Merge pull request #589 from InfiniTensor/issue/566

issue/566: 添加 Python 算子测试

test/infinicore/ops/fmin.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+from framework.utils import is_broadcast
+
+# Test cases format: (shape, a_strides, b_strides)
+_TEST_CASES_DATA = [
+    ((6, 8), None, None),
+    ((8, 4), (16, 1), None),
+    ((5, 5), None, (10, 1)),
+    ((3, 7), (14, 1), (14, 1)),
+    ((10, 3), None, None),
+    ((2, 16), (32, 1), (32, 1)),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 0, "rtol": 1e-2},
+    infinicore.float32: {"atol": 0, "rtol": 1e-3},
+    infinicore.bfloat16: {"atol": 0, "rtol": 5e-2},
+}
+
+_TENSOR_DTYPES = [infinicore.float16, infinicore.bfloat16, infinicore.float32]
+
+
+def parse_test_cases():
+    test_cases = []
+    for data in _TEST_CASES_DATA:
+        shape, a_strides, b_strides = data
+
+        a_inplace = not is_broadcast(a_strides)
+        b_inplace = not is_broadcast(b_strides)
+
+        for dtype in _TENSOR_DTYPES:
+            tol = _TOLERANCE_MAP.get(dtype, {"atol": 0, "rtol": 1e-3})
+
+            a_spec = TensorSpec.from_tensor(shape, a_strides, dtype)
+            b_spec = TensorSpec.from_tensor(shape, b_strides, dtype)
+
+            test_cases.append(
+                TestCase(
+                    inputs=[a_spec, b_spec],
+                    kwargs={},
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tol,
+                    description="fmin - OUT",
+                )
+            )
+
+            if a_inplace:
+                test_cases.append(
+                    TestCase(
+                        inputs=[a_spec, b_spec],
+                        kwargs={"out": 0},
+                        output_spec=None,
+                        comparison_target=0,
+                        tolerance=tol,
+                        description="fmin - INPLACE(a)",
+                    )
+                )
+
+            if b_inplace:
+                test_cases.append(
+                    TestCase(
+                        inputs=[a_spec, b_spec],
+                        kwargs={"out": 1},
+                        output_spec=None,
+                        comparison_target=1,
+                        tolerance=tol,
+                        description="fmin - INPLACE(b)",
+                    )
+                )
+
+    return test_cases
+
+
+class OpTest(BaseOperatorTest):
+    """FMin operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("FMin")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.fmin(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.fmin(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/fmod.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+from framework.utils import is_broadcast
+
+# Test cases format: (in_shape, in_strides_or_None, other_shape_or_None)
+# infinicore.fmod(input, other)
+
+_TEST_CASES_DATA = [
+    ((2, 3, 4), None, None),
+    ((1, 4, 8), (32, 8, 1), None),
+    ((3, 2, 5, 7), None, None),
+    ((2, 1, 16), None, None),
+    ((1, 8, 9, 11), (792, 99, 11, 1), None),
+    ((2, 6, 10), None, None),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 1e-3, "rtol": 1e-2},
+    infinicore.float32: {"atol": 1e-5, "rtol": 1e-4},
+}
+_TENSOR_DTYPES = [infinicore.float16, infinicore.float32]
+
+
+def parse_test_cases():
+    cases = []
+    for in_shape, in_strides, other_shape in _TEST_CASES_DATA:
+        for dtype in _TENSOR_DTYPES:
+            tol = _TOLERANCE_MAP[dtype]
+            a_spec = TensorSpec.from_tensor(in_shape, in_strides, dtype)
+            b_spec = TensorSpec.from_tensor(
+                in_shape if other_shape is None else other_shape, None, dtype
+            )
+
+            cases.append(
+                TestCase(
+                    inputs=[a_spec, b_spec],
+                    kwargs={},
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tol,
+                    description="fmod_out_of_place",
+                )
+            )
+
+            out_spec = TensorSpec.from_tensor(in_shape, None, dtype)
+            cases.append(
+                TestCase(
+                    inputs=[a_spec, b_spec],
+                    kwargs={},
+                    output_spec=out_spec,
+                    comparison_target="out",
+                    tolerance=tol,
+                    description="fmod_explicit_out",
+                )
+            )
+
+            if not is_broadcast(a_spec.strides):
+                cases.append(
+                    TestCase(
+                        inputs=[a_spec, b_spec],
+                        kwargs={"out": 0},
+                        output_spec=None,
+                        comparison_target=0,
+                        tolerance=tol,
+                        description="fmod_inplace_a",
+                    )
+                )
+            if not is_broadcast(b_spec.strides):
+                cases.append(
+                    TestCase(
+                        inputs=[a_spec, b_spec],
+                        kwargs={"out": 1},
+                        output_spec=None,
+                        comparison_target=1,
+                        tolerance=tol,
+                        description="fmod_inplace_b",
+                    )
+                )
+
+    return cases
+
+
+class OpTest(BaseOperatorTest):
+    """Fmod operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("Fmod")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.fmod(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.fmod(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/fold.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+
+# Test cases format: (in_shape, in_strides_or_None, output_size, kernel_size, dilation, padding, stride)
+# For fold/unfold: input is the output of unfold; fold reconstructs image from patches.
+
+_TEST_CASES_DATA = [
+    # ((2, 6, 9), None, (4, 4), (2, 2), 1, 0, (2, 2)),
+    # ((1, 8, 16), None, (8, 8), (4, 4), 1, 0, (2, 2)),
+    # ((3, 4, 12), None, (6, 6), (3, 2), 1, 1, (1, 2)),
+    # ((2, 2, 20), None, (10, 2), (2, 2), 1, 0, (2, 1)),
+    # ((1, 3, 25), None, (5, 5), (5, 5), 1, 0, (5, 5)),
+    # ((2, 5, 18), (90, 18, 1), (9, 2), (3, 2), 1, 0, (2, 1)),
+    # 原来对应 ((2,3,8,8), None, (3,3), 1, 0, (1,1))
+    # 计算得到 L=6*6=36, channels_for_fold = 3*3*3 = 27
+    ((2, 27, 36), None, (8, 8), (3, 3), 1, 0, (1, 1)),
+    # 原来对应 ((1,4,10,12), None, (5,3), 1, 1, (2,1))
+    # L = 4 * 12 = 48, channels = 4*5*3 = 60
+    ((1, 60, 48), None, (10, 12), (5, 3), 1, 1, (2, 1)),
+    # 原来对应 ((2,2,16,16), (512,256,16,1), (4,4), 1, 0, (4,4))
+    # L = 4 * 4 = 16, channels = 2*4*4 = 32
+    ((2, 32, 16), None, (16, 16), (4, 4), 1, 0, (4, 4)),
+    # 原来对应 ((3,6,7,9), None, (3,2), 1, 0, (1,1))
+    # L = 5 * 8 = 40, channels = 6*3*2 = 36
+    ((3, 36, 40), None, (7, 9), (3, 2), 1, 0, (1, 1)),
+    # 原来对应 ((1,8,9,11), None, (2,3), 1, 1, (1,2))
+    # L = 10 * 6 = 60, channels = 8*2*3 = 48
+    ((1, 48, 60), None, (9, 11), (2, 3), 1, 1, (1, 2)),
+    # 原来对应 ((2,5,12,6), (360,72,6,1), (3,3), 1, 0, (2,1))
+    # L = 5 * 4 = 20, channels = 5*3*3 = 45
+    ((2, 45, 20), None, (12, 6), (3, 3), 1, 0, (2, 1)),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 1e-3, "rtol": 1e-2},
+    infinicore.float32: {"atol": 1e-5, "rtol": 1e-4},
+}
+
+_TENSOR_DTYPES = [infinicore.float16, infinicore.float32]
+
+
+def parse_test_cases():
+    cases = []
+    for (
+        in_shape,
+        in_strides,
+        output_size,
+        kernel_size,
+        dilation,
+        padding,
+        stride,
+    ) in _TEST_CASES_DATA:
+        for dtype in _TENSOR_DTYPES:
+            tol = _TOLERANCE_MAP.get(dtype)
+            in_spec = TensorSpec.from_tensor(in_shape, in_strides, dtype)
+
+            kwargs = {
+                "output_size": output_size,
+                "kernel_size": kernel_size,
+                "dilation": dilation,
+                "padding": padding,
+                "stride": stride,
+            }
+
+            cases.append(
+                TestCase(
+                    inputs=[in_spec],
+                    kwargs=kwargs,
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tol,
+                    description="Fold - OUT_OF_PLACE",
+                )
+            )
+
+    return cases
+
+
+class OpTest(BaseOperatorTest):
+    """Fold operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("Fold")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.nn.functional.fold(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.nn.functional.fold(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/frac.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+from framework.utils import is_broadcast
+
+# Test cases format: (in_shape, in_strides_or_None)
+# infinicore.frac(input)
+
+_TEST_CASES_DATA = [
+    ((2, 3), None),
+    ((1, 4, 8), (32, 8, 1)),
+    ((3, 2, 5, 7), None),
+    ((2, 1, 16), None),
+    ((1, 8, 9, 11), (792, 99, 11, 1)),
+    ((2, 6, 10), None),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 1e-3, "rtol": 1e-2},
+    infinicore.float32: {"atol": 1e-5, "rtol": 1e-4},
+}
+_TENSOR_DTYPES = [infinicore.float16, infinicore.float32]
+
+
+def parse_test_cases():
+    cases = []
+    for shape, strides in _TEST_CASES_DATA:
+        for dtype in _TENSOR_DTYPES:
+            tol = _TOLERANCE_MAP[dtype]
+            in_spec = TensorSpec.from_tensor(shape, strides, dtype)
+
+            cases.append(
+                TestCase(
+                    inputs=[in_spec],
+                    kwargs={},
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tol,
+                    description="frac_out",
+                )
+            )
+            out_spec = TensorSpec.from_tensor(shape, None, dtype)
+            cases.append(
+                TestCase(
+                    inputs=[in_spec],
+                    kwargs={},
+                    output_spec=out_spec,
+                    comparison_target="out",
+                    tolerance=tol,
+                    description="frac_out_explicit",
+                )
+            )
+            if not is_broadcast(in_spec.strides):
+                cases.append(
+                    TestCase(
+                        inputs=[in_spec],
+                        kwargs={"out": 0},
+                        output_spec=None,
+                        comparison_target=0,
+                        tolerance=tol,
+                        description="frac_inplace",
+                    )
+                )
+
+    return cases
+
+
+class OpTest(BaseOperatorTest):
+    """Frac operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("Frac")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.frac(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.frac(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/fractional_max_pool2d.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+
+# Test cases format: (in_shape, in_strides_or_None, kernel_size, output_size_or_None, return_indices)
+# Note: fractional_max_pool may return values and indices; PyTorch accepts optional random samples. We avoid
+# explicit _random_samples and focus on default behavior. Indices (if returned) form a separate output; tests
+# here exercise the value-returning path.
+
+_TEST_CASES_DATA = [
+    ((2, 3, 15, 15), None, (3, 3), (5, 5), False),
+    ((1, 4, 16, 14), (896, 224, 14, 1), (4, 3), (4, 5), False),
+    ((2, 2, 17, 19), None, (5, 5), (7, 6), False),
+    ((3, 6, 9, 11), None, (2, 2), (4, 5), False),
+    ((1, 8, 20, 20), (3200, 400, 20, 1), (3, 3), (6, 6), False),
+    ((2, 5, 12, 10), None, (4, 3), (3, 3), False),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 1e-3, "rtol": 1e-2},
+    infinicore.float32: {"atol": 1e-5, "rtol": 1e-4},
+}
+_TENSOR_DTYPES = [infinicore.float16, infinicore.float32]
+
+
+def parse_test_cases():
+    cases = []
+    for in_shape, in_strides, kernel_size, out_size, return_indices in _TEST_CASES_DATA:
+        for dtype in _TENSOR_DTYPES:
+            tol = _TOLERANCE_MAP[dtype]
+            in_spec = TensorSpec.from_tensor(in_shape, in_strides, dtype)
+            kwargs = {
+                "kernel_size": kernel_size,
+                "output_size": out_size,
+                "return_indices": return_indices,
+            }
+            cases.append(
+                TestCase(
+                    inputs=[in_spec],
+                    kwargs=kwargs,
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tol,
+                    description="FractionalMaxPool2d - OUT_OF_PLACE",
+                )
+            )
+
+    return cases
+
+
+class OpTest(BaseOperatorTest):
+    """FractionalMaxPool2d operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("FractionalMaxPool2d")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.nn.functional.fractional_max_pool2d(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.nn.functional.fractional_max_pool2d(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/fractional_max_pool3d.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+
+# Test cases format: (in_shape, in_strides_or_None, kernel_size, output_size_or_None, return_indices)
+# Note: PyTorch fractional_max_pool3d behaves similarly to fractional_max_pool2d; we avoid _random_samples.
+
+_TEST_CASES_DATA = [
+    ((2, 3, 9, 9, 9), None, (3, 3, 3), (4, 4, 4), False),
+    ((1, 4, 8, 10, 12), None, (2, 3, 2), (4, 4, 6), False),
+    ((2, 2, 7, 11, 5), (770, 110, 55, 5, 1), (3, 2, 3), (3, 4, 2), False),
+    ((3, 6, 5, 6, 7), None, (2, 2, 2), (3, 3, 4), False),
+    ((1, 8, 10, 10, 10), None, (4, 3, 2), (5, 4, 5), False),
+    ((2, 5, 12, 8, 6), None, (3, 3, 2), (4, 3, 2), False),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 1e-3, "rtol": 1e-2},
+    infinicore.float32: {"atol": 1e-5, "rtol": 1e-4},
+}
+_TENSOR_DTYPES = [infinicore.float16, infinicore.float32]
+
+
+def parse_test_cases():
+    cases = []
+    for in_shape, in_strides, kernel_size, out_size, return_indices in _TEST_CASES_DATA:
+        for dtype in _TENSOR_DTYPES:
+            tol = _TOLERANCE_MAP[dtype]
+            in_spec = TensorSpec.from_tensor(in_shape, in_strides, dtype)
+            kwargs = {
+                "kernel_size": kernel_size,
+                "output_size": out_size,
+                "return_indices": return_indices,
+            }
+            cases.append(
+                TestCase(
+                    inputs=[in_spec],
+                    kwargs=kwargs,
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tol,
+                    description="FractionalMaxPool3d - OUT_OF_PLACE",
+                )
+            )
+
+    return cases
+
+
+class OpTest(BaseOperatorTest):
+    """FractionalMaxPool3d operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("FractionalMaxPool3d")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.nn.functional.fractional_max_pool3d(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.nn.functional.fractional_max_pool3d(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/frexp.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+
+# Test cases format: (in_shape, in_strides_or_None)
+
+_TEST_CASES_DATA = [
+    ((2, 3), None),
+    ((1, 4, 8), (32, 8, 1)),
+    ((3, 2, 5, 7), None),
+    ((2, 1, 16), None),
+    ((1, 8, 9, 11), (792, 99, 11, 1)),
+    ((2, 6, 10), None),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 1e-3, "rtol": 1e-2},
+    infinicore.float32: {"atol": 1e-5, "rtol": 1e-4},
+}
+_TENSOR_DTYPES = [infinicore.float16, infinicore.float32]
+
+
+def parse_test_cases():
+    cases = []
+    for shape, strides in _TEST_CASES_DATA:
+        for dtype in _TENSOR_DTYPES:
+            tol = _TOLERANCE_MAP[dtype]
+            in_spec = TensorSpec.from_tensor(shape, strides, dtype)
+
+            # Return-value tests only
+            cases.append(
+                TestCase(
+                    inputs=[in_spec],
+                    kwargs={},
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tol,
+                    description="frexp_out_tuple",
+                )
+            )
+
+    return cases
+
+
+class OpTest(BaseOperatorTest):
+    """Frexp operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("Frexp")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.frexp(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.frexp(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/full.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+from framework.datatypes import to_torch_dtype
+
+# Test cases format: (shape, fill_value, dtype)
+_TEST_CASES_DATA = [
+    ((3, 4), 0.0, infinicore.float32),
+    ((6, 2), 1.5, infinicore.float16),
+    ((5, 5), -2.0, infinicore.float32),
+    ((1, 7), 3.14, infinicore.bfloat16),
+    ((8, 3), 42.0, infinicore.float32),
+    ((2, 2, 2), 0.5, infinicore.float16),
+]
+
+_TOLERANCE_MAP = {infinicore.float32: {"atol": 1e-5, "rtol": 1e-4}}
+_TENSOR_DTYPES = [infinicore.float32, infinicore.float16, infinicore.bfloat16]
+
+
+def parse_test_cases():
+    test_cases = []
+    for shape, val, dtype in _TEST_CASES_DATA:
+        out_spec = TensorSpec.from_tensor(shape, None, dtype)
+        # kwargs = {"fill_value": val, "size": shape, "dtype": to_torch_dtype(dtype)}
+        kwargs = {"fill_value": val, "size": shape, "dtype": dtype}
+
+        test_cases.append(
+            TestCase(
+                inputs=[],
+                kwargs=kwargs,
+                output_spec=out_spec,
+                comparison_target="out",
+                tolerance=_TOLERANCE_MAP.get(dtype, {"atol": 1e-5, "rtol": 1e-4}),
+                description=f"full - OUT_OF_PLACE",
+            )
+        )
+
+    return test_cases
+
+
+class OpTest(BaseOperatorTest):
+    """Full operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("Full")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        if "fill_value" not in kwargs:
+            raise TypeError("full test did not provide 'fill_value' parameter")
+        fill_value = kwargs.pop("fill_value")
+
+        if "size" not in kwargs:
+            raise TypeError("full test did not provide 'size' parameter")
+        size = kwargs.pop("size")
+
+        if "dtype" not in kwargs:
+            raise TypeError("full test did not provide 'dtype' parameter")
+        dtype_torch = to_torch_dtype(kwargs.pop("dtype"))
+        if dtype_torch is None:
+            raise TypeError("full test provided unsupported 'dtype' parameter")
+
+        # 支持测试框架通过 kwargs 注入 out 参数
+        out = kwargs.pop("out", None)
+
+        if out is not None:
+            return torch.full(tuple(size), fill_value, dtype=dtype_torch, out=out)
+        else:
+            return torch.full(tuple(size), fill_value, dtype=dtype_torch)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.full(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/full_like.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+from framework.datatypes import to_torch_dtype
+
+# Test cases format: (base_shape, base_strides_or_None, fill_value, dtype_or_None)
+_TEST_CASES_DATA = [
+    ((3, 4), None, 0.0, None),
+    ((6, 2), (12, 1), 1.5, infinicore.float16),
+    ((5, 5), None, -2.0, infinicore.float32),
+    ((1, 7), None, 3.14, infinicore.bfloat16),
+    ((8, 3), (24, 1), 42.0, None),
+    ((2, 2, 2), None, 0.5, infinicore.float16),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float32: {"atol": 0, "rtol": 0},
+    infinicore.float16: {"atol": 1e-3, "rtol": 1e-3},
+    infinicore.bfloat16: {"atol": 1e-2, "rtol": 1e-2},
+}
+
+_TENSOR_DTYPES = [infinicore.float32, infinicore.float16, infinicore.bfloat16]
+
+
+def parse_test_cases():
+    test_cases = []
+    for base_shape, base_strides, val, dtype in _TEST_CASES_DATA:
+        for input_dtype in _TENSOR_DTYPES:
+            base_spec = TensorSpec.from_tensor(base_shape, base_strides, input_dtype)
+
+            kwargs = {"fill_value": val}
+            if dtype is not None:
+                kwargs["dtype"] = dtype
+
+            # torch.full_like does not accept an `out=` kwarg in most PyTorch
+            # versions; call out-of-place and compare the return value instead.
+            test_cases.append(
+                TestCase(
+                    inputs=[base_spec],
+                    kwargs=kwargs,
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=_TOLERANCE_MAP.get(dtype, {"atol": 1e-5, "rtol": 1e-4}),
+                    description=f"full_like - OUT_OF_PLACE",
+                )
+            )
+
+    return test_cases
+
+
+class OpTest(BaseOperatorTest):
+    """FullLike operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("FullLike")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        if "fill_value" not in kwargs:
+            raise TypeError("full_like test did not provide 'fill_value' parameter")
+        fill_value = kwargs.pop("fill_value")
+
+        if "dtype" not in kwargs:
+            dtype_torch = None
+        else:
+            dtype_torch = to_torch_dtype(kwargs.pop("dtype"))
+        return torch.full_like(*args, fill_value=fill_value, dtype=dtype_torch)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore full_like implementation (operator not yet available)."""
+    #     # Mirror the torch_operator signature/behavior but call infinicore implementation.
+    #     if "fill_value" not in kwargs:
+    #         raise TypeError("full_like test did not provide 'fill_value' parameter")
+    #     fill_value = kwargs.pop("fill_value")
+    #
+    #     if "dtype" not in kwargs:
+    #         dtype_infinicore = None
+    #     else:
+    #         # tests pass infinicore dtypes directly, so forward as-is
+    #         dtype_infinicore = kwargs.pop("dtype")
+    #     return infinicore.full_like(*args, fill_value=fill_value, dtype=dtype_infinicore)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/gather.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+from framework.tensor import TensorInitializer
+from framework.utils import is_broadcast
+
+# Test cases format: (input_shape, input_strides_or_None, dim, index_shape)
+_TEST_CASES_DATA = [
+    ((3, 4), None, 1, (3, 2)),
+    ((5, 6), (30, 1), 0, (2, 6)),
+    ((2, 3, 4), None, 2, (2, 3, 2)),
+    ((4, 4), None, -1, (4, 2)),
+    ((6, 2), (12, 1), 1, (6, 1)),
+    ((3, 5), None, 0, (1, 5)),
+]
+
+_TOLERANCE_MAP = {infinicore.float32: {"atol": 1e-5, "rtol": 1e-4}}
+_TENSOR_DTYPES = [infinicore.float32]
+
+
+def parse_test_cases():
+    test_cases = []
+    for shape, strides, dim, idx_shape in _TEST_CASES_DATA:
+        input_spec = TensorSpec.from_tensor(shape, strides, infinicore.float32)
+        input_ndim = len(shape)
+        if dim < 0:
+            actual_dim = input_ndim + dim
+        else:
+            actual_dim = dim
+
+        size_dim = shape[actual_dim]
+
+        # index tensor spec：值必须在 [0, size_dim)
+        index_spec = TensorSpec.from_tensor(
+            idx_shape,
+            None,
+            infinicore.int64,
+            init_mode=TensorInitializer.RANDINT,
+            low=0,
+            high=size_dim,  # exclusive bound
+        )
+
+        # gather returns same dtype as input
+        kwargs = {"dim": dim}
+
+        test_cases.append(
+            TestCase(
+                inputs=[input_spec, index_spec],
+                kwargs=kwargs,
+                output_spec=None,
+                comparison_target=None,
+                tolerance=_TOLERANCE_MAP[infinicore.float32],
+                description=f"gather - OUT_OF_PLACE",
+            )
+        )
+
+    return test_cases
+
+
+class OpTest(BaseOperatorTest):
+    """Gather operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("Gather")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        if "dim" not in kwargs:
+            raise TypeError("gather test did not provide 'dim' parameter")
+        dim = kwargs.pop("dim")
+        input = args[0]
+        index = args[1]
+
+        return torch.gather(input, dim, index)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.gather(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/gaussian_nll_loss.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+
+# Test cases format: (input_shape, var_present_bool, full_or_None, eps_or_None, input_strides_or_None)
+# infinicore.nn.functional.gaussian_nll_loss(input, target, var, full=False, eps=1e-6, reduction='mean')
+
+_TEST_CASES_DATA = [
+    ((4, 5), True, None, None, None),
+    ((8, 8), True, True, 1e-6, (512, 64)),
+    ((1, 10), True, False, 1e-3, None),
+    ((16, 100), True, None, None, None),
+    ((3, 7), True, True, 1e-5, None),
+    ((2, 2), True, False, None, None),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 1e-2, "rtol": 1e-1},
+    infinicore.float32: {"atol": 1e-5, "rtol": 1e-4},
+    infinicore.bfloat16: {"atol": 1e-2, "rtol": 5e-2},
+}
+
+_TENSOR_DTYPES = [infinicore.float16, infinicore.bfloat16, infinicore.float32]
+
+
+def parse_test_cases():
+    test_cases = []
+    for shape, var_present, full, eps, strides in _TEST_CASES_DATA:
+        for dtype in _TENSOR_DTYPES:
+            tol = _TOLERANCE_MAP.get(dtype, {"atol": 1e-5, "rtol": 1e-4})
+            inp = TensorSpec.from_tensor(shape, strides, dtype)
+            tgt = TensorSpec.from_tensor(shape, None, dtype)
+            var = TensorSpec.from_tensor(shape, None, dtype) if var_present else None
+
+            inputs = [inp, tgt]
+            if var is not None:
+                inputs.append(var)
+
+            kwargs = {}
+            if full is not None:
+                kwargs["full"] = full
+            if eps is not None:
+                kwargs["eps"] = eps
+
+            test_cases.append(
+                TestCase(
+                    inputs=inputs,
+                    kwargs=kwargs,
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tol,
+                    description="gaussian_nll_loss - OUT_OF_PLACE",
+                )
+            )
+
+    return test_cases
+
+
+class OpTest(BaseOperatorTest):
+    """gaussian_nll_loss operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("gaussian_nll_loss")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.nn.functional.gaussian_nll_loss(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.nn.functional.gaussian_nll_loss(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/gcd.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+from framework.utils import is_broadcast
+
+# Test cases format: (shape, a_strides, b_strides)
+_TEST_CASES_DATA = [
+    ((6, 8), None, None),
+    ((8, 4), (16, 1), None),
+    ((5, 5), None, (10, 1)),
+    ((3, 7), (14, 1), (14, 1)),
+    ((10, 3), None, None),
+    ((2, 16), (32, 1), (32, 1)),
+]
+
+_TOLERANCE_MAP = {infinicore.int64: {"atol": 0, "rtol": 0}}
+_TENSOR_DTYPES = [infinicore.int64]
+
+
+def parse_test_cases():
+    test_cases = []
+    for data in _TEST_CASES_DATA:
+        shape, a_strides, b_strides = data
+
+        a_inplace = not is_broadcast(a_strides)
+        b_inplace = not is_broadcast(b_strides)
+
+        for dtype in _TENSOR_DTYPES:
+            tol = _TOLERANCE_MAP.get(dtype)
+
+            a_spec = TensorSpec.from_tensor(shape, a_strides, dtype)
+            b_spec = TensorSpec.from_tensor(shape, b_strides, dtype)
+
+            # Out-of-place
+            test_cases.append(
+                TestCase(
+                    inputs=[a_spec, b_spec],
+                    kwargs={},
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tol,
+                    description="gcd - OUT",
+                )
+            )
+
+            if a_inplace:
+                test_cases.append(
+                    TestCase(
+                        inputs=[a_spec, b_spec],
+                        kwargs={"out": 0},
+                        output_spec=None,
+                        comparison_target=0,
+                        tolerance=tol,
+                        description="gcd - INPLACE(a)",
+                    )
+                )
+
+            if b_inplace:
+                test_cases.append(
+                    TestCase(
+                        inputs=[a_spec, b_spec],
+                        kwargs={"out": 1},
+                        output_spec=None,
+                        comparison_target=1,
+                        tolerance=tol,
+                        description="gcd - INPLACE(b)",
+                    )
+                )
+
+    return test_cases
+
+
+class OpTest(BaseOperatorTest):
+    """GCD operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("GCD")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.gcd(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.gcd(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/glu.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+from framework.utils import is_broadcast
+
+# Test cases format: (in_shape, in_strides_or_None, dim_or_None)
+
+_TEST_CASES_DATA = [
+    ((13, 4), None, -1),
+    ((8, 6), (48, 1), -1),  # last dim divisible by 2
+    ((4, 8, 6), (48, 8, 1), 1),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 1e-3, "rtol": 1e-2},
+    infinicore.float32: {"atol": 1e-5, "rtol": 1e-4},
+    infinicore.bfloat16: {"atol": 1e-2, "rtol": 5e-2},
+}
+
+_TENSOR_DTYPES = [infinicore.float16, infinicore.bfloat16, infinicore.float32]
+
+
+def parse_test_cases():
+    """GLU: infinicore.nn.functional.glu(input, dim=-1)
+
+    GLU splits last dimension in half; ensure last dim is even in test shapes.
+    No inplace argument in functional API, so only out-of-place tests.
+    """
+    test_cases = []
+
+    for data in _TEST_CASES_DATA:
+        shape = data[0]
+        in_strides = data[1] if len(data) > 1 else None
+        dim = data[2] if len(data) > 2 else -1
+
+        for dtype in _TENSOR_DTYPES:
+            tolerance = _TOLERANCE_MAP.get(dtype, {"atol": 1e-5, "rtol": 1e-4})
+
+            input_spec = TensorSpec.from_tensor(shape, in_strides, dtype)
+
+            kwargs = {"dim": dim}
+
+            test_cases.append(
+                TestCase(
+                    inputs=[input_spec],
+                    kwargs=kwargs,
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tolerance,
+                    description=f"GLU - OUT_OF_PLACE",
+                )
+            )
+
+    return test_cases
+
+
+class OpTest(BaseOperatorTest):
+    """GLU operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("GLU")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.nn.functional.glu(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.nn.functional.glu(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/group_norm.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+
+# Test cases format: (input_shape, input_strides_or_None, num_groups, weight_bias_present_bool, eps_or_None)
+# infinicore.nn.functional.group_norm(input, num_groups, weight=None, bias=None, eps=1e-5)
+
+_TEST_CASES_DATA = [
+    ((4, 8, 16, 16), None, 4, True, None),
+    ((2, 6, 8, 8), (768, 128, 1, 1), 3, False, 1e-3),
+    ((1, 3, 10, 10), None, 3, True, None),
+    ((8, 12, 6, 6), None, 6, True, 1e-4),
+    ((6, 4, 7, 7), None, 2, False, None),
+    ((3, 2, 9, 9), None, 1, True, None),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 1e-2, "rtol": 1e-1},
+    infinicore.float32: {"atol": 1e-5, "rtol": 1e-4},
+    infinicore.bfloat16: {"atol": 1e-2, "rtol": 5e-2},
+}
+
+_TENSOR_DTYPES = [infinicore.float16, infinicore.bfloat16, infinicore.float32]
+
+
+def parse_test_cases():
+    test_cases = []
+    for shape, strides, num_groups, wb_present, eps in _TEST_CASES_DATA:
+        C = shape[1]
+        for dtype in _TENSOR_DTYPES:
+            tol = _TOLERANCE_MAP.get(dtype, {"atol": 1e-5, "rtol": 1e-4})
+            inp = TensorSpec.from_tensor(shape, strides, dtype)
+
+            # infinicore.nn.functional.group_norm(input, num_groups, weight=None, bias=None, eps=1e-5)
+            # pass num_groups as positional argument to avoid duplicate kwarg issue
+            inputs = [inp, num_groups]
+            kwargs = {}
+            if wb_present:
+                weight = TensorSpec.from_tensor((C,), None, dtype)
+                bias = TensorSpec.from_tensor((C,), None, dtype)
+                inputs.append(weight)
+                inputs.append(bias)
+            else:
+                # explicit None placeholders for weight and bias
+                inputs.append(None)
+                inputs.append(None)
+
+            if eps is not None:
+                kwargs["eps"] = eps
+
+            test_cases.append(
+                TestCase(
+                    inputs=inputs,
+                    kwargs=kwargs,
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tol,
+                    description="group_norm - OUT_OF_PLACE",
+                )
+            )
+
+    return test_cases
+
+
+class OpTest(BaseOperatorTest):
+    """group_norm operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("group_norm")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.nn.functional.group_norm(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.nn.functional.group_norm(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/gt.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+from framework.utils import is_broadcast
+
+# Test cases format: (shape, a_strides_or_None, b_strides_or_None, out_strides_or_None)
+# greater-than comparison
+
+_TEST_CASES_DATA = [
+    ((8, 8), None, None, None),
+    ((8, 8), (16, 1), (16, 1), None),
+    ((8, 8), None, (0, 1), None),
+    ((1, 8), None, None, (8, 1)),
+    ((2, 3, 4), None, None, None),
+    ((32, 64), (128, 1), (128, 1), None),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 0, "rtol": 1e-2},
+    infinicore.float32: {"atol": 0, "rtol": 1e-3},
+    infinicore.int32: {"atol": 0, "rtol": 0},
+}
+
+_TENSOR_DTYPES = [infinicore.float16, infinicore.float32, infinicore.int32]
+
+
+def parse_test_cases():
+    test_cases = []
+    for data in _TEST_CASES_DATA:
+        shape, a_strides, b_strides, out_strides = data[0], data[1], data[2], data[3]
+
+        a_supports_inplace = not is_broadcast(a_strides)
+        b_supports_inplace = not is_broadcast(b_strides)
+        out_supports_inplace = not is_broadcast(out_strides)
+
+        for dtype in _TENSOR_DTYPES:
+            tol = _TOLERANCE_MAP.get(dtype, {"atol": 0, "rtol": 1e-3})
+            a_spec = TensorSpec.from_tensor(shape, a_strides, dtype)
+            b_spec = TensorSpec.from_tensor(shape, b_strides, dtype)
+            out_spec = TensorSpec.from_tensor(shape, out_strides, infinicore.bool)
+
+            test_cases.append(
+                TestCase(
+                    inputs=[a_spec, b_spec],
+                    kwargs={},
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tol,
+                    description="GT - OUT_OF_PLACE",
+                )
+            )
+
+            if out_supports_inplace:
+                test_cases.append(
+                    TestCase(
+                        inputs=[a_spec, b_spec],
+                        kwargs=None,
+                        output_spec=out_spec,
+                        comparison_target="out",
+                        tolerance=tol,
+                        description="GT - INPLACE(out)",
+                    )
+                )
+
+            if a_supports_inplace:
+                test_cases.append(
+                    TestCase(
+                        inputs=[a_spec, b_spec],
+                        kwargs={"out": 0},
+                        output_spec=None,
+                        comparison_target=0,
+                        tolerance=tol,
+                        description="GT - INPLACE(a)",
+                    )
+                )
+
+            if b_supports_inplace:
+                test_cases.append(
+                    TestCase(
+                        inputs=[a_spec, b_spec],
+                        kwargs={"out": 1},
+                        output_spec=None,
+                        comparison_target=1,
+                        tolerance=tol,
+                        description="GT - INPLACE(b)",
+                    )
+                )
+
+    return test_cases
+
+
+class OpTest(BaseOperatorTest):
+    """GT operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("GT")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.gt(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.gt(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/gumbel_softmax.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+from framework.utils import is_broadcast
+
+# Test cases format: (in_shape, in_strides_or_None, tau_or_None, hard_or_None, dim_or_None)
+
+_TEST_CASES_DATA = [
+    ((4, 10), None, 1.0, False, -1),
+    ((8, 20), (160, 1), 0.5, False, -1),
+    ((2, 5, 6), None, 1.5, True, 2),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 1e-3, "rtol": 1e-2},
+    infinicore.float32: {"atol": 1e-5, "rtol": 1e-4},
+    infinicore.bfloat16: {"atol": 1e-2, "rtol": 5e-2},
+}
+
+_TENSOR_DTYPES = [infinicore.float16, infinicore.bfloat16, infinicore.float32]
+
+
+def parse_test_cases():
+    """
+    gumbel_softmax: infinicore.nn.functional.gumbel_softmax(input, tau=1, hard=False, eps=1e-10, dim=-1)
+    """
+    test_cases = []
+
+    for data in _TEST_CASES_DATA:
+        shape = data[0]
+        in_strides = data[1] if len(data) > 1 else None
+        tau = data[2] if len(data) > 2 else 1.0
+        hard = data[3] if len(data) > 3 else False
+        dim = data[4] if len(data) > 4 else -1
+
+        for dtype in _TENSOR_DTYPES:
+            tolerance = _TOLERANCE_MAP.get(dtype, {"atol": 1e-5, "rtol": 1e-4})
+
+            input_spec = TensorSpec.from_tensor(shape, in_strides, dtype)
+
+            kwargs = {"tau": tau, "hard": hard, "dim": dim}
+
+            test_cases.append(
+                TestCase(
+                    inputs=[input_spec],
+                    kwargs=kwargs,
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tolerance,
+                    description=f"GumbelSoftmax - OUT_OF_PLACE",
+                )
+            )
+
+    return test_cases
+
+
+class OpTest(BaseOperatorTest):
+    """GumbelSoftmax operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("GumbelSoftmax")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.nn.functional.gumbel_softmax(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.nn.functional.gumbel_softmax(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/hardshrink.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+from framework.utils import is_broadcast
+
+# Test cases format: (in_shape, in_strides_or_None, lambd_or_None)
+
+_TEST_CASES_DATA = [
+    ((13, 4), None, None),
+    ((13, 4), (10, 1), 0.5),
+    ((8, 8, 8), None, 1.0),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 1e-3, "rtol": 1e-2},
+    infinicore.float32: {"atol": 1e-5, "rtol": 1e-4},
+    infinicore.bfloat16: {"atol": 1e-2, "rtol": 5e-2},
+}
+
+_TENSOR_DTYPES = [infinicore.float16, infinicore.bfloat16, infinicore.float32]
+
+
+def parse_test_cases():
+    """
+    hardshrink(input, lambd=0.5)
+    """
+    test_cases = []
+
+    for data in _TEST_CASES_DATA:
+        shape = data[0]
+        in_strides = data[1] if len(data) > 1 else None
+        lambd = data[2] if len(data) > 2 else 0.5
+
+        for dtype in _TENSOR_DTYPES:
+            tolerance = _TOLERANCE_MAP.get(dtype, {"atol": 1e-5, "rtol": 1e-4})
+            input_spec = TensorSpec.from_tensor(shape, in_strides, dtype)
+
+            kwargs = {}
+            if lambd is not None:
+                kwargs["lambd"] = lambd
+
+            test_cases.append(
+                TestCase(
+                    inputs=[input_spec],
+                    kwargs=kwargs,
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tolerance,
+                    description=f"Hardshrink - OUT_OF_PLACE",
+                )
+            )
+
+    return test_cases
+
+
+class OpTest(BaseOperatorTest):
+    """Hardshrink operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("Hardshrink")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.nn.functional.hardshrink(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.nn.functional.hardshrink(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/hardsigmoid.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+from framework.utils import is_broadcast
+
+# Test cases format: (in_shape, in_strides_or_None)
+
+_TEST_CASES_DATA = [
+    ((13, 4), None),
+    ((13, 4), (10, 1)),
+    ((8, 8, 8), None),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 1e-3, "rtol": 1e-2},
+    infinicore.float32: {"atol": 1e-5, "rtol": 1e-4},
+    infinicore.bfloat16: {"atol": 1e-2, "rtol": 5e-2},
+}
+
+_TENSOR_DTYPES = [infinicore.float16, infinicore.bfloat16, infinicore.float32]
+
+
+def parse_test_cases():
+    """hardsigmoid(input)
+
+    No inplace arg on functional API; only out-of-place tests.
+    """
+    test_cases = []
+
+    for data in _TEST_CASES_DATA:
+        shape = data[0]
+        in_strides = data[1] if len(data) > 1 else None
+
+        for dtype in _TENSOR_DTYPES:
+            tolerance = _TOLERANCE_MAP.get(dtype, {"atol": 1e-5, "rtol": 1e-4})
+            input_spec = TensorSpec.from_tensor(shape, in_strides, dtype)
+
+            test_cases.append(
+                TestCase(
+                    inputs=[input_spec],
+                    kwargs={},
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tolerance,
+                    description=f"Hardsigmoid - OUT_OF_PLACE",
+                )
+            )
+
+    return test_cases
+
+
+class OpTest(BaseOperatorTest):
+    """Hardsigmoid operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("Hardsigmoid")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.nn.functional.hardsigmoid(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.nn.functional.hardsigmoid(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/hardswish.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+from framework.utils import is_broadcast
+
+# Test cases format: (in_shape, in_strides_or_None)
+
+_TEST_CASES_DATA = [
+    ((13, 4), None),
+    ((8, 8), (64, 1)),
+    ((2, 3, 6), None),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 1e-3, "rtol": 1e-2},
+    infinicore.float32: {"atol": 1e-5, "rtol": 1e-4},
+    infinicore.bfloat16: {"atol": 1e-2, "rtol": 5e-2},
+}
+
+_TENSOR_DTYPES = [infinicore.float16, infinicore.bfloat16, infinicore.float32]
+
+
+def parse_test_cases():
+    """
+    hardswish(input)
+    """
+    test_cases = []
+
+    for data in _TEST_CASES_DATA:
+        shape = data[0]
+        in_strides = data[1] if len(data) > 1 else None
+
+        for dtype in _TENSOR_DTYPES:
+            tolerance = _TOLERANCE_MAP.get(dtype, {"atol": 1e-5, "rtol": 1e-4})
+            input_spec = TensorSpec.from_tensor(shape, in_strides, dtype)
+
+            test_cases.append(
+                TestCase(
+                    inputs=[input_spec],
+                    kwargs={},
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tolerance,
+                    description=f"Hardswish - OUT_OF_PLACE",
+                )
+            )
+
+    return test_cases
+
+
+class OpTest(BaseOperatorTest):
+    """Hardswish operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("Hardswish")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.nn.functional.hardswish(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.nn.functional.hardswish(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()

test/infinicore/ops/hardtanh.py

+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+import torch
+import infinicore
+from framework.base import BaseOperatorTest, TensorSpec, TestCase
+from framework.runner import GenericTestRunner
+from framework.utils import is_broadcast
+
+# Test cases format: (in_shape, in_strides_or_None, min_val_or_None, max_val_or_None)
+
+_TEST_CASES_DATA = [
+    ((13, 4), None, -1.0, 1.0),
+    ((13, 4), (10, 1), -0.5, 0.5),
+    ((8, 8, 8), None, -2.0, 2.0),
+]
+
+_TOLERANCE_MAP = {
+    infinicore.float16: {"atol": 1e-3, "rtol": 1e-2},
+    infinicore.float32: {"atol": 1e-5, "rtol": 1e-4},
+    infinicore.bfloat16: {"atol": 1e-2, "rtol": 5e-2},
+}
+
+_TENSOR_DTYPES = [infinicore.float16, infinicore.bfloat16, infinicore.float32]
+
+
+def parse_test_cases():
+    """hardtanh(input, min_val=-1.0, max_val=1.0, inplace=False)"""
+    test_cases = []
+
+    for data in _TEST_CASES_DATA:
+        shape = data[0]
+        in_strides = data[1] if len(data) > 1 else None
+        minv = data[2] if len(data) > 2 else -1.0
+        maxv = data[3] if len(data) > 3 else 1.0
+
+        input_supports_inplace = not is_broadcast(in_strides)
+
+        for dtype in _TENSOR_DTYPES:
+            tolerance = _TOLERANCE_MAP.get(dtype, {"atol": 1e-5, "rtol": 1e-4})
+            input_spec = TensorSpec.from_tensor(shape, in_strides, dtype)
+
+            kwargs = {"min_val": minv, "max_val": maxv}
+
+            test_cases.append(
+                TestCase(
+                    inputs=[input_spec],
+                    kwargs=kwargs,
+                    output_spec=None,
+                    comparison_target=None,
+                    tolerance=tolerance,
+                    description=f"Hardtanh - OUT_OF_PLACE",
+                )
+            )
+
+            if input_supports_inplace:
+                inplace_kwargs = {"min_val": minv, "max_val": maxv, "inplace": True}
+                test_cases.append(
+                    TestCase(
+                        inputs=[input_spec],
+                        kwargs=inplace_kwargs,
+                        output_spec=None,
+                        comparison_target=0,
+                        tolerance=tolerance,
+                        description=f"Hardtanh - INPLACE",
+                    )
+                )
+
+    return test_cases
+
+
+class OpTest(BaseOperatorTest):
+    """Hardtanh operator test with simplified implementation"""
+
+    def __init__(self):
+        super().__init__("Hardtanh")
+
+    def get_test_cases(self):
+        return parse_test_cases()
+
+    def torch_operator(self, *args, **kwargs):
+        return torch.nn.functional.hardtanh(*args, **kwargs)
+
+    # def infinicore_operator(self, *args, **kwargs):
+    #     """InfiniCore implementation (operator not yet available)."""
+    #     return infinicore.nn.functional.hardtanh(*args, **kwargs)
+
+
+def main():
+    """Main entry point"""
+    runner = GenericTestRunner(OpTest)
+    runner.run_and_exit()
+
+
+if __name__ == "__main__":
+    main()