slyubomirsky opened a new issue, #13619:
URL: https://github.com/apache/tvm/issues/13619
Consider the following Relay program:
```python
import tvm
from tvm import relay
from tvm.relay.prelude import Prelude
p = prelude.Prelude().mod
list_var = p.get_global_type_var("List")
list_data = p.type_definitions[list_var]
nil_ctor = None
for ctor in list_data.constructors:
if ctor.name_hint == "Nil":
nil_ctor = ctor
assert nil_ctor != None
p["main"] = relay.Function(
[],
relay.Match(
relay.Call(nil_ctor, [], type_args=[relay.TensorType((),
dtype="int64")]),
[relay.Clause(relay.PatternWildcard(), relay.const(1,
dtype="int64"))],
),
ret_type=relay.TensorType((), dtype="int64"),
)
```
If we print the module using the pretty-printer, we obtain the following
(omitting parts of the module that aren't relevant):
```
type List[A] {
Cons(A, List[A]),
Nil,
}
# snip
def @main() -> int64 {
%10 = Nil;
match (%10) {
_ => {
1i64
},
}
}
```
It all looks as it should, right? However, if we parse back the module and
compare it to the original, the comparison fails:
```python
text = p.astext(show_meta_data=True)
with open(os.path.join(out_dir, "prog.rly"), "w") as fp:
fp.write(text)
new_mod = tvm.parser.fromtext(text)
assert tvm.ir.structural_equal(p["main"], new_mod["main"],
map_free_vars=True)
```
The reason for this is the type args field on the call to `Nil`.
```python
print(new_mod["main"].body.data.type_args) # prints "[IncompleteTypeNode(0,
0x556f6335e570)]"
```
The text format is not serializing the type args field. In this case, type
inference cannot unambiguously infer the type arg for the constructor (indeed,
there is not enough information for it).
(Bug discovered in the process of fuzzer development.)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
