Arm backend: Generate random op test inputs lazily

Several Arm operator tests were creating random inputs at module import
time. The Arm test seed is applied later by an autouse pytest fixture,
so those tensors were not actually controlled by ARM_TEST_SEED.

That made tests nondeterministic across fresh pytest processes and could
expose different quantization behavior from run to run. Generate the
affected inputs lazily inside each test case so the existing seed
fixture makes them reproducible and ARM_TEST_SEED=RANDOM can rerandomize
the intended data.

Signed-off-by: Per Held <per.held@arm.com>
Change-Id: Ic4414da5e84b7fb19275e04399634289b10a0a19

Author: perheld

backends/arm/test/ops/test_addmm.py

@@ -28,73 +28,91 @@
 
 
 test_data_suite = {
-    "basic": [
+    "basic": lambda: [
         torch.tensor([[1.0, 2.0], [3.0, 4.0]]),
         torch.tensor([[1.0, 0.0], [0.0, 1.0]]),
         torch.tensor([[1.0, 2.0], [3.0, 4.0]]),
         1.0,
         1.0,
     ],
-    "zeros": [torch.zeros(2, 2), torch.zeros(2, 3), torch.zeros(3, 2), 1.0, 1.0],
-    "beta_only": [
+    "zeros": lambda: [
+        torch.zeros(2, 2),
+        torch.zeros(2, 3),
+        torch.zeros(3, 2),
+        1.0,
+        1.0,
+    ],
+    "beta_only": lambda: [
         torch.tensor([[10.0, 20.0], [30.0, 40.0]]),
         torch.randn(2, 3),
         torch.randn(3, 2),
         0.0,
         1.0,
     ],
-    "alpha_only": [
+    "alpha_only": lambda: [
         torch.tensor([[10.0, 20.0], [30.0, 40.0]]),
         torch.randn(2, 3),
         torch.randn(3, 2),
         1.0,
         0.0,
     ],
-    "scaled": [
+    "scaled": lambda: [
         torch.ones(2, 2),
         torch.tensor([[1.0, 2.0], [3.0, 4.0]]),
         torch.tensor([[5.0, 6.0], [7.0, 8.0]]),
         0.5,
         2.0,
     ],
-    "negative_scalars": [
+    "negative_scalars": lambda: [
         torch.tensor([[1.0, -1.0], [-1.0, 1.0]]),
         torch.tensor([[2.0, 0.0], [0.0, 2.0]]),
         torch.tensor([[1.0, 1.0], [1.0, 1.0]]),
         -1.0,
         -1.0,
     ],
-    "non_square": [torch.ones(3, 4), torch.rand(3, 2), torch.rand(2, 4), 1.0, 1.0],
-    "large_values": [
+    "non_square": lambda: [
+        torch.ones(3, 4),
+        torch.rand(3, 2),
+        torch.rand(2, 4),
+        1.0,
+        1.0,
+    ],
+    "large_values": lambda: [
         torch.full((2, 2), 1e6),
         torch.full((2, 3), 1e3),
         torch.full((3, 2), 1e3),
         1.0,
         1.0,
     ],
-    "small_values": [
+    "small_values": lambda: [
         torch.full((2, 2), 1e-6),
         torch.full((2, 3), 1e-3),
         torch.full((3, 2), 1e-3),
         1.0,
         1.0,
     ],
-    "random": [torch.randn(4, 5), torch.randn(4, 3), torch.randn(3, 5), 1.0, 1.0],
-    "broadcast_bias_row": [
+    "random": lambda: [
+        torch.randn(4, 5),
+        torch.randn(4, 3),
+        torch.randn(3, 5),
+        1.0,
+        1.0,
+    ],
+    "broadcast_bias_row": lambda: [
         torch.randn(1, 2),
         torch.randn(3, 4),
         torch.randn(4, 2),
         1.0,
         1.0,
     ],
-    "row_bias": [
+    "row_bias": lambda: [
         torch.randn(3, 1),
         torch.randn(3, 4),
         torch.randn(4, 4),
         1.0,
         1.0,
     ],
-    "scalar_bias": [
+    "scalar_bias": lambda: [
         torch.tensor(2.0),
         torch.randn(5, 3),
         torch.randn(3, 6),
@@ -120,7 +138,7 @@ def forward(
 def test_addmm_tosa_FP(test_data: Tuple):
     pipeline = TosaPipelineFP[input_t1](
         Addmm(),
-        (*test_data,),
+        (*test_data(),),
         aten_op=aten_op,
         exir_op=exir_op,
     )
@@ -131,7 +149,7 @@ def test_addmm_tosa_FP(test_data: Tuple):
 def test_addmm_tosa_INT(test_data: Tuple):
     pipeline = TosaPipelineINT[input_t1](
         Addmm(),
-        (*test_data,),
+        (*test_data(),),
         aten_op=[],
         exir_op=exir_op,
     )
@@ -143,7 +161,7 @@ def test_addmm_tosa_INT(test_data: Tuple):
 def test_addmm_u55_INT(test_data: Tuple):
     pipeline = EthosU55PipelineINT[input_t1](
         Addmm(),
-        (*test_data,),
+        (*test_data(),),
         aten_ops=[],
         exir_ops=exir_op,
     )
@@ -155,7 +173,7 @@ def test_addmm_u55_INT(test_data: Tuple):
 def test_addmm_u85_INT(test_data: Tuple):
     pipeline = EthosU85PipelineINT[input_t1](
         Addmm(),
-        (*test_data,),
+        (*test_data(),),
         aten_ops=[],
         exir_ops=exir_op,
     )
@@ -167,7 +185,7 @@ def test_addmm_u85_INT(test_data: Tuple):
 def test_addmm_vgf_no_quant(test_data: input_t1):
     pipeline = VgfPipeline[input_t1](
         Addmm(),
-        (*test_data,),
+        (*test_data(),),
         aten_op=aten_op,
         exir_op=exir_op,
         quantize=False,
@@ -180,7 +198,7 @@ def test_addmm_vgf_no_quant(test_data: input_t1):
 def test_addmm_vgf_quant(test_data: input_t1):
     pipeline = VgfPipeline[input_t1](
         Addmm(),
-        (*test_data,),
+        (*test_data(),),
         aten_op=[],
         exir_op=exir_op,
         quantize=True,
@@ -197,7 +215,7 @@ def test_addmm_16a8w_tosa_INT(test_data: input_t1):
 
     pipeline = TosaPipelineINT[input_t1](
         Addmm(),
-        (*test_data,),
+        (*test_data(),),
         aten_op=[],
         exir_op=[],
         per_channel_quantization=per_channel_quantization,
@@ -223,7 +241,7 @@ def test_addmm_16a8w_u55_INT(test_data: input_t1):
 
     pipeline = EthosU55PipelineINT[input_t1](
         Addmm(),
-        (*test_data,),
+        (*test_data(),),
         aten_ops=[],
         exir_ops=[],
         per_channel_quantization=per_channel_quantization,
@@ -245,7 +263,7 @@ def test_addmm_16a8w_u85_INT(test_data: input_t1):
 
     pipeline = EthosU85PipelineINT[input_t1](
         Addmm(),
-        (*test_data,),
+        (*test_data(),),
         aten_ops=[],
         exir_ops=[],
         per_channel_quantization=per_channel_quantization,

backends/arm/test/ops/test_atan.py

@@ -1,4 +1,4 @@
-# Copyright 2025 Arm Limited and/or its affiliates.
+# Copyright 2025-2026 Arm Limited and/or its affiliates.
 #
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
@@ -22,14 +22,14 @@
 input_t1 = Tuple[torch.Tensor]
 
 test_data_suite = {
-    "zeros": torch.zeros(1, 10, 10, 10),
-    "zeros_alt_shape": torch.zeros(1, 10, 3, 5),
-    "ones": torch.ones(10, 10, 10),
-    "rand": torch.rand(10, 10) - 0.5,
-    "rand_alt_shape": torch.rand(1, 10, 3, 5) - 0.5,
-    "randn_pos": torch.randn(10) + 10,
-    "randn_neg": torch.randn(10) - 10,
-    "ramp": torch.arange(-16, 16, 0.2),
+    "zeros": lambda: torch.zeros(1, 10, 10, 10),
+    "zeros_alt_shape": lambda: torch.zeros(1, 10, 3, 5),
+    "ones": lambda: torch.ones(10, 10, 10),
+    "rand": lambda: torch.rand(10, 10) - 0.5,
+    "rand_alt_shape": lambda: torch.rand(1, 10, 3, 5) - 0.5,
+    "randn_pos": lambda: torch.randn(10) + 10,
+    "randn_neg": lambda: torch.randn(10) - 10,
+    "ramp": lambda: torch.arange(-16, 16, 0.2),
 }
 
 
@@ -43,7 +43,7 @@ def forward(self, x: torch.Tensor):
 def test_atan_tosa_FP(test_data: Tuple):
     pipeline = TosaPipelineFP[input_t1](
         Atan(),
-        (test_data,),
+        (test_data(),),
         aten_op=aten_op,
         exir_op=exir_op,
     )
@@ -54,7 +54,7 @@ def test_atan_tosa_FP(test_data: Tuple):
 def test_atan_tosa_INT(test_data: Tuple):
     pipeline = TosaPipelineINT[input_t1](
         Atan(),
-        (test_data,),
+        (test_data(),),
         aten_op=aten_op,
         exir_op=exir_op,
     )
@@ -66,7 +66,7 @@ def test_atan_tosa_INT(test_data: Tuple):
 def test_atan_u55_INT(test_data: Tuple):
     pipeline = EthosU55PipelineINT[input_t1](
         Atan(),
-        (test_data,),
+        (test_data(),),
         aten_ops=aten_op,
         exir_ops=exir_op,
     )
@@ -78,7 +78,7 @@ def test_atan_u55_INT(test_data: Tuple):
 def test_atan_u85_INT(test_data: Tuple):
     pipeline = EthosU85PipelineINT[input_t1](
         Atan(),
-        (test_data,),
+        (test_data(),),
         aten_ops=aten_op,
         exir_ops=exir_op,
     )
@@ -90,7 +90,7 @@ def test_atan_u85_INT(test_data: Tuple):
 def test_atan_vgf_no_quant(test_data: Tuple):
     pipeline = VgfPipeline[input_t1](
         Atan(),
-        (test_data,),
+        (test_data(),),
         aten_op,
         exir_op,
         quantize=False,
@@ -103,7 +103,7 @@ def test_atan_vgf_no_quant(test_data: Tuple):
 def test_atan_vgf_quant(test_data: Tuple):
     pipeline = VgfPipeline[input_t1](
         Atan(),
-        (test_data,),
+        (test_data(),),
         aten_op,
         exir_op,
         quantize=True,

backends/arm/test/ops/test_atanh.py

@@ -24,13 +24,13 @@
 
 
 test_data_suite = {
-    "zeros": torch.zeros(1, 10, 10, 10),
-    "zeros_alt_shape": torch.zeros(1, 10, 3, 5),
-    "rand": torch.rand(10, 10) - 0.5,
-    "rand_alt_shape": torch.rand(1, 10, 3, 5) - 0.5,
-    "ramp": torch.arange(-1, 1, 0.2),
-    "near_bounds": torch.tensor([-0.99, -0.9, 0.9, 0.99]),
-    "on_bounds": torch.tensor([-1.0, 1.0]),
+    "zeros": lambda: torch.zeros(1, 10, 10, 10),
+    "zeros_alt_shape": lambda: torch.zeros(1, 10, 3, 5),
+    "rand": lambda: torch.rand(10, 10) - 0.5,
+    "rand_alt_shape": lambda: torch.rand(1, 10, 3, 5) - 0.5,
+    "ramp": lambda: torch.arange(-1, 1, 0.2),
+    "near_bounds": lambda: torch.tensor([-0.99, -0.9, 0.9, 0.99]),
+    "on_bounds": lambda: torch.tensor([-1.0, 1.0]),
 }
 
 
@@ -43,7 +43,7 @@ def forward(self, x: torch.Tensor):
 def test_atanh_tosa_FP(test_data: Tuple):
     pipeline = TosaPipelineFP[input_t1](
         Atanh(),
-        (test_data,),
+        (test_data(),),
         aten_op=aten_op,
         exir_op=exir_op,
     )
@@ -52,13 +52,14 @@ def test_atanh_tosa_FP(test_data: Tuple):
 
 @common.parametrize("test_data", test_data_suite)
 def test_atanh_tosa_INT(test_data: Tuple):
+    input_data = test_data()
     pipeline = TosaPipelineINT[input_t1](
         Atanh(),
-        (test_data,),
+        (input_data,),
         aten_op=aten_op,
         exir_op=exir_op,
     )
-    if torch.any(test_data >= 1) or torch.any(test_data <= -1):
+    if torch.any(input_data >= 1) or torch.any(input_data <= -1):
         # The quantized model will saturate to max/min values while the
         # original model will return inf/-inf, so comparison wont be valid here.
         pipeline.pop_stage("run_method_and_compare_outputs.original_model")
@@ -70,7 +71,7 @@ def test_atanh_tosa_INT(test_data: Tuple):
 def test_atanh_u55_INT(test_data: Tuple):
     pipeline = EthosU55PipelineINT[input_t1](
         Atanh(),
-        (test_data,),
+        (test_data(),),
         aten_ops=aten_op,
         exir_ops=exir_op,
     )
@@ -82,7 +83,7 @@ def test_atanh_u55_INT(test_data: Tuple):
 def test_atanh_u85_INT(test_data: Tuple):
     pipeline = EthosU85PipelineINT[input_t1](
         Atanh(),
-        (test_data,),
+        (test_data(),),
         aten_ops=aten_op,
         exir_ops=exir_op,
     )
@@ -94,7 +95,7 @@ def test_atanh_u85_INT(test_data: Tuple):
 def test_atanh_vgf_no_quant(test_data: input_t1):
     pipeline = VgfPipeline[input_t1](
         Atanh(),
-        (test_data,),
+        (test_data(),),
         aten_op=aten_op,
         exir_op=exir_op,
         quantize=False,
@@ -107,7 +108,7 @@ def test_atanh_vgf_no_quant(test_data: input_t1):
 def test_atanh_vgf_quant(test_data: input_t1):
     pipeline = VgfPipeline[input_t1](
         Atanh(),
-        (test_data,),
+        (test_data(),),
         aten_op=aten_op,
         exir_op=exir_op,
         quantize=True,