SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Fix important models CI (#39576)

Browse Source 
* relax test boundaries and fix from config

* eager is always supported.
This commit is contained in:
Pablo Montalvo
2025-07-23 16:24:29 +02:00
committed by GitHub
parent 0fe03afeb8
commit ea56eb6bed
4 changed files with 7 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
tests/causal_lm_tester.py
View File

@@ -451,5 +451,4 @@ class CausalLMModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterM
logits = outputs.hidden_states[-1] logits = outputs.hidden_states[-1]
logits_fa = outputs_fa.hidden_states[-1] logits_fa = outputs_fa.hidden_states[-1]
torch.testing.assert_close(logits_fa, logits, atol=3e-2, rtol=3e-2)
assert torch.allclose(logits_fa, logits, atol=2e-3)

2
tests/generation/test_utils.py
View File

@@ -2309,7 +2309,7 @@ class GenerationTesterMixin:
set_model_tester_for_less_flaky_test(self) set_model_tester_for_less_flaky_test(self)
for model_class in self.all_generative_model_classes: for model_class in self.all_generative_model_classes:
if not getattr(model_class, support_flag[attn_implementation]): if attn_implementation != "eager" and not getattr(model_class, support_flag[attn_implementation]):
self.skipTest(f"{model_class.__name__} does not support `attn_implementation={attn_implementation}`") self.skipTest(f"{model_class.__name__} does not support `attn_implementation={attn_implementation}`")
config, original_inputs_dict = self.prepare_config_and_inputs_for_generate() config, original_inputs_dict = self.prepare_config_and_inputs_for_generate()

2
tests/models/kyutai_speech_to_text/test_modeling_kyutai_speech_to_text.py
View File

@@ -552,7 +552,7 @@ class KyutaiSpeechToTextModelTest(ModelTesterMixin, GenerationTesterMixin, Pipel
} }
for model_class in self.all_generative_model_classes: for model_class in self.all_generative_model_classes:
if not getattr(model_class, support_flag[attn_implementation]): if attn_implementation != "eager" and not getattr(model_class, support_flag[attn_implementation]):
self.skipTest(f"{model_class.__name__} does not support `attn_implementation={attn_implementation}`") self.skipTest(f"{model_class.__name__} does not support `attn_implementation={attn_implementation}`")
config, original_inputs_dict = self.prepare_config_and_inputs_for_generate() config, original_inputs_dict = self.prepare_config_and_inputs_for_generate()

6
tests/test_modeling_common.py
View File

@@ -4107,7 +4107,9 @@ class ModelTesterMixin:
config, _ = self.model_tester.prepare_config_and_inputs_for_common() config, _ = self.model_tester.prepare_config_and_inputs_for_common()
cls = self._torch_compile_train_cls # e.g. LlamaFroCausalLM cls = self._torch_compile_train_cls # e.g. LlamaFroCausalLM
model = cls(config, attn_implementation="flash_attention_2").to(device=torch_device, dtype=torch_dtype) model = cls._from_config(config, attn_implementation="flash_attention_2").to(
device=torch_device, dtype=torch_dtype
)
inputs = { inputs = {
"input_ids": torch.randint(low=1, high=model.config.vocab_size, size=(2, 10), device=torch_device), "input_ids": torch.randint(low=1, high=model.config.vocab_size, size=(2, 10), device=torch_device),
@@ -4137,7 +4139,7 @@ class ModelTesterMixin:
} }
for model_class in self.all_generative_model_classes: for model_class in self.all_generative_model_classes:
if not getattr(model_class, support_flag[attn_implementation]): if attn_implementation != "eager" and not getattr(model_class, support_flag[attn_implementation]):
self.skipTest(f"{model_class.__name__} does not support {attn_implementation}") self.skipTest(f"{model_class.__name__} does not support {attn_implementation}")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()