primitive). We set the alignment to the `constant.o` directly so that there's no need to copy the data to an aligned memory for CPU (when using `--rename-section`, the original section name would need to be specified for `--set-section-alignment`). cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
Apr 18, 2024
ghstack-source-id: f62a7a4690367880c6e5f4b0ea8dc1008058f51f Pull Request resolved: #124272
This was referenced Apr 18, 2024
|
Somewhat relevant to #124034, but there we need to align each tensor. |
jgong5
approved these changes
Apr 19, 2024
|
@desertfire could you help review this PR? |
desertfire
approved these changes
Apr 24, 2024
|
@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 |
alat-rights
pushed a commit
to alat-rights/pytorch
that referenced
this pull request
Apr 26, 2024
GPU copies the constant blob to aligned memory ([RAII_cudaMalloc](https://github.com/pytorch/pytorch/blob/d0211e207c78fafac2edaf2e14954f668e898b4a/torch/csrc/inductor/aoti_runtime/model.h#L46 ), [64-alignment](https://github.com/pytorch/pytorch/blob/d0211e207c78fafac2edaf2e14954f668e898b4a/torch/csrc/inductor/aoti_runtime/model.h#L324)) while CPU doesn't have this copy procedure for constant blob, which may result in sub-optimal performance when we want to directly use the constant blob buffer in the computation (for example when these constant blobs are the weight tensor to the oneDNN primitive). We set the alignment to the `constant.o` directly so that there's no need to copy the data to an aligned memory for CPU (when using `--rename-section`, the original section name would need to be specified for `--set-section-alignment`). Pull Request resolved: pytorch#124272 Approved by: https://github.com/jgong5, https://github.com/desertfire
|
This causes a failures with older objcopy: |
andoorve
pushed a commit
to andoorve/pytorch
that referenced
this pull request
May 1, 2024
GPU copies the constant blob to aligned memory ([RAII_cudaMalloc](https://github.com/pytorch/pytorch/blob/d0211e207c78fafac2edaf2e14954f668e898b4a/torch/csrc/inductor/aoti_runtime/model.h#L46 ), [64-alignment](https://github.com/pytorch/pytorch/blob/d0211e207c78fafac2edaf2e14954f668e898b4a/torch/csrc/inductor/aoti_runtime/model.h#L324)) while CPU doesn't have this copy procedure for constant blob, which may result in sub-optimal performance when we want to directly use the constant blob buffer in the computation (for example when these constant blobs are the weight tensor to the oneDNN primitive). We set the alignment to the `constant.o` directly so that there's no need to copy the data to an aligned memory for CPU (when using `--rename-section`, the original section name would need to be specified for `--set-section-alignment`). Pull Request resolved: pytorch#124272 Approved by: https://github.com/jgong5, https://github.com/desertfire
I saw this as well, is this fixed? Oh looks like this is the fix: https://github.com/pytorch/torchchat/pull/497 thanks |
pytorch-bot bot
pushed a commit
that referenced
this pull request
May 3, 2024
GPU copies the constant blob to aligned memory ([RAII_cudaMalloc](https://github.com/pytorch/pytorch/blob/d0211e207c78fafac2edaf2e14954f668e898b4a/torch/csrc/inductor/aoti_runtime/model.h#L46 ), [64-alignment](https://github.com/pytorch/pytorch/blob/d0211e207c78fafac2edaf2e14954f668e898b4a/torch/csrc/inductor/aoti_runtime/model.h#L324)) while CPU doesn't have this copy procedure for constant blob, which may result in sub-optimal performance when we want to directly use the constant blob buffer in the computation (for example when these constant blobs are the weight tensor to the oneDNN primitive). We set the alignment to the `constant.o` directly so that there's no need to copy the data to an aligned memory for CPU (when using `--rename-section`, the original section name would need to be specified for `--set-section-alignment`). Pull Request resolved: #124272 Approved by: https://github.com/jgong5, https://github.com/desertfire
chunyuan-w
added a commit
that referenced
this pull request
May 31, 2024
#124272 set the alignment to the consts_o but if there're `data_size` of tensor in the consts_o non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes, ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
May 31, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
May 31, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
Jun 3, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
Jun 3, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
Jun 3, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
Jun 3, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
Jun 3, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
Jun 3, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
Jun 3, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
Jun 3, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
Jun 6, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). This PR will improve the performance on CPU for 4 models in HF, 7 models in TIMM and 1 model in Torchbench. For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
Jun 6, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). This PR will improve the performance on CPU for 4 models in HF, 7 models in TIMM and 1 model in Torchbench. For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
Jun 7, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). This PR will improve the performance on CPU for 4 models in HF, 7 models in TIMM and 1 model in Torchbench. For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
Jun 7, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). This PR will improve the performance on CPU for 4 models in HF, 7 models in TIMM and 1 model in Torchbench. For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
Jun 7, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). This PR will improve the performance on CPU for 4 models in HF, 7 models in TIMM and 1 model in Torchbench. For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
chunyuan-w
added a commit
that referenced
this pull request
Jun 7, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). This PR will improve the performance on CPU for 4 models in HF, 7 models in TIMM and 1 model in Torchbench. For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
pytorchmergebot
pushed a commit
that referenced
this pull request
Jun 8, 2024
#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). This PR will improve the performance on CPU for 4 models in HF, 7 models in TIMM and 1 model in Torchbench. For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` Pull Request resolved: #127610 Approved by: https://github.com/jgong5, https://github.com/desertfire
TharinduRusira
pushed a commit
to TharinduRusira/pytorch
that referenced
this pull request
Jun 14, 2024
pytorch#124272 set the alignment to the `consts_o` but if there're `data_size` of tensor in the `consts_o` non divisible by the alignment, the following tensors are not aligned anymore, resulting in poor performance on CPU. We align the `data_size` as well in this PR and pad the serialized bytes. Since `size` of the tensor instead of the `data_size` is used when creating tensor from the serialized bytes ([link](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L236-L259)), there won't be correctness issue. `data_size` is only used to record the [bytes_read](https://github.com/pytorch/pytorch/blob/f4d7cdc5e63c786b1f6588eafa53bbc6d33c3826/torch/csrc/inductor/aoti_runtime/model.h#L217). This PR will improve the performance on CPU for 4 models in HF, 7 models in TIMM and 1 model in Torchbench. For the unit test, I add a bias value the original `data_size` of which is not divisible by the alignment to test the correctness: ``` constants_info_[0].dtype = static_cast<int32_t>(at::kFloat); constants_info_[0].data_size = 64; # was 40 before this PR constants_info_[0].shape = {10}; constants_info_[1].dtype = static_cast<int32_t>(at::kFloat); ...... ``` Pull Request resolved: pytorch#127610 Approved by: https://github.com/jgong5, https://github.com/desertfire
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):
GPU copies the constant blob to aligned memory (RAII_cudaMalloc, 64-alignment) while CPU doesn't have this copy procedure for constant blob, which may result in sub-optimal performance when we want to directly use the constant blob buffer in the computation (for example when these constant blobs are the weight tensor to the oneDNN primitive).
We set the alignment to the
constant.odirectly so that there's no need to copy the data to an aligned memory for CPU (when using--rename-section, the original section name would need to be specified for--set-section-alignment).cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @peterbell10 @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @aakhundov @ColinPeppler @amjames @desertfire @chauhang