Reset Dynamo Before Export for all XNNPACK Tests (#9561)

https://github.com/pytorch/executorch/actions/runs/14047575373/job/39331644423

There seems to be some CI issues with:
```
torch._dynamo.exc.FailOnRecompileLimitHit: recompile_limit reached with one_graph=True. Excessive recompilations can degrade performance due to the compilation overhead of each recompilation. To monitor recompilations, enable TORCH_LOGS=recompiles. If recompilations are expected, consider increasing
```

To help resolve this we reset dynamo at setup for all unittests. Let's
see if this helps

Author: mcr229

backends/xnnpack/test/models/deeplab_v3.py

@@ -23,6 +23,9 @@ def forward(self, *args):
 
 
 class TestDeepLabV3(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     dl3 = DL3Wrapper()
     dl3 = dl3.eval()
     model_inputs = (torch.randn(1, 3, 224, 224),)

backends/xnnpack/test/models/edsr.py

@@ -14,6 +14,9 @@
 
 
 class TestEDSR(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     edsr = edsr_r16f64(2, False).eval()  # noqa
     model_inputs = (torch.randn(1, 3, 224, 224),)
 

backends/xnnpack/test/models/emformer_rnnt.py

@@ -13,6 +13,9 @@
 
 
 class TestEmformerModel(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class EmformerRnnt(torch.nn.Module):
         def __init__(self):
             super().__init__()

backends/xnnpack/test/models/inception_v3.py

@@ -13,6 +13,9 @@
 
 
 class TestInceptionV3(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     ic3 = models.inception_v3(weights="IMAGENET1K_V1").eval()  # noqa
     model_inputs = (torch.randn(1, 3, 224, 224),)
 

backends/xnnpack/test/models/inception_v4.py

@@ -12,6 +12,9 @@
 
 
 class TestInceptionV4(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     ic4 = inception_v4(pretrained=False).eval()
     model_inputs = (torch.randn(3, 299, 299).unsqueeze(0),)
 

backends/xnnpack/test/models/llama2_et_example.py

@@ -13,6 +13,9 @@
 
 
 class TestLlama2ETExample(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     def test_f32(self):
         self._test()
 

backends/xnnpack/test/models/mobilebert.py

@@ -12,6 +12,9 @@
 
 
 class TestMobilebert(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     # pyre-ignore
     mobilebert = MobileBertModel(MobileBertConfig()).eval()
     example_inputs = (torch.tensor([[101, 7592, 1010, 2026, 3899, 2003, 10140, 102]]),)

backends/xnnpack/test/models/mobilenet_v2.py

@@ -14,6 +14,9 @@
 
 
 class TestMobileNetV2(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     mv2 = models.mobilenetv2.mobilenet_v2(weights=MobileNet_V2_Weights)
     mv2 = mv2.eval()
     model_inputs = (torch.randn(1, 3, 224, 224),)

backends/xnnpack/test/models/mobilenet_v3.py

@@ -13,6 +13,9 @@
 
 
 class TestMobileNetV3(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     mv3 = models.mobilenetv3.mobilenet_v3_small(pretrained=True)
     mv3 = mv3.eval()
     model_inputs = (torch.randn(1, 3, 224, 224),)

backends/xnnpack/test/models/resnet.py

@@ -13,6 +13,9 @@
 
 
 class TestResNet18(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     inputs = (torch.randn(1, 3, 224, 224),)
     dynamic_shapes = (
         {

backends/xnnpack/test/models/torchvision_vit.py

@@ -12,6 +12,9 @@
 
 
 class TestViT(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     vit = models.vision_transformer.vit_b_16(weights="IMAGENET1K_V1")
     vit = vit.eval()
     model_inputs = (torch.randn(1, 3, 224, 224),)

backends/xnnpack/test/models/very_big_model.py

@@ -11,6 +11,9 @@
 
 
 class TestVeryBigModel(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class BigModel(torch.nn.Module):
         def __init__(self):
             super().__init__()

backends/xnnpack/test/models/w2l.py

@@ -12,6 +12,9 @@
 
 
 class TestW2L(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     batch_size = 10
     input_frames = 700
     vocab_size = 4096

backends/xnnpack/test/ops/test_abs.py

@@ -11,6 +11,9 @@
 
 
 class TestAbs(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class Abs(torch.nn.Module):
         def __init__(self):
             super().__init__()

backends/xnnpack/test/ops/test_add.py

@@ -11,6 +11,9 @@
 
 
 class TestAdd(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class Add(torch.nn.Module):
         def __init__(self):
             super().__init__()

backends/xnnpack/test/ops/test_avgpool2d.py

@@ -11,6 +11,9 @@
 
 
 class TestAvgPool2d(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class AvgPool2d(torch.nn.Module):
         def __init__(
             self, count_include_pad=False, ceil_mode=False, divisor_override=None

backends/xnnpack/test/ops/test_bilinear2d.py

@@ -14,6 +14,9 @@
 
 
 class TestUpsampleBilinear2d(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class StaticResizeBilinear2dModule(torch.nn.Module):
         def forward(self, x):
             a = torch.nn.functional.interpolate(

backends/xnnpack/test/ops/test_bmm.py

@@ -11,6 +11,9 @@
 
 
 class TestBMM(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class BMM(torch.nn.Module):
         def __init__(self):
             super().__init__()

backends/xnnpack/test/ops/test_cat.py

@@ -13,6 +13,9 @@
 
 
 class TestCat(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class Cat(torch.nn.Module):
         def __init__(self, dim=0):
             super().__init__()

backends/xnnpack/test/ops/test_ceil.py

@@ -11,6 +11,9 @@
 
 
 class TestCeil(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class Ceil(torch.nn.Module):
         def __init__(self):
             super().__init__()

backends/xnnpack/test/ops/test_check_quant_params.py

@@ -14,6 +14,9 @@
 
 
 class TestCheckQuantParams(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     def create_invalid_value_injector(
         self, invalid_value, is_per_channel=False, is_zp=False
     ):
@@ -46,6 +49,7 @@ def inject_invalid_scale_in_per_tensor(aten):
         return inject_invalid_scale_in_per_tensor
 
     def _test_check_quant_message(self, ep_modifier, expected_message):
+        torch._dynamo.reset()
         mod = torch.nn.Linear(10, 10)
         quantizer = XNNPACKQuantizer()
         captured = export_for_training(mod, (torch.randn(1, 10),)).module()

backends/xnnpack/test/ops/test_clamp.py

@@ -11,6 +11,9 @@
 
 
 class TestClamp(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class Clamp(torch.nn.Module):
         def __init__(self, min_val=None, max_val=None):
             super().__init__()

backends/xnnpack/test/ops/test_conv1d.py

@@ -19,6 +19,9 @@
 
 
 class TestConv1d(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class Conv1d(torch.nn.Module):
         def __init__(self, dtype: torch.dtype = torch.float):
             groups = 1

backends/xnnpack/test/ops/test_conv2d.py

@@ -170,6 +170,9 @@ def get_inputs(self):
 
 
 class TestConv2d(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     def _test(
         self,
         m: torch.nn.Module,

backends/xnnpack/test/ops/test_div.py

@@ -11,6 +11,9 @@
 
 
 class TestDiv(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class Div(torch.nn.Module):
         def __init__(self):
             super().__init__()

backends/xnnpack/test/ops/test_elu.py

@@ -11,6 +11,9 @@
 
 
 class TestElu(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class ELU(torch.nn.Module):
         def __init__(self):
             super().__init__()

backends/xnnpack/test/ops/test_floor.py

@@ -11,6 +11,9 @@
 
 
 class TestFloor(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class Floor(torch.nn.Module):
         def __init__(self):
             super().__init__()

backends/xnnpack/test/ops/test_hardswish.py

@@ -11,6 +11,9 @@
 
 
 class TestHardswish(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class Hardswish(torch.nn.Module):
         def __init__(self):
             super().__init__()

backends/xnnpack/test/ops/test_hardtanh.py

@@ -11,6 +11,9 @@
 
 
 class TestHardTanh(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class HardTanh(torch.nn.Module):
         def __init__(self, min_val=-1.0, max_val=1.0):
             super().__init__()

backends/xnnpack/test/ops/test_leaky_relu.py

@@ -11,6 +11,9 @@
 
 
 class TestLeakyRelu(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class LeakyReLU(torch.nn.Module):
         def __init__(self, **kwargs):
             super().__init__()

backends/xnnpack/test/ops/test_linear.py

@@ -219,6 +219,9 @@ class TestLinear(unittest.TestCase):
           should produce strictly better results compared to Per-Tensor Quantization
     """
 
+    def setUp(self):
+        torch._dynamo.reset()
+
     @staticmethod
     def _get_4b_dqconfig() -> QuantizationConfig:
         # Returns a QuantizationConfig for 4b dynamic quantization for XNNPACK.

backends/xnnpack/test/ops/test_lstm.py

@@ -14,6 +14,9 @@
 
 
 class TestLSTM(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class LSTMLinear(torch.nn.Module):
         def __init__(self, input_size, hidden_size, out_size):
             super().__init__()

backends/xnnpack/test/ops/test_max_dim.py

@@ -11,6 +11,9 @@
 
 
 class TestMaxDim(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class Max(torch.nn.Module):
         def forward(self, x):
             max_values_1, max_indices_1 = torch.max(x, dim=2, keepdim=True)

backends/xnnpack/test/ops/test_maximum.py

@@ -11,6 +11,9 @@
 
 
 class TestMaximum(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class Maximum(torch.nn.Module):
         def __init__(self):
             super().__init__()

backends/xnnpack/test/ops/test_maxpool2d.py

@@ -15,6 +15,9 @@
 
 
 class TestMaxPool2d(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class MaxPool2d(torch.nn.Module):
         def __init__(self, kernel_size=3, stride=1, padding=0, dilation=1):
             super().__init__()

backends/xnnpack/test/ops/test_mean_dim.py

@@ -11,6 +11,9 @@
 
 
 class TestMeanDim(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class MeanDim(torch.nn.Module):
         def __init__(self, dims):
             super().__init__()

backends/xnnpack/test/ops/test_minimum.py

@@ -11,6 +11,9 @@
 
 
 class TestMinimum(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class Minimum(torch.nn.Module):
         def __init__(self):
             super().__init__()

backends/xnnpack/test/ops/test_multiply.py

@@ -11,6 +11,9 @@
 
 
 class TestMul(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class Mul(torch.nn.Module):
         def forward(self, x, y):
             z = x * y

backends/xnnpack/test/ops/test_negate.py

@@ -11,6 +11,9 @@
 
 
 class TestNegate(unittest.TestCase):
+    def setUp(self):
+        torch._dynamo.reset()
+
     class Negate(torch.nn.Module):
         def __init__(self):
             super().__init__()