gistfile1.txt

diff --git a/torch/testing/_internal/common_methods_invocations.py b/torch/testing/_internal/common_methods_invocations.py
index 938cb7dd97a..d3ac1369e6a 100644
--- a/torch/testing/_internal/common_methods_invocations.py
+++ b/torch/testing/_internal/common_methods_invocations.py
@@ -7443,6 +7443,57 @@ def reference_inputs_clone_contiguous(op, device, dtype, requires_grad, **kwargs
     yield SampleInput(a, kwargs={'memory_format': torch.channels_last_3d})
 
 
+def sample_inputs_copy(op_info, device, dtype, requires_grad, **kwargs):
+    """Sample inputs for copy and copy_ operations.
+    
+    copy(dst_tensor, src_tensor, non_blocking=False) copies data from src to dst.
+    For copy_: dst_tensor.copy_(src_tensor, non_blocking=False)
+    """
+    make_arg = partial(make_tensor, dtype=dtype, device=device, requires_grad=requires_grad)
+    
+    # Basic same-shape copies
+    dst = make_arg((S, M))
+    src = make_arg((S, M))
+    yield SampleInput(dst, args=(src,))
+    
+    # Scalar cases
+    dst_scalar = make_arg(())
+    src_scalar = make_arg(())
+    yield SampleInput(dst_scalar, args=(src_scalar,))
+    
+    # Broadcasting cases - src can be broadcast to dst shape
+    dst_broadcast = make_arg((S, M, L))
+    src_broadcast1 = make_arg((1, M, L))  # broadcast in first dim
+    src_broadcast2 = make_arg((S, 1, L))  # broadcast in second dim  
+    src_broadcast3 = make_arg((1,))       # broadcast scalar to tensor
+    yield SampleInput(dst_broadcast, args=(src_broadcast1,))
+    yield SampleInput(dst_broadcast, args=(src_broadcast2,))
+    yield SampleInput(dst_broadcast, args=(src_broadcast3,))
+    
+    # Different dtypes that are compatible
+    if dtype != torch.bool:  # bool has limited conversion support
+        if dtype.is_floating_point:
+            # Float to float conversions
+            other_dtype = torch.float64 if dtype != torch.float64 else torch.float32
+        elif dtype.is_complex:
+            # Complex to complex conversions  
+            other_dtype = torch.complex128 if dtype != torch.complex128 else torch.complex64
+        else:
+            # Int to int conversions
+            other_dtype = torch.int64 if dtype != torch.int64 else torch.int32
+            
+        if other_dtype != dtype:
+            dst_convert = make_arg((S, M))
+            src_convert = make_tensor((S, M), dtype=other_dtype, device=device, requires_grad=False)
+            yield SampleInput(dst_convert, args=(src_convert,))
+    
+    # Non-blocking parameter
+    dst_nonblock = make_arg((S, M))
+    src_nonblock = make_arg((S, M))
+    yield SampleInput(dst_nonblock, args=(src_nonblock,), kwargs={'non_blocking': True})
+    yield SampleInput(dst_nonblock, args=(src_nonblock,), kwargs={'non_blocking': False})
+
+
 def sample_inputs_sum_to_size(op_info, device, dtype, requires_grad, **kwargs):
     make_arg = partial(make_tensor, dtype=dtype, device=device, requires_grad=requires_grad)
 
@@ -12847,6 +12898,31 @@ op_db: list[OpInfo] = [
                DecorateInfo(unittest.expectedFailure, 'TestCommon', 'test_numpy_ref'),
                DecorateInfo(unittest.expectedFailure, 'TestCommon', 'test_numpy_ref_mps'),
            ),),
+    OpInfo('copy_',
+           # copy_ is the in-place tensor method: tensor.copy_(src, non_blocking=False)
+           op=lambda x, src, non_blocking=False: x.copy_(src, non_blocking=non_blocking),
+           inplace_variant=True,
+           dtypes=all_types_and_complex_and(torch.bool, torch.bfloat16, torch.float16, torch.chalf),
+           sample_inputs_func=sample_inputs_copy,
+           supports_forward_ad=True,
+           supports_fwgrad_bwgrad=True,
+           supports_out=False,
+           skips=(
+               # copy_ requires compatible dtypes/devices which may not be satisfied in all test cases
+               DecorateInfo(unittest.expectedFailure, 'TestCommon', 'test_variant_consistency_eager'),
+           ),),
+    OpInfo('copy',
+           # aten::copy functional version: copy(dst_tensor, src_tensor, non_blocking=False) 
+           op=torch.ops.aten.copy.default,
+           dtypes=all_types_and_complex_and(torch.bool, torch.bfloat16, torch.float16, torch.chalf),
+           sample_inputs_func=sample_inputs_copy,
+           supports_forward_ad=True,
+           supports_fwgrad_bwgrad=True,
+           supports_out=False,
+           skips=(
+               # copy requires compatible dtypes/devices which may not be satisfied in all test cases
+               DecorateInfo(unittest.expectedFailure, 'TestCommon', 'test_variant_consistency_eager'),
+           ),),
     OpInfo('contiguous',
            op=lambda x, *args, **kwargs: x.contiguous(*args, **kwargs),
            dtypes=all_types_and_complex_and(torch.bool, torch.bfloat16, torch.float16, torch.chalf),