diff --git a/pyproject.toml b/pyproject.toml
index 0950b2682..19cdc0a26 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -65,7 +65,8 @@ dev = [
     "orjson>=3.11.7", # to make v1/embeddings API fast
     "voxcpm>=2.0.2", # for VoxCPM2 path-locating in tests
     "torchdiffeq>=0.2.5", # for covo_audio Test
-    "FlagEmbedding" # for hunyuan_image3 tests
+    "FlagEmbedding", # for hunyuan_image3 tests
+    "kernels<0.15.0" # for accuracy tests comparison with diffusers output (0.15.0 incompatible with diffusers 0.38.0)
 ]
 
 demo = [
diff --git a/tests/e2e/accuracy/test_diffusers_backend_similarity.py b/tests/e2e/accuracy/test_diffusers_backend_similarity.py
index 26b39ea78..ad4dd5726 100644
--- a/tests/e2e/accuracy/test_diffusers_backend_similarity.py
+++ b/tests/e2e/accuracy/test_diffusers_backend_similarity.py
@@ -218,7 +218,7 @@ def _run_diffusers_qwen_image(*, model: str, output_path: Path) -> tuple[Image.I
             torch_dtype=torch.bfloat16,
             trust_remote_code=True,
         ).to("cuda")
-
+        pipe.transformer.set_attention_backend("_flash_3_hub")
         _diffusers_dummy_run(pipe)
 
         generator = torch.Generator(device="cuda").manual_seed(SEED)
diff --git a/tests/e2e/accuracy/test_qwen_image.py b/tests/e2e/accuracy/test_qwen_image.py
index c9a09e6de..2b513ecda 100644
--- a/tests/e2e/accuracy/test_qwen_image.py
+++ b/tests/e2e/accuracy/test_qwen_image.py
@@ -106,6 +106,7 @@ def _run_diffusers_qwen_image(*, model: str, output_path: Path) -> Image.Image:
             trust_remote_code=True,
             local_files_only=_local_files_only(model),
         ).to("cuda")
+        pipe.transformer.set_attention_backend("_flash_3_hub")
         generator = torch.Generator(device="cuda").manual_seed(SEED)
         result = pipe(  # pyright: ignore[reportCallIssue]
             prompt=PROMPT,
@@ -167,6 +168,7 @@ def _run_diffusers_qwen_image_2512(*, model: str, output_path: Path) -> Image.Im
             trust_remote_code=True,
             local_files_only=_local_files_only(model),
         ).to("cuda")
+        pipe.transformer.set_attention_backend("_flash_3_hub")
         generator = torch.Generator(device="cuda").manual_seed(SEED_2512)
         result = pipe(  # pyright: ignore[reportCallIssue]
             prompt=PROMPT_2512,