[Relay] error in type inference#1984
Conversation
| // slow path with path compression. | ||
| TypeNode* root = this; | ||
| while (root->parent != nullptr) { | ||
| CHECK(root->parent != root) << "should not refer back to itself"; |
There was a problem hiding this comment.
Why not return root if you detect root->parent == root? This way, you won't throw (unnecessary?) exception (BTW the new unit test failed). The following snippet seems normal program to me and we probably should support it?
"""Program:
def f(x : Tensor[(10, 10), f32]) {
return log(x);
}
"""
from tvm import relay
a = relay.TypeVar("a")
x = relay.var("x", a)
sb = relay.ScopeBuilder()
f = relay.Function([x], x)
fx = relay.Call(f, [x])
assert relay.ir_pass.infer_type(x).checked_type == a
assert relay.ir_pass.infer_type(f).checked_type == relay.FuncType([a], a)
assert relay.ir_pass.infer_type(fx).checked_type == aThere was a problem hiding this comment.
yes, I did, and it is breaking invariant somewhere else (so did not get merged).
see #1983
There was a problem hiding this comment.
It is supposed to fail. This is more of a issue (but issue wont get ci).
There was a problem hiding this comment.
Oh OK, I'm interested in how such self reference structure is created cause the test case is trivial. FYI, I tried my fix and it correctly pass the above tests.
There was a problem hiding this comment.
That's good. I have no time to investigate into type inference internal right now, so we might have this on hold until I have more time slot (after 20days). IDK.
There was a problem hiding this comment.
Actually for this particular case, it's simply because we don't check type equality before we do unification(and we should always do the quality check before we try to unify them):
https://github.com/dmlc/tvm/blob/cbf4fdbbbf4993849342906722e776d332f2e97d/src/relay/pass/type_solver.cc#L71
I have a test test that failed master and works after the fix. Could you help review?
There was a problem hiding this comment.
This problem is due to there not being an occurs check: https://en.wikipedia.org/wiki/Occurs_check
|
fixed by #2012 |
4 participants
Thanks for contributing to TVM! Please refer to guideline https://docs.tvm.ai/contribute/ for useful information and tips. After the pull request is submitted, please request code reviews from Reviewers.