Thomas Viehmann

I think you'd hit the numerical stability issues pretty hard, too. But I understand CuDNN has an even shorter limit?

Hmm. I'm decidedly not seeing segfaults. Is that the same script you use? Did you use the exact same compiler between PyTorch and warp-ctc? Mixing versions gets you funny results....

Thank you for the PR. If I'm not mistaken, the latest release 0.4.1 still has the old way. I would prefer if we could keep compatibility with that - if...

Re PyTorch's own CTC: I've not published anything for that yet. You do have a logsoftmax? The native GPU implementation makes a similar precision / speed tradeoff as cudnn, but...

Trying to produce a minimal example, I got: ``` RuntimeError: IValue is not a Scalar The above operation failed in interpreter, with the following stack trace: at :3:12 @torch.jit.script def...

So the first error seems related to `prim::Constant` being declared fusible. As far as I understand, `prim::Constant` nodes are not usually declared fusible but instead are pulled in (but duplicated)...

Yes, sorry, I have been meaning to send a PR, but you simply need to remove `prim::Constant` fusion: https://github.com/pytorch/tvm/blob/7d3fa3fbd88855a1655c174cc43caac7fe6b954e/torch_tvm/operators.cpp#L417 I see, however, that #72 is moving away from the CustomGraphFuser...

This isn't PDF-specific. I'm using [nbsphinx](https://github.com/spatialaudio/nbsphinx) and it seems like the conversion to HTML (though rst?) has the same effect. Similarly it seems that the Python markdown package doesn't like...

I would prefer that we functionalize the RNG state handling within thunder and I wonder if this could be achieved with moderate effort (so the problem is similar to #145...

Yeah, as @tfogal , points out, I think something is up in the tests.