[tests] expand flex-attn test for vision models (#38434)

* expand the test for VLMs

* typo

* mark models `supports_flex` + expand test for additional kwargs

* flex attn for refactored vision models

* fix copies

* fix

* unskip

* style

* address comments
This commit is contained in:
Raushan Turganbay
2025-06-03 09:40:44 +02:00
committed by GitHub
parent de4cf5a38e
commit bf68dd9e6e
45 changed files with 429 additions and 195 deletions

View File

@@ -351,12 +351,6 @@ class Gemma3Vision2TextModelTest(ModelTesterMixin, GenerationTesterMixin, unitte
def test_initialization(self):
pass
@unittest.skip(
reason="Siglip has no FLEX attention, and we don't have a proper way to set/test attn in VLMs. TODO @raushan"
)
def test_flex_attention_with_grads(self):
pass
def test_automodelforcausallm(self):
"""
Regression test for #36741/#36917 -- make sure `AutoModelForCausalLM` works with a Gemma3 config, i.e. that