check: add test cases to the scripts/python_test.py one-click test script

Author: tianyuxbear

scripts/python_test.py

@@ -24,6 +24,16 @@ def run_tests(args):
         "rope.py",
         "sub.py",
         "swiglu.py",
+        "silu.py",
+        "div.py",
+        "logical_and.py",
+        "logical_or.py",
+        "equal.py",
+        "all_equal.py",
+        "relu_backward.py",
+        "gelu.py",
+        "gelu_backward.py",
+        "cross_entropy_loss_backward.py"
     ]:
         result = subprocess.run(
             f"python {test} {args} --debug", text=True, encoding="utf-8", shell=True

test/infiniop/gelu.py

@@ -19,6 +19,8 @@
     infiniopOperatorDescriptor_t,
     profile_operation,
     test_operator,
+    to_torch_dtype,
+    torch_device_map,
 )
 
 # ==============================================================================
@@ -101,8 +103,14 @@ def test(
         f"dtype:{InfiniDtypeNames[dtype]} inplace:{inplace}"
     )
 
-    new_tensor = torch.nn.functional.gelu(input.torch_tensor(), approximate="tanh")
-    output.update_torch_tensor(new_tensor)
+    # ans的shape对齐至input，而input可能存在广播维度
+    ans = torch.nn.functional.gelu(input.torch_tensor(), approximate="tanh")
+    # 利用add(+)计算的自动广播机制，确保ouput的torch_tensor与actual_tensor shape一致，以通过debug模式的shape检查
+    zero = torch.zeros(
+        *shape, dtype=to_torch_dtype(dtype), device=torch_device_map[device]
+    )
+    new_output = ans + zero
+    output.update_torch_tensor(new_output)
 
     if sync is not None:
         sync()

test/infiniop/libinfiniop/utils.py

@@ -110,7 +110,7 @@ def __init__(
             )
         else:
             raise ValueError("Unsupported mode")
-        
+
         if is_bool:
             self._torch_tensor = self._torch_tensor > 0.5
 
@@ -367,6 +367,11 @@ def debug(actual, desired, atol=0, rtol=1e-2, equal_nan=False, verbose=True):
         actual = actual.to(torch.float32)
         desired = desired.to(torch.float32)
 
+    # 如果是BOOL，全部转成FP32再比对
+    if actual.dtype == torch.bool or desired.dtype == torch.bool:
+        actual = actual.to(torch.float32)
+        desired = desired.to(torch.float32)
+
     print_discrepancy(actual, desired, atol, rtol, equal_nan, verbose)
     np.testing.assert_allclose(
         actual.cpu(), desired.cpu(), rtol, atol, equal_nan, verbose=True

test/infiniop/silu.py

@@ -19,6 +19,8 @@
     infiniopOperatorDescriptor_t,
     profile_operation,
     test_operator,
+    to_torch_dtype,
+    torch_device_map,
 )
 
 # ==============================================================================
@@ -101,7 +103,13 @@ def test(
         f"dtype:{InfiniDtypeNames[dtype]} inplace:{inplace}"
     )
 
-    new_output = torch.nn.functional.silu(input.torch_tensor())
+    # ans的shape对齐至input，而input可能存在广播维度
+    ans = torch.nn.functional.silu(input.torch_tensor())
+    # 利用add(+)计算的自动广播机制，确保ouput的torch_tensor与actual_tensor shape一致，以通过debug模式的shape检查
+    zero = torch.zeros(
+        *shape, dtype=to_torch_dtype(dtype), device=torch_device_map[device]
+    )
+    new_output = ans + zero
     output.update_torch_tensor(new_output)
 
     if sync is not None: