javierdejesusda commented on code in PR #19679:
URL: https://github.com/apache/tvm/pull/19679#discussion_r3368198037
##########
python/tvm/relax/frontend/torch/base_fx_graph_translator.py:
##########
@@ -391,6 +391,17 @@ def _log_softmax(self, node: fx.Node) -> relax.Var:
dim = node.args[1] if len(node.args) > 1 else node.kwargs.get("dim",
-1)
return self.block_builder.emit(relax.op.nn.log_softmax(x, dim))
+ def _logical_and(self, node: fx.Node) -> relax.Var:
+ lhs = self.env[node.args[0]]
+ rhs = self.env[node.args[1]]
+ # torch.logical_and accepts any dtype (treating nonzero as True) and
returns bool, but
+ # relax.op.logical_and requires boolean inputs, so cast non-bool
inputs to bool first.
+ if lhs.struct_info.dtype != "bool":
+ lhs = self.block_builder.emit(relax.op.astype(lhs, "bool"))
+ if rhs.struct_info.dtype != "bool":
+ rhs = self.block_builder.emit(relax.op.astype(rhs, "bool"))
Review Comment:
I kept the direct `struct_info.dtype` check to stay
consistent with the existing `_logical_not` converter just below, which
accesses
`struct_info.dtype` the same way without an `isinstance` check. Both
operands come
from `torch.logical_and`, which only accepts tensor inputs, so their
`struct_info`
is always a `TensorStructInfo` and the access is safe. Adding the guard here
would
make this converter diverge from `_logical_not`, so I left it consistent
with the
surrounding code.
--
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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
