needed for mock

test/mockgpu/mockgpu.py

@@ -29,6 +29,8 @@ class TrackedMemoryView:
     self.mv = self.mv.cast('B').cast(new_type, **kwargs)
     return self
 
+  @property
+  def obj(self): return self.mv.obj
   @property
   def nbytes(self): return self.mv.nbytes
   def __len__(self): return len(self.mv)

tinygrad/device.py

@@ -201,7 +201,10 @@ class Buffer:
     return self._trace_num
 
   def _host_mv(self) -> memoryview|None:
-    if self.is_allocated() and hasattr(host:=self.get_storage()[1], 'mv'): return unwrap(host).view(fmt='B').mv
+    if self.is_allocated() and hasattr(host:=self.get_storage()[1], 'mv'):
+      mv = unwrap(host).view(fmt='B').mv
+      mv.obj._buffer = self  # raw ctypes views do not own their memory; keep the allocation alive for asynchronous copies
+      return mv
     if self.is_allocated() and hasattr(self.allocator, '_as_buffer'): return self.allocator._as_buffer(self._buf)
     return None
 

tinygrad/tensor.py

@@ -312,7 +312,7 @@ class Tensor(RandMixin):
     if not isinstance(data, UOp): raise RuntimeError(f"can't create Tensor from {data!r} with type {type(data)}")
 
     # data might be on a different device
-    self.uop:UOp = data if data.device is None or data.device == _device else data.copy_to_device(_device)
+    self.uop:UOp = data if data.device is None or data.device == _device else data.copy_to_device(_device).clone()
     # cast on the target device, the source may not hold the dtype (numpy has no fp8/bfloat16) or be able to compute it (DISK)
     if _dtype is not None: self.uop = self.uop.cast(_dtype)
 
@@ -545,9 +545,8 @@ class Tensor(RandMixin):
     """
     if self.uop.device is None: return self
     if (device:=canonicalize_device(device)) == self.device: return self
-    # a copy to disk wants to persist, so it inserts a clone: the disk buffer is the storage of the copied value
-    if isinstance(device, str) and device.startswith("DISK"): ret = Tensor(self.uop.clone(device))
-    else: ret = Tensor(self.uop.copy_to_device(device))
+    # The transfer owns its destination from construction; COPY itself only describes the transfer.
+    ret = Tensor(self.uop.copy_to_device(device).clone())
     if self.grad is not None: ret.grad = self.grad.to(device)
     return ret.is_param_(self.is_param)