fix repeat_interleave bug & update benchmark setting for inplace operators (flagos-ai#1227)

Author: kiddyjinjin

benchmark/performance_utils.py

@@ -262,12 +262,7 @@ def set_gems(self, gems_op):
         self.gems_op = gems_op
 
     def get_latency(self, op, *args, **kwargs):
-        if self.is_inplace:
-            fn = lambda: op(
-                *[x.clone() if torch.is_tensor(x) else x for x in args], **kwargs
-            )
-        else:
-            fn = lambda: op(*args, **kwargs)
+        fn = lambda: op(*args, **kwargs)
         if self.is_backward:
             out = fn()
             dout = torch.randn_like(out)

tests/test_special_ops.py

@@ -1004,7 +1004,7 @@ def test_accuracy_repeat_interleave_self_int(shape, dim, dtype):
 
     ref_out = torch.repeat_interleave(ref_inp, repeats, dim)
     with flag_gems.use_gems():
-        res_out = torch.repeat_interleave(ref_inp, repeats, dim)
+        res_out = torch.repeat_interleave(inp, repeats, dim)
     gems_assert_equal(res_out, ref_out)
 
 
@@ -1019,7 +1019,7 @@ def test_accuracy_repeat_interleave_self_int_non_contiguous(shape, dim, dtype):
 
     ref_out = torch.repeat_interleave(ref_inp, repeats, dim)
     with flag_gems.use_gems():
-        res_out = torch.repeat_interleave(ref_inp, repeats, dim)
+        res_out = torch.repeat_interleave(inp, repeats, dim)
     gems_assert_equal(res_out, ref_out)