Masaki Kozuki

Results 167 comments of Masaki Kozuki

`xentropy_cuda` is not compiled with `--cuda_ext` option, but `--xentropy`: https://github.com/NVIDIA/apex#custom-ccuda-extensions-and-install-options

that sounds right, would you mind opening a pull request?

Personally I recommend using `--no-build-isolation` as even `packaging` is installed, I guess it would be a bit tricky to install the same torch as in the environment into a build...

hmm, I haven't found myself in the same situation. what if the latest pip and multiple `--config-settigns`?

one way (I wouldn't recommend though) to dodge pyproject.toml dependency management could be to use `python setup.py install ---cpp_ext --cuda_ext ...` to avoid pip being called

A tidy reproducible code is shared, as in the description. I confirmed that we can reproduce the error (with the slightly different KeyError message, with high probability)

Alternative: have a check in test and skip accordingly

https://github.com/Lightning-AI/lightning-thunder/pull/2633 would be related or could be a reference point

> PyTorch decorates the `_init_group` method of every optimizer class with a wrapper that disables Dynamo to trace it. Thus `_init_group` is always executed in the eager mode. I'd expect...

Can you check llama 4, deep seek v3.1, and qwen3 next? I understand the models are quite heavy though