masahi edited a comment on pull request #5779: URL: https://github.com/apache/incubator-tvm/pull/5779#issuecomment-643509762
I don't have any objection regarding dtype handling in this PR. At the moment we assume fp32 everywhere by default, but to support doubles we need to somehow pass dtype information from user. I think we can pass an optional list of dtypes, corresponding to the entries in `input_shapes` (a required argument). If not passed we can fill in the dtype list with fp32. I think allowing `input_shape` to be optional is an orthogonal change to dtype issues that we can discuss elsewhere. I've already explained my reasoning, but to reiterate: making it optional is technically possible since Torch maintains names and traced Torch modules know the input shape. But it doesn't work for scripted modules. If names are omitted, user need to be aware of (or we need to explain) how to correctly figure out the names Torch maintains. Since this is not trivial and Torch may change the way they handle naming on a whim in the future, we shouldn't rely on naming chosen by Torch. On top of above, I think it is better to keep the API as close as possible to other frontends. They all require input names and shapes to be passed explicitly. ---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected]
