actually fix early alloc

tinygrad/llops/ops_gpu.py

@@ -198,7 +198,6 @@ class GPUBuffer(ExplicitExecAST):
 
   def __init__(self, shape:Union[ShapeTracker, Tuple[int, ...]], hostbuf:Optional[GPUBuffer]=None, backing:Optional[np.ndarray]=None):
     super().__init__(shape, hostbuf)
-    if hostbuf is not None and hostbuf._backing is None: hostbuf.cl  # does the allocation of unbacked buffer, pylint: disable=W0104
     self._buf : Optional[CLBuffer] = hostbuf._buf if hostbuf is not None else None
     self._base_shape : Tuple[int, ...] = hostbuf._base_shape if hostbuf is not None else self.shape
     self._backing : Optional[np.ndarray] = hostbuf._backing if hostbuf is not None else backing

tinygrad/ops.py

@@ -118,6 +118,7 @@ class ASTKernel:
 
     # create the buffer we are returning (as the same type as the input buffers) and add it as the first buffer
     self.ret = type(self.bufs[0])(output_shape if output_shape else self.info.shape)
+    if hasattr(self.ret, "cl"): self.ret.cl  # does the allocation of unbacked buffer, pylint: disable=W0104
     self.bufs = [type(self.ret)(self.info.shape, hostbuf=self.ret)] + self.bufs
 
     # check valid AST kernel