Delete deletion-eligible test_kahan_summation_cpu - #554
Closed
twmoveon wants to merge 4 commits into
Closed
Conversation
…#551) Summary: Oncall test cleanup for the `stl_multimodal` rotation. TestX flags these four tests in `tests/models/albef/test_albef.py` as eligible for deletion — all sit in `DISABLED_FAILING` trunk state with open `FAILURE` issues, well past the 60-day zero-value threshold: - `test_albef_image_embeddings` (`844425197297248`, `844425197101365`) - `test_albef_image_embeddings_momentum` (`562950220612232`, `844425197101313`) - `test_albef_multimodal_embeddings` (`844425197297177`, `844425197101408`) - `test_albef_multimodal_embeddings_momentum` (`562950220612287`, `281475243658978`) Each asserted a hard-coded 2x2x3 activation tensor to `atol=1e-4` against a seeded forward pass — golden values that drift whenever RNG or kernel behavior changes underneath them. Re-baselining numbers nobody has validated since they broke would restore a green signal without restoring any real assurance, so they are deleted instead. The `albef_model_output` fixture and the surrounding tests are untouched: `test_albef_text_embeddings` and `test_albef_text_embeddings_momentum` still exercise the same forward pass, and `test_similarity`, `test_neg_embeddings`, `test_dequeue_and_enqueue`, `test_momentum_update`, and `test_copy_params_momentum_models` still cover `ALBEFModelWithSimilarity`. Differential Revision: D114397069
…rch#552) Summary: Oncall test cleanup for the `stl_multimodal` rotation. TestX flags `TestALBEFVisionEncoder.test_vision_transformer` (test IDs `281475243855620` and `281475243658858`) as eligible for deletion — it sits in `DISABLED_FAILING` trunk state with an open `FAILURE` issue, well past the 60-day zero-value threshold. It was another seeded golden-value test, asserting a hard-coded 1x2x3 tensor to `atol=1e-4`, so it is deleted rather than re-baselined. `assert_expected` and `Tensor` were only used by the deleted test, so both imports are dropped. `set_rng_seed` stays — it is still called at class scope. The four negative tests remain and are the more durable half of this file: `test_invalid_input_length`, `test_invalid_image_channel_dim`, `test_invalid_image_height`, and `test_invalid_image_width` still pin `ALBEFVisionEncoder` input validation. Differential Revision: D114397059
…ch#553) Summary: Oncall test cleanup for the `stl_multimodal` rotation. TestX flags these three tests in `tests/models/test_omnivore.py` as eligible for deletion — all sit in `DISABLED_FAILING` trunk state with open `FAILURE` issues, well past the 60-day zero-value threshold: - `test_omnivore_swin_t_forward` (`562950220612371`, `844425197101258`) - `test_omnivore_swin_s_forward` (`562950220612245`, `281475243658871`) - `test_omnivore_swin_b_forward` (`281475243855615`, `844425197101257`) Each asserted hard-coded output magnitudes (for example `184.01417`, `239.73104`, `278.06488`) from a seeded forward pass across the image, rgbd, and video heads. These are golden values with no independently derived expectation, so they are deleted rather than re-baselined. The `omnivore_swin_s_model` and `omnivore_swin_b_model` fixtures existed only to serve the deleted tests and are removed with them. `assert_expected` was likewise only used by the deleted tests, so that import is dropped; `set_rng_seed` stays because the `device` fixture still calls it. `test_omnivore_forward_wrong_input_type` remains and still covers the `input_type` validation path via `omnivore_swin_t_model`. Differential Revision: D114397066
Summary: Oncall test cleanup for the `stl_multimodal` rotation. TestX flags `TestAnyPrecisionOptimizer.test_kahan_summation_cpu` (test IDs `281475243659032` and `281475243855546`) as eligible for deletion — it sits in `DISABLED_FAILING` trunk state with an open `FAILURE` issue, well past the 60-day zero-value threshold. The test hard-coded four bfloat16 compensation-buffer tensors captured from a seeded two-step optimizer run and compared them at `atol=1e-4`. At bfloat16 precision those buffers hold values near `1e-5`, so the assertion tolerance is wider than the quantities being checked and the test was pinning RNG output rather than Kahan-summation correctness. Deleting is the right call here rather than re-recording the golden buffers. `test_kahan_summation_gpu` is left in place: it is `gpu_test()`-gated, was not flagged, and still covers the same `use_kahan_summation=True` path. `_test_adam_equivalence`, `test_adam_equivalence_cpu`, `_test_bfloat16_states`, and `test_bfloat16_states_cpu` continue to cover `AnyPrecisionAdamW` on CPU. Differential Revision: D114397064
meta-cla
Bot
added
the
CLA Signed
|
@twmoveon has exported this pull request. If you are a Meta employee, you can view the originating Diff in D114397064. |
|
This pull request has been merged in 77c3110. |
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.
Summary:
Oncall test cleanup for the
stl_multimodalrotation. TestX flagsTestAnyPrecisionOptimizer.test_kahan_summation_cpu(test IDs281475243659032and281475243855546) as eligible for deletion — it sits inDISABLED_FAILINGtrunk state with an openFAILUREissue, well past the60-day zero-value threshold.
The test hard-coded four bfloat16 compensation-buffer tensors captured from a
seeded two-step optimizer run and compared them at
atol=1e-4. At bfloat16precision those buffers hold values near
1e-5, so the assertion tolerance iswider than the quantities being checked and the test was pinning RNG output
rather than Kahan-summation correctness. Deleting is the right call here
rather than re-recording the golden buffers.
test_kahan_summation_gpuis left in place: it isgpu_test()-gated, was notflagged, and still covers the same
use_kahan_summation=Truepath._test_adam_equivalence,test_adam_equivalence_cpu,_test_bfloat16_states,and
test_bfloat16_states_cpucontinue to coverAnyPrecisionAdamWon CPU.Differential Revision: D114397064