[Bug] concat([x], axis=1) return random results

[ganler]

import tvm
from tvm import relay
import numpy as np

x1_shape = (2, 1)

x1 = relay.var("x1", shape=x1_shape, dtype='int64')
concatenate = relay.op.concatenate([x1], axis=1)

f = relay.Function([x1], concatenate)

x1_val = np.array([[6], [5]], dtype='int64')

op_res = relay.create_executor('graph', device=tvm.cpu(), target='llvm').evaluate(f)(x1_val)
print(op_res)
print(f'should be equal to {x1_val}')

Expected behavior

op_res should be same as [[6], [5]].

Actual behavior

1st try:

2nd try:

3rd try:

Environment

ed638ef on Ubuntu 20.04.

cc: @masahi

[lileidev]

I tried the sample code to reproduce, but get bellow error:
Traceback (most recent call last):
File "5.py", line 1, in
import tvm
File "/home/larry/.wine/drive_d/tvm/python/tvm/init.py", line 42, in
from .ir import IRModule
File "/home/larry/.wine/drive_d/tvm/python/tvm/ir/init.py", line 33, in
from .memory_pools import (
File "/home/larry/.wine/drive_d/tvm/python/tvm/ir/memory_pools.py", line 38, in
class PoolInfoProperties(Object):
File "/home/larry/.wine/drive_d/tvm/python/tvm/_ffi/registry.py", line 69, in register
check_call(_LIB.TVMObjectTypeKey2Index(c_str(object_name), ctypes.byref(tidx)))
File "/home/larry/.wine/drive_d/tvm/python/tvm/_ffi/base.py", line 348, in check_call
raise get_last_ffi_error()
tvm._ffi.base.TVMError: Traceback (most recent call last):
81: 0xffffffffffffffff
80: _start
79: __libc_start_main
at ../csu/libc-start.c:308
78: Py_BytesMain
77: Py_RunMain
76: PyRun_SimpleFileExFlags
75: 0x000000000067e470
74: 0x000000000067e3ce
73: 0x000000000067e350
72: PyEval_EvalCode
71: _PyEval_EvalCodeWithName
70: _PyEval_EvalFrameDefault
69: PyImport_ImportModuleLevelObject
68: _PyObject_CallMethodIdObjArgs
67: 0x00000000005f4400
66: _PyFunction_Vectorcall
65: _PyEval_EvalFrameDefault
64: _PyFunction_Vectorcall
63: _PyEval_EvalFrameDefault
62: _PyFunction_Vectorcall
61: _PyEval_EvalFrameDefault
60: _PyFunction_Vectorcall
59: _PyEval_EvalFrameDefault
58: _PyFunction_Vectorcall
57: _PyEval_EvalCodeWithName
56: _PyEval_EvalFrameDefault
55: PyVectorcall_Call
54: 0x00000000005c4a3f
53: 0x00000000006003a3
52: PyEval_EvalCode
51: _PyEval_EvalCodeWithName
50: _PyEval_EvalFrameDefault
49: PyImport_ImportModuleLevelObject
48: _PyObject_CallMethodIdObjArgs
47: 0x00000000005f4400
46: _PyFunction_Vectorcall
45: _PyEval_EvalFrameDefault
44: _PyFunction_Vectorcall
43: _PyEval_EvalFrameDefault
42: _PyFunction_Vectorcall
41: _PyEval_EvalFrameDefault
40: _PyFunction_Vectorcall
39: _PyEval_EvalFrameDefault
38: _PyFunction_Vectorcall
37: _PyEval_EvalCodeWithName
36: _PyEval_EvalFrameDefault
35: PyVectorcall_Call
34: 0x00000000005c4a3f
33: 0x00000000006003a3
32: PyEval_EvalCode
31: _PyEval_EvalCodeWithName
30: _PyEval_EvalFrameDefault
29: PyImport_ImportModuleLevelObject
28: _PyObject_CallMethodIdObjArgs
27: 0x00000000005f4400
26: _PyFunction_Vectorcall
25: _PyEval_EvalFrameDefault
24: _PyFunction_Vectorcall
23: _PyEval_EvalFrameDefault
22: _PyFunction_Vectorcall
21: _PyEval_EvalFrameDefault
20: _PyFunction_Vectorcall
19: _PyEval_EvalFrameDefault
18: _PyFunction_Vectorcall
17: _PyEval_EvalCodeWithName
16: _PyEval_EvalFrameDefault
15: PyVectorcall_Call
14: 0x00000000005c4a3f
13: 0x00000000006003a3
12: PyEval_EvalCode
11: _PyEval_EvalCodeWithName
10: _PyEval_EvalFrameDefault
9: _PyFunction_Vectorcall
8: _PyEval_EvalCodeWithName
7: _PyEval_EvalFrameDefault
6: _PyObject_MakeTpCall
5: 0x00007fb5add2a9db
4: _ctypes_callproc
3: 0x00007fb5aeaba409
2: 0x00007fb5aeabaff4
1: TVMObjectTypeKey2Index
0: tvm::runtime::Object::TypeKey2Index(std::__cxx11::basic_string<char, std::char_traits, std::allocator > const&)
File "/home/larry/.wine/drive_d/tvm/src/runtime/object.cc", line 165
TVMError:

An error occurred during the execution of TVM.
For more information, please see: https://tvm.apache.org/docs/errors.html

Check failed: (it != type_key2index_.end()) is false: Cannot find type ir.PoolInfoProperties. Did you forget to register the node by TVM_REGISTER_NODE_TYPE ?

[ganler]

I doubled checked that but I can still reproduce it (in the commit version I attached which is very recent). I think the operators used here are very common that it is not likely to be not registering them. You can try more recent TVM if you want to reproduce it.

[masahi]

what happens if you revert the commits #11341 and #11800?

[ganler]

Just tried the commit right before #11800, this issue is still there. am trying #11341 ~1 now.

[ganler]

@masahi So a1d95ec (#11341 ~1) is correct as follows:

I also directly tried #11341 it fails:

Therefore, it is highly likely that this bug is introduced by #11341.

cc: @shtinsa

[masahi]

@shtinsa Can you take a look?

[shtinsa]

Hello, the fix is in this PR: #11907