Wrong result when using new numpy ffi in deferred compute

[hgt312]

Find it in CI result of #17958.

Reproduce:

Sample code, use master branch

import mxnet as mx
import mxnet._deferred_compute as dc
from mxnet import np, npx
npx.set_np()
with dc.context():
    a = np.ones((2, 2))
    b = np.tril(a, 1)
    c = np.tril(a, -1)

sym = dc.get_symbol([b, c], sym_cls=mx.sym.np._Symbol)
res = sym.bind(mx.context.current_context(), args={}).forward()
res

Results:

[
[[1. 1.]
 [1. 1.]]
<NDArray 2x2 @cpu(0)>,
[[1. 1.]
 [1. 1.]]
<NDArray 2x2 @cpu(0)>]

If replace _api_internal.tril by _npi.tril, the result is right.

@leezu Can you take a look at this?

[hgt312]

@hzfan

[leezu]

Thanks for reporting the issue. I can confirm that disabling the new FFI for tril via below patch avoids the issue. I'll take a look why the two FFIs behave differently

diff --git a/python/mxnet/ndarray/numpy/_op.py b/python/mxnet/ndarray/numpy/_op.py
index 4820f56f3..8ff7ba073 100644
--- a/python/mxnet/ndarray/numpy/_op.py
+++ b/python/mxnet/ndarray/numpy/_op.py
@@ -2051,7 +2051,7 @@ def tril(m, k=0):
            [ 7.,  8.,  0.],
            [10., 11., 12.]])
     """
-    return _api_internal.tril(m, k)
+    return _npi.tril(m, k)


 def _unary_func_helper(x, fn_array, fn_scalar, out=None, **kwargs):

[haojin2]

@leezu Have you been able to find the root cause for this bug?

[leezu]

@haojin2 sorry, I didn't get to dive deep on this issue yet. I'll work on this during our bug bash in the coming week.

[leezu]

It's due to common expression elimination. As workaround one could set MXNET_ELIMINATE_COMMON_EXPR=0.

https://github.com/apache/incubator-mxnet/blob/ab4f7f6a7335e88034edcf61402aec170cdca5fd/src/executor/eliminate_common_expr_pass.cc#L59-L94

For the old FFI, n->attrs.dict != m->attrs.dict and the two nodes are preserved.
But in the new FFI, n->attrs.dict == m->attrs.dict and the np.tril(a, -1) is replaced by np.tril(a, 1).

For the new FFI, the reason is

https://github.com/apache/incubator-mxnet/blob/7bef85ecbf9cb064d45479efa5cc46c6f8a4e947/src/api/operator/utils.h#L52-L57

only parses attributes if autograd is enabled, whereas the old FFI always parses attributes. This was an oversight when I rebased the DC PR on top of the FFI update PR.