Aaron Orenstein
Aaron Orenstein
Tacotron2 causes massive loop unrolling resulting in very large graphs (26k nodes) which was causing inductor (and tracing itself) to choke. The unrolling size is controlled by the environment variable...
A small medley of fixes: - When validating sparse tensor indices don't check numel() if it's symbolic. - When validating sparse tensor indices if the indices are a FakeTensor then...
`Tensor.__repr__` calls functions which can perform logging which ends up logging `self` (with `__repr__`) causing an infinite loop. Detect this in `__repr__` and early-out instead of recursing. Another possible fix...
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #124545
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #125312 * __->__ #124226 * #124225 * #124224 * #124223 * #122911
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #125312 * #124226 * __->__ #124225 * #124224 * #124223 * #122911
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #125312 * #124226 * #124225 * __->__ #124224 * #124223 * #122911
When dispatching a fake tensor op we cache the result with `(op, args)` as the key. There are some args (such as one with a dynamic output shape) where the...
When constructing a `FakeTensorMode`, instead of immediately formatting a full stack trace, grab the traceback and only format it on demand. 4.2% FakeTensor perf win on the microbenchmark. ``` import...
We save and restore the DynamicLayerStack during frame eval but since fx graph has no way to express a try/finally we just assume it will happen. If we throw an...