Use different conv layout optimization heuristics for inference #114600
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/114600
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (5 Unrelated Failures)As of commit 70140dc with merge base 56a95af ( FLAKY - The following jobs failed but were likely due to flakiness present on trunk:
BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
…rence" cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler [ghstack-poisoned]
…rence" cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler [ghstack-poisoned]
…rence" While many models regress in training when converted to channels last, in inference the results are quite different. Almost all of the models experienced a speedup when converted to channels last. There were a few big regressions in torchbench - `timm_regnet` from `1.4343 → 1.0573` and `timm_resnet` from `1.7484 → 1.2868`. I used a modified script of the operator benchmarks [here](https://gist.github.com/eellison/e11dc645412f52e8b45fb26ba6f9f6a1) to measure the average speedup of convolutions across all of the input shapes found in torchbench according to the existing classifications that shunting314 used - grouped convs, small channel convs, convolution with larger in-channel than out-channel. Only grouped convolutions benchmarked as a slowdown in inference. I updated the inference heuristic to multiply the flops of each conv with its predicted speedup/slowdown in channels last. With this heuristic the two previously regressing models no longer regress. Speeds up inference for torchbench ~8% and timm ~6%. The motivating model here was SDXL which now hits channels last and improves 10%. There were some models that were sped up in training when forcing channels last (along with a number of regressions), it's possible there is some speedup there to be had. We could also have more granular classification/predictions. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler [ghstack-poisoned]
eellison
requested review from
Chillee,
jansel and
shunting314
and removed request for
Chillee,
jansel and
shunting314
…rence" While many models regress in training when converted to channels last, in inference the results are quite different. Almost all of the models experienced a speedup when converted to channels last. There were a few big regressions in torchbench - `timm_regnet` from `1.4343 → 1.0573` and `timm_resnet` from `1.7484 → 1.2868`. I used a modified script of the operator benchmarks [here](https://gist.github.com/eellison/e11dc645412f52e8b45fb26ba6f9f6a1) to measure the average speedup of convolutions across all of the input shapes found in torchbench according to the existing classifications that shunting314 used - grouped convs, small channel convs, convolution with larger in-channel than out-channel. Only grouped convolutions benchmarked as a slowdown in inference. I updated the inference heuristic to multiply the flops of each conv with its predicted speedup/slowdown in channels last. With this heuristic the two previously regressing models no longer regress. Speeds up inference for torchbench ~8% and timm ~6%. The motivating model here was SDXL which now hits channels last and improves 10%. There were some models that were sped up in training when forcing channels last (along with a number of regressions), it's possible there is some speedup there to be had. We could also have more granular classification/predictions. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler [ghstack-poisoned]
eellison
requested review from
Chillee,
jansel,
jgong5 and
shunting314
and removed request for
jgong5 and
shunting314
…rence" While many models regress in training when converted to channels last, in inference the results are quite different. Almost all of the models experienced a speedup when converted to channels last. There were a few big regressions in torchbench - `timm_regnet` from `1.4343 → 1.0573` and `timm_resnet` from `1.7484 → 1.2868`. I used a modified script of the operator benchmarks [here](https://gist.github.com/eellison/e11dc645412f52e8b45fb26ba6f9f6a1) to measure the average speedup of convolutions across all of the input shapes found in torchbench according to the existing classifications that shunting314 used - grouped convs, small channel convs, convolution with larger in-channel than out-channel. Only grouped convolutions benchmarked as a slowdown in inference. I updated the inference heuristic to multiply the flops of each conv with its predicted speedup/slowdown in channels last. With this heuristic the two previously regressing models no longer regress. Speeds up inference for torchbench ~8% and timm ~6%. The motivating model here was SDXL which now hits channels last and improves 10%. There were some models that were sped up in training when forcing channels last (along with a number of regressions), it's possible there is some speedup there to be had. We could also have more granular classification/predictions. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler [ghstack-poisoned]
jansel
requested changes
Nov 28, 2023
| "test_kwargs_dynamic_shapes": TestFailure(("cpu",)), | ||
| # calling div on only symint args | ||
| "test_AllenaiLongformerBase_repro_dynamic_shapes": TestFailure(("cpu", "cuda")), | ||
| "test_conv_inference_heuristics_dynamic_shapes": TestFailure("cuda"), |
There was a problem hiding this comment.
We turn off channels last optimization for dynamic shapes. That is blocked on persistent reduction perf not being enabled for dynamic shapes. CC @peterbell10 @shunting314 was one of you working on persistent reduction with dynamic shapes ?
…rence" While many models regress in training when converted to channels last, in inference the results are quite different. Almost all of the models experienced a speedup when converted to channels last. There were a few big regressions in torchbench - `timm_regnet` from `1.4343 → 1.0573` and `timm_resnet` from `1.7484 → 1.2868`. I used a modified script of the operator benchmarks [here](https://gist.github.com/eellison/e11dc645412f52e8b45fb26ba6f9f6a1) to measure the average speedup of convolutions across all of the input shapes found in torchbench according to the existing classifications that shunting314 used - grouped convs, small channel convs, convolution with larger in-channel than out-channel. Only grouped convolutions benchmarked as a slowdown in inference. I updated the inference heuristic to multiply the flops of each conv with its predicted speedup/slowdown in channels last. With this heuristic the two previously regressing models no longer regress. Speeds up inference for torchbench ~8% and timm ~6%. The motivating model here was SDXL which now hits channels last and improves 10%. There were some models that were sped up in training when forcing channels last (along with a number of regressions). It's possible there is some speedup in training to be had with additional heuristics. We could also have more granular classification/predictions which might benefit both training and inference. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler [ghstack-poisoned]
jansel
approved these changes
Nov 28, 2023
shunting314
approved these changes
Nov 28, 2023
Is this speedup between using the new inference specific heuristics v.s. using the previous heuristics for both training and inference? |
|
@shunting314 the speedup is for inference relative to main branch. |
|
@pytorchbot merge |
Merge failedReason: This PR needs a If not, please add the To add a label, you can comment to pytorchbot, for example For more information, see Details for Dev Infra teamRaised by workflow job |
|
@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 |
|
Is there any reason as to why this speedup/slowdown occurs and also why the behavior changes during training and inference> |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
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):
While many models regress in training when converted to channels last, in inference the results are quite different. Almost all of the models experienced a speedup when converted to channels last. There were a few big regressions in torchbench -
timm_regnetfrom1.4343 → 1.0573andtimm_resnetfrom1.7484 → 1.2868.I used a modified script of the operator benchmarks here to measure the average speedup of convolutions across all of the input shapes found in torchbench according to the existing classifications that @shunting314 used - grouped convs, small channel convs, convolution with larger in-channel than out-channel. Only grouped convolutions benchmarked as a slowdown in inference.
I updated the inference heuristic to multiply the flops of each conv with its predicted speedup/slowdown in channels last. With this heuristic the two previously regressing models no longer regress.
Speeds up inference for torchbench ~8% and timm ~6%. The motivating model here was SDXL which now hits channels last and improves 10%.
There were some models that were sped up in training when forcing channels last (along with a number of regressions). It's possible there is some speedup in training to be had with additional heuristics. We could also have more granular classification/predictions which might benefit both training and inference.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @peterbell10 @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @aakhundov @ColinPeppler