Forward rank and world size info to Torchbench models when using dynamo runner #108438
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/108438
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 20aad90 with merge base 8851603 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
xmfan
force-pushed
the
xmfan/distributed_torchbench
branch
2 times, most recently
from
39b6fcc to
e5ba94d
Compare
xmfan
changed the title
Chillee
approved these changes
Sep 5, 2023
|
@pytorchbot label "topic: not user facing" |
|
Failures about unhandled extra_args should be resolved once the changes from torchbenchmark/util/extra_args.py https://github.com/pytorch/benchmark/pull/1867/files#diff-6ccf656b90a64ee9ee4e55aec794320710e717b65271baeae74e69940524bb6a land |
H-Huang
reviewed
Sep 5, 2023
| try: | ||
| with tqdm(desc="loading model"): | ||
| extra_args = [] | ||
| if hasattr(args, "rank") and hasattr(args, "world_size"): |
There was a problem hiding this comment.
I think we also need to update the hash with your torchbench changes on they land in order for the CI to pick it up
https://github.com/pytorch/pytorch/blob/main/.github/ci_commit_pins/torchbench.txt
facebook-github-bot
pushed a commit
to pytorch/benchmark
that referenced
this pull request
Sep 8, 2023
Summary: Adds simple_gpt + DTensor implemented in meta-pytorch/simple_gpt#7 to torchbench Tested via `python benchmarks/dynamo/torchbench.py -d cuda --output-directory=benchmark_logs --output=performance.csv --inference --performance --timing --print-memory --multiprocess --nothing --only simple_gpt`. Note: --nothing is used here to disable compile, since DTensor + compile isn't yet supported in main ``` dev,name,batch_size,speedup,abs_latency,compilation_latency,compression_ratio,eager_peak_mem,dynamo_peak_mem,calls_captured,unique_graphs,graph_breaks,unique_graph_breaks cuda,simple_gpt,1,0.966153,196.819773,-0.059319,1.000000,4.576880,4.576880,0,0,0,0 cuda,simple_gpt,1,0.967389,196.608152,-0.058833,1.000000,4.577404,4.577404,0,0,0,0 cuda,simple_gpt,1,0.973152,196.093583,-0.059316,1.000000,4.593133,4.593133,0,0,0,0 cuda,simple_gpt,1,0.973087,196.124046,-0.075580,1.000000,4.611483,4.611483,0,0,0,0 cuda,simple_gpt,1,0.967908,193.998484,-0.040192,1.000000,4.593133,4.593133,0,0,0,0 cuda,simple_gpt,1,0.968949,193.798088,-0.028878,1.000000,4.593133,4.593133,0,0,0,0 ``` 2 changes were required to the model: - decorate torch.no_grad() on the caches, previously this was done outside the model, the entire eval call was wrapped in a torch.no_grad() context. After using torchbench, I notice even with only inference mode, we don't disable gradient calculations - rank/world size, added support from torchbench side in pytorch/pytorch#108438 and updated model to fetch from the provided extra_args Pull Request resolved: #1867 Reviewed By: msaroufim Differential Revision: D49065244 Pulled By: xmfan fbshipit-source-id: d4709fa3997c6a25c75e87eff7c13492b370b1af
xmfan
force-pushed
the
xmfan/distributed_torchbench
branch
from
266257b to
20aad90
Compare
|
@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 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 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.
Adding support to pass rank and world_size to torchbench model, via its extra_args parameter: https://github.com/pytorch/benchmark/blob/main/torchbenchmark/util/model.py#L83C80-L83C90
This is used for models which distribute over multiple GPUs e.g. simple_gpt pytorch/benchmark#1867
Also add an option to skip multiprocess only gpu models
Testing via
python benchmarks/dynamo/torchbench.py -d cuda --output=benchmark_logs/performance.csv --inference --performance --timing --print-memory --multiprocess --only simple_gptcc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx @chenyang78 @aakhundov @kadeng @anijain2305