assign after copy shouldn't contig (#14847)

* assign after copy shouldn't contig

* fix assign copy

test/unit/test_assign.py

@@ -35,6 +35,14 @@ class TestAssign(unittest.TestCase):
     a.realize()
     np.testing.assert_allclose(b.numpy(), 0)
 
+  def test_assign_copy(self):
+    a = Tensor([1.,2,3], device="PYTHON")
+    c = Tensor.empty(3).assign(a.to(None))
+    # it should copy into the empty buffer
+    GlobalCounters.reset()
+    c.realize()
+    self.assertEqual(GlobalCounters.kernel_count, 1)
+
   def test_assign_add(self):
     for T in (1, 2, 10):#, 100): # this crashes in CI, not sure why
       x = Tensor([0]).realize()

tinygrad/schedule/indexing.py

@@ -167,6 +167,11 @@ def run_rangeify(tsink:UOp, debug:bool=False) -> tuple[UOp, IndexingContext]:
 
   # get ops to realize
   graph_rewrite(tsink, pm_generate_realize_map, ctx=rctx.realize_map, bottom_up=True, name="get realize")
+  # don't realize COPY/BUFFER_VIEW/ENCDEC when they are the direct source of ASSIGN — the ASSIGN target buffer is the output
+  for u in tsink.toposort():
+    if u.op is Ops.ASSIGN and u.src[1].op in {Ops.COPY, Ops.BUFFER_VIEW, Ops.ENCDEC} and u.src[1] in rctx.realize_map \
+       and not u.src[0].op_in_backward_slice_with_self(Ops.SHRINK, Ops.PERMUTE, Ops.FLIP, Ops.PAD):
+      del rctx.realize_map[u.src[1]]
 
   # get the consumer map
   with cpu_profile("consumer map in rangeify", "TINY"):