Add PyTorchAlpaka mini-batching support - #50794
Conversation
|
type ngt |
|
cms-bot internal usage |
|
-code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-50794/49105
Code check has found code style and quality issues which could be resolved by applying following patch(s)
|
|
+code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-50794/49114
|
|
Pull request #50794 was updated. |
86423c6 to
bfe41a0
Compare
|
-code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-50794/49118
Code check has found code style and quality issues which could be resolved by applying following patch(s)
|
bfe41a0 to
6999803
Compare
|
+code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-50794/49119
|
2aa7a7f to
43bafa2
Compare
|
+code-checks Logs: https://cmssdt.cern.ch/SDT/code-checks/cms-sw-PR-50794/49166
|
|
@fwyzard , thanks for the review. I addressed the other comments, the only ones left are:
Since PR #50790 is also modifying the |
|
It's fine by me if those two points are addressed in a separate PR. |
|
please test |
|
+1 Size: This PR adds an extra 24KB to repository Comparison SummarySummary:
|
@fwyzard do you mind signing again? |
|
+heterogeneous |
|
+ml |
|
This pull request is fully signed and it will be integrated in one of the next master IBs (tests are also fine). This pull request will now be reviewed by the release team before it's merged. @ftenchini, @sextonkennedy, @mandrenguyen (and backports should be raised in the release meeting by the corresponding L2) |
|
+1 |
This PR results from a collaboration with @valsdav. Thanks also to @Electricks94 for his help.
PR description:
This PR introduces native support for splitting a single SoA into multiple logical batches.
TensorCollectionnow supports(batch_size, total_size)construction and computes per-batch offsets internally.Validation
Validation is performed via dedicated mini-batch inference producers:
SimpleNetMiniBatchTinyResNetMiniBatchThe latter reproduces the same inference as the non-batched implementations using the new batching API.
InspectionSinkhas been extended to compare outputs between the standard and the batched inference.A relative difference check is enforced:
|ref - batched| / ref < 1e-5It can be run by:
Notes / limitations