ci(speech): split L0_Unit_Tests_GPU_ASR into 5 parallel buckets#15654
Open
ci(speech): split L0_Unit_Tests_GPU_ASR into 5 parallel buckets#15654
Conversation
The monolithic job ran 40+ min total. Each bucket targets ≤10 min, distributed by observed wall-clock time from run 25112807335. Bucket mapping (approx times): 1 (~6.6m): confidence/ + fast decoding tests 2 (~9.8m): streaming/rnnt decoding + inference/ + k2/ + mixins/ 3 (~8.6m): numba/ + hybrid/interctc/local_attn models 4 (~10.6m): test_asr_multitask_model_bpe.py (single file, irreducible) 5 (~4.9m): rnnt encoder models + remaining small tests + utils/ Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> Signed-off-by: oliver könig <okoenig@nvidia.com>
… flaky batchnorm test ASR_1/ASR_2 were failing because the split scripts were missing TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1, which is required by all other model-loading test scripts (Core, Common, TTS). The .nemo checkpoints on /home/TestData use a legacy PyTorch storage format incompatible with weights_only=True in PyTorch 2.6+. ASR_5 was failing because test_from_batchnorm is order-dependent: the monolithic run consumed random state from ~1641 prior tests, whereas the isolated split starts fresh. Fix: add torch.manual_seed(0) for determinism and use atol=1e-5 to reflect the float32 rounding difference between fused (x*W+B) and standard ((x-mean)/std*w+b) formulations. Signed-off-by: Oliver Koenig <okoenig@nvidia.com> Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> Signed-off-by: oliver könig <okoenig@nvidia.com>
…ripts The original L0_Unit_Tests_GPU_ASR.sh launched with: python -c "from nemo.collections.asr.models import ASRModel" && ... which performs the required module initialization before running the test suite. All five splits were missing this prefix, causing failures in model-loading tests (kenlm, RNNT decoding) that depend on the import side-effects triggered by ASRModel initialization. Signed-off-by: Oliver Koenig <okoenig@nvidia.com> Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> Signed-off-by: oliver könig <okoenig@nvidia.com>
Signed-off-by: Oliver Koenig <okoenig@nvidia.com> Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> Signed-off-by: oliver könig <okoenig@nvidia.com>
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
Claude summary
L0_Unit_Tests_GPU_ASRwas a single job running the entiretests/collections/asr/directory sequentially, taking ~40 minutes wall-clock. This PR splits it into 5 parallel jobs by grouping test files based on observed durations from run 25112807335.L0_Unit_Tests_GPU_ASR_1confidence/+ fast decoding testsL0_Unit_Tests_GPU_ASR_2inference/+k2/+mixins/L0_Unit_Tests_GPU_ASR_3numba/+ hybrid/interctc/local-attn modelsL0_Unit_Tests_GPU_ASR_4test_asr_multitask_model_bpe.py(single file, irreducible)L0_Unit_Tests_GPU_ASR_5utils/Note: Bucket 4 (
test_asr_multitask_model_bpe.py) marginally exceeds the 10-minute target. Shrinking it further would require splitting within the test class.Each job timeout is set to 15 minutes (down from 60).