hiyufan commented on PR #20376: URL: https://github.com/apache/tvm/pull/20376#issuecomment-5715999640
The first CI run went red on `cpu` — `test_prod[float32]` and `test_prod[bool]` in both frontend test files — and that was this PR, fixed in `9cb5ff9`. `_prod` emitted `R.astype(x, dtype=...)` for an explicit `dtype=` even when it already matched the input, so `torch.prod(x, dtype=torch.float32)` on a float32 input lowered to `astype` + `prod` where the test expects a bare `R.prod`. It now skips the cast when the dtype is unchanged; the int64 promotion for integer inputs without a `dtype=` is untouched (`test_prod_dim_and_integer_accumulation` still pins the `astype int64`). I should say why my before/after check missed it: `test_prod` cannot run in my local environment (Python 3.14, the TVMScript parser does not see the parametrised `relax_dtype` name), so it is in the "pre-existing failure" set on `main` and on this branch alike, and a same-set comparison is blind to a test that never ran. CI was the first place it could fail. Force-pushed the amended commit; the body's verification section otherwise stands. -- 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]
