[bugfix]: restore TF32 flags in transformer test teardown on torch 2.12

torch 2.12 computes torch.backends.cuda.matmul.allow_tf32 on cuBLASModule
instead of storing it, so monkeypatch.setattr cannot undo it and pytest's
teardown raised AttributeError for every Qwen-Image-2.1 attention and Hunyuan
batch test. A shared tf32_disabled fixture saves and restores the legacy flags.

fastvideo/tests/conftest.py

@@ -45,3 +45,21 @@ def env_overrides():
     """
     with contextlib.ExitStack() as stack:
         yield stack
+
+
+@pytest.fixture
+def tf32_disabled():
+    """Disable TF32 for CUDA matmuls and cuDNN during the test, then restore the previous flags.
+
+    On torch 2.12 ``torch.backends.cuda.matmul.allow_tf32`` is computed by ``cuBLASModule`` rather
+    than stored on it, so ``monkeypatch.setattr`` cannot undo it. Restore through the same legacy
+    flags: mixing them with the per-backend ``fp32_precision`` API makes later ``allow_tf32`` reads raise.
+    """
+    matmul, cudnn = torch.backends.cuda.matmul, torch.backends.cudnn
+    previous = (matmul.allow_tf32, cudnn.allow_tf32)
+    matmul.allow_tf32 = False
+    cudnn.allow_tf32 = False
+    try:
+        yield
+    finally:
+        matmul.allow_tf32, cudnn.allow_tf32 = previous

fastvideo/tests/transformers/test_hunyuan_batch.py

@@ -25,7 +25,7 @@ pytestmark = pytest.mark.skipif(not torch.cuda.is_available(), reason="Requires
 
 
 @pytest.fixture
-def hunyuan_setup(monkeypatch, request):
+def hunyuan_setup(request):
     # Environment writes go through the registry helpers (docs/contributing/env_vars.md).
     with contextlib.ExitStack() as stack:
         # Respect the port reserved by the CI runner when sharing a GPU host.
@@ -37,8 +37,7 @@ def hunyuan_setup(monkeypatch, request):
         if "MASTER_ADDR" not in os.environ:
             stack.enter_context(envs.override_external("MASTER_ADDR", "127.0.0.1"))
         stack.enter_context(envs.FASTVIDEO_ATTENTION_BACKEND.override("TORCH_SDPA"))
-        monkeypatch.setattr(torch.backends.cuda.matmul, "allow_tf32", False)
-        monkeypatch.setattr(torch.backends.cudnn, "allow_tf32", False)
+        request.getfixturevalue("tf32_disabled")
         request.getfixturevalue("distributed_setup")
         yield
 

fastvideo/tests/transformers/test_qwen_image21_attention.py

@@ -24,12 +24,11 @@ from fastvideo.layers.linear import ReplicatedLinear
 from fastvideo.platforms import AttentionBackendEnum
 
 @pytest.fixture
-def qwen_setup(env_overrides, monkeypatch, request):
+def qwen_setup(env_overrides, request):
     if not torch.cuda.is_available():
         pytest.skip("Requires one CUDA GPU")
     env_overrides.enter_context(envs.FASTVIDEO_ATTENTION_BACKEND.override("TORCH_SDPA"))
-    monkeypatch.setattr(torch.backends.cuda.matmul, "allow_tf32", False)
-    monkeypatch.setattr(torch.backends.cudnn, "allow_tf32", False)
+    request.getfixturevalue("tf32_disabled")
     request.getfixturevalue("distributed_setup")
     with torch.no_grad(), sdpa_kernel(SDPBackend.MATH):
         yield torch.device("cuda")