Tests: upgrade test_eager_matches_sdpa_generate
#34386
Merged
Conversation
This file contains 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
gante
commented
Oct 24, 2024
| @@ -2046,6 +2047,81 @@ def test_inherits_generation_mixin(self): | |||
| for model_class in self.all_generative_model_classes: | |||
| self.assertTrue("GenerationMixin" in str(model_class.__bases__)) | |||
|
|
|||
| @require_torch_sdpa | |||
There was a problem hiding this comment.
(this is mostly copy-paste, going to comment the sections that are changed)
gante
commented
Oct 24, 2024
Comment on lines
+2059
to
+2067
| config, original_inputs_dict = self.prepare_config_and_inputs_for_generate() | ||
| inputs_dict = {} | ||
| for input_name, input_data in original_inputs_dict.items(): | ||
| if isinstance(input_data, torch.Tensor) and input_data.dtype in [torch.float32, torch.bfloat16]: | ||
| inputs_dict[input_name] = input_data.to(torch.float16) | ||
| else: | ||
| inputs_dict[input_name] = input_data | ||
| main_input = inputs_dict[model_class.main_input_name] |
There was a problem hiding this comment.
Uses self.prepare_config_and_inputs_for_generate() instead, which enables us to pass a dictionary of inputs to generate (better input control than simply using inputs_dict[model_class.main_input_name])
gante
commented
Oct 24, 2024
tests/generation/test_utils.py
Outdated
Comment on lines
2090
to
2098
| # Just test that a large cache works as expected | ||
| generate_kwargs = { | ||
| "max_new_tokens": max_new_tokens, | ||
| "do_sample": False, | ||
| "return_dict_in_generate": True, | ||
| "output_scores": True, | ||
| } | ||
| res_eager = model_eager.generate(**inputs_dict, **generate_kwargs) | ||
| res_sdpa = model_sdpa.generate(**inputs_dict, **generate_kwargs) |
There was a problem hiding this comment.
Uses dictionaries -> more compact
gante
commented
Oct 24, 2024
Comment on lines
+2100
to
+2127
| # Eager and SDPA are very similar, but not exactly the same. Because we are using random models, this | ||
| # test would be flaky if we only checked the sequences. Two situations in which this test passes: | ||
| # 1. The sequences are the same | ||
| # 2. The sequences are different, but the scores up until the first mismatch are nearly identical | ||
| output_matches = res_eager.sequences == res_sdpa.sequences | ||
| has_matching_outputs = output_matches.all() | ||
| has_matching_scores = None | ||
| if not has_matching_outputs: | ||
| input_length = main_input.shape[1] | ||
| for batch_idx in range(res_eager.sequences.shape[0]): | ||
| batch_matches = output_matches[batch_idx] | ||
| if batch_matches.all(): | ||
| continue | ||
| first_mismatch_idx = batch_matches.int().argmin() # gets the index of the first False | ||
| first_mismatch_idx -= input_length # scores doesn't include data regarding input tokens | ||
| sdpa_first_mismatch_scores = res_sdpa.scores[first_mismatch_idx][batch_idx] | ||
| eager_first_mismatch_scores = res_eager.scores[first_mismatch_idx][batch_idx] | ||
| has_matching_scores = torch.allclose( | ||
| sdpa_first_mismatch_scores, eager_first_mismatch_scores, rtol=1e-3, atol=1e-3 | ||
| ) | ||
| if not has_matching_scores: | ||
| break |
There was a problem hiding this comment.
flakiness handling as explained in the PR header
ArthurZucker
approved these changes
Oct 24, 2024
tests/generation/test_utils.py
Outdated
Comment on lines
2077
to
2101
| model_sdpa = model_class.from_pretrained( | ||
| tmpdirname, | ||
| torch_dtype=torch.float16, | ||
| low_cpu_mem_usage=True, | ||
| ).to(torch_device) | ||
|
|
||
| model_eager = model_class.from_pretrained( | ||
| tmpdirname, | ||
| torch_dtype=torch.float16, | ||
| low_cpu_mem_usage=True, | ||
| attn_implementation="eager", | ||
| ).to(torch_device) |
There was a problem hiding this comment.
with #34282 we won't have to init 2 models.
Also this is memory hungry
There was a problem hiding this comment.
TBH we can do it one model at a time, going to change the test.
After Flex attention becomes the norm, we probably won't need this test
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
ydshieh
approved these changes
Oct 24, 2024
gante
force-pushed
the
test_eager_matches_sdpa_generate
branch
from
6ffa18d to
415d009
Compare
Merged
3 tasks
BernardZach
pushed a commit
to BernardZach/transformers
that referenced
this pull request
Dec 5, 2024
ylacombe
added a commit
that referenced
this pull request
Dec 10, 2024
* Support BatchNorm in Hubert pos_conv_emb as in fairseq * Correct the new defaults (#34377) * Correct the new defaults * CIs * add check * Update utils.py * Update utils.py * Add the max_length in generate test checking shape without passing length * style * CIs * fix fx CI issue * [auto. ping] Avoid sending empty info + add more team members (#34383) * update * update --------- Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> * Fix glm (#34388) * Fix duplicated * fix import * Use non nested images and batched text Idefics2/3 (#34222) * add support for non nested images and add tests * add tests error scenario * fix style * added single and no image to error tests * Fix onnx non-expotable inplace aten op (#34376) * fix onnx non-expotable inplace op * mistral, qwen2, qwen2_vl, starcoder2 * fixup copies * Fix right padding in LLaVA models (#34305) * fix right pad llavas * device mismatch * no filter (#34391) * no filter * no filter * no filter --------- Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> * SynthID: better example (#34372) * better example * Update src/transformers/generation/configuration_utils.py * Update src/transformers/generation/logits_process.py * nits * Tests: upgrade `test_eager_matches_sdpa_generate` (#34386) * Fix bnb training test failure (#34414) * Fix bnb training test: compatibility with OPTSdpaAttention * Avoid check expected exception when it is on CUDA (#34408) * update * update --------- Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> * Fix typos in agents_advanced.md (#34405) * [docs] Cache implementations (#34325) cache * [run-slow] hubert * Support BatchNorm in Hubert pos_conv_emb as in fairseq Add conversion integration test, and make batchnorm explicit variable * Support BatchNorm in Hubert pos_conv_emb as in fairseq fix make fixup styling changes * [run-slow] hubert * Support BatchNorm in Hubert pos_conv_emb as in fairseq * [run-slow] hubert * Support BatchNorm in Hubert pos_conv_emb as in fairseq Add conversion integration test, and make batchnorm explicit variable * Support BatchNorm in Hubert pos_conv_emb as in fairseq fix make fixup styling changes * [run-slow] hubert * [run-slow] hubert --------- Co-authored-by: Cyril Vallez <cyril.vallez@huggingface.co> Co-authored-by: Yih-Dar <2521628+ydshieh@users.noreply.github.com> Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> Co-authored-by: Yoni Gozlan <74535834+yonigozlan@users.noreply.github.com> Co-authored-by: Ilyas Moutawwakil <57442720+IlyasMoutawwakil@users.noreply.github.com> Co-authored-by: Raushan Turganbay <raushan@huggingface.co> Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com> Co-authored-by: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com> Co-authored-by: Rudy Delouya <rudy.delouya@gmail.com> Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com> Co-authored-by: Yoach Lacombe <52246514+ylacombe@users.noreply.github.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
What does this PR do?
test_eager_matches_sdpa_generatehas failed in our new failure reporting system (here, cc @ydshieh )Having a look at the test, the cause for flakiness was clear: we are using random models with
generate, and tiny perturbations can result in a different sampled token, causing generation to go in a different direction and ultimately failing the check (eager generate == sdpa generate).This PR:
GenerationTesterMixin, as it callsgenerateis_flaky().The following test commands were run:
RUN_SLOW=1 py.test tests/models/ -k test_eager_matches_sdpa_generate✅RUN_SLOW=1 py.test tests/models/gpt2/test_modeling_gpt2.py::GPT2ModelTest::test_eager_matches_sdpa_generate --flake-finder --flake-runs 500✅