diff --git a/test/external/external_test_onnx_backend.py b/test/external/external_test_onnx_backend.py
index 73303d4f6a..c5cf06257f 100644
--- a/test/external/external_test_onnx_backend.py
+++ b/test/external/external_test_onnx_backend.py
@@ -168,10 +168,6 @@ if Device.DEFAULT == "METAL" or (OSX and Device.DEFAULT == "GPU"):
   backend_test.exclude('test_mish_cpu')
   backend_test.exclude('test_mish_expanded_cpu')
 
-if Device.DEFAULT == 'METAL':
-  # with default fast math enabled, padding -inf does not work
-  backend_test.exclude('test_MaxPool3d_stride_padding_cpu')
-
 # TODO: llvm has problems with inf
 if Device.DEFAULT in ['LLVM']:
   backend_test.exclude('test_isinf_cpu')
@@ -179,7 +175,7 @@ if Device.DEFAULT in ['LLVM']:
   backend_test.exclude('test_isinf_positive_cpu')
 
 # # TODO: problems with nan
-if Device.DEFAULT in ['LLVM', 'METAL']:
+if Device.DEFAULT in ['LLVM']:
   backend_test.exclude('test_isnan_float16_cpu')
   backend_test.exclude('test_isnan_cpu')
 
diff --git a/test/test_dtype_alu.py b/test/test_dtype_alu.py
index be27bd7e8c..35ce598874 100644
--- a/test/test_dtype_alu.py
+++ b/test/test_dtype_alu.py
@@ -69,10 +69,7 @@ def universal_test_unary(a, dtype, op):
   tensor_value = out.numpy()
   numpy_value = op[1](np.array([a]).astype(dtype.np))
   if dtype in dtypes_float:
-    atol = 2 if (Device.DEFAULT == "METAL" or getenv("PTX")) and op[0] == Tensor.sin else 1e-3
-    rtol = 2 if Device.DEFAULT == "METAL" and op[0] == Tensor.sin else 1e-4 if dtype == dtypes.float32 else 1e-2
-    # exp and log and sin are approximations (in METAL, the default fast-math versions are less precise)
-    np.testing.assert_allclose(tensor_value, numpy_value, atol=atol, rtol=rtol)
+    np.testing.assert_allclose(tensor_value, numpy_value, atol=1e-3, rtol=1e-2)
   else: np.testing.assert_equal(tensor_value, numpy_value)
   if op[0] != Tensor.reciprocal: # reciprocal is not supported in most backends
     op = [x for x in ast.lazyops if x.op in UnaryOps][0]
diff --git a/test/test_ops.py b/test/test_ops.py
index f03071999b..3a304697b8 100644
--- a/test/test_ops.py
+++ b/test/test_ops.py
@@ -351,12 +351,10 @@ class TestOps(unittest.TestCase):
     helper_test_op([()], lambda x: x/2)
     helper_test_op([()], lambda x: 2/x)
 
-  @unittest.skipIf(Device.DEFAULT in ["METAL", "WEBGPU"], "METAL has issues with -inf")
   def test_mul_naninf(self):
     helper_test_op([(45,65)], lambda x: x*math.inf)
     helper_test_op([(45,65)], lambda x: x*-math.inf)
     helper_test_op([(45,65)], lambda x: x*math.nan)
-  @unittest.skipIf(Device.DEFAULT in ["METAL", "WEBGPU"], "METAL has issues with -inf")
   def test_div_naninf(self):
     helper_test_op([(45,65)], lambda x: x/math.inf)
     helper_test_op([(45,65)], lambda x: x/-math.inf)
@@ -474,8 +472,7 @@ class TestOps(unittest.TestCase):
 
   def test_gelu(self):
     helper_test_op([(45,65)], lambda x: torch.nn.functional.gelu(x, approximate="tanh"), Tensor.gelu)
-    if not (CI and Device.DEFAULT == "METAL"):
-      helper_test_op([(45,65)], lambda x: torch.nn.functional.gelu(x, approximate="tanh"), Tensor.gelu, low=300, high=303)
+    helper_test_op([(45,65)], lambda x: torch.nn.functional.gelu(x, approximate="tanh"), Tensor.gelu, low=300, high=303)
     helper_test_op([(45,65)], lambda x: torch.nn.functional.gelu(x, approximate="tanh"), Tensor.gelu, low=-300, high=-297)
   def test_quick_gelu(self):
     helper_test_op([(45,65)], lambda x: x * torch.sigmoid(1.702 * x), Tensor.quick_gelu)