[tests] Parameterized test_eager_matches_sdpa_inference (#36650)

This commit is contained in:
Joao Gante
2025-03-14 14:41:27 +00:00
committed by GitHub
parent 9215cc62d4
commit 42ebb6c23e
16 changed files with 285 additions and 1900 deletions

View File

@@ -20,7 +20,6 @@ from typing import ClassVar
import torch
from datasets import load_dataset
from parameterized import parameterized
from tests.test_configuration_common import ConfigTester
from tests.test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
@@ -32,7 +31,6 @@ from transformers.models.colpali.modeling_colpali import ColPaliForRetrieval, Co
from transformers.models.colpali.processing_colpali import ColPaliProcessor
from transformers.testing_utils import (
require_torch,
require_torch_sdpa,
require_vision,
slow,
torch_device,
@@ -271,14 +269,6 @@ class ColPaliForRetrievalModelTest(ModelTesterMixin, unittest.TestCase):
def test_training_gradient_checkpointing_use_reentrant_false(self):
pass
@require_torch_sdpa
@slow
@parameterized.expand([("float16",), ("bfloat16",), ("float32",)])
def test_eager_matches_sdpa_inference(self, torch_dtype: str):
self.skipTest(
"Due to custom causal mask, there is a slightly too big difference between eager and sdpa in bfloat16."
)
@unittest.skip(
reason="From PaliGemma: Some undefined behavior encountered with test versions of this model. Skip for now."
)