Unable to run Llama3.1_(8B)-GRPO.ipynb on Titan V

[simusid]

Trying to run the notebook announced here on a recently updated Ubuntu 22.04 system. GPUs are two Titan V with Volta architecture, not Ampere.

The notebook runs fine unmodified all the way to trainer.train() where I get the error:
python: /project/lib/Analysis/Allocation.cpp:47: std::pair<llvm::SmallVector, llvm::SmallVector > mlir::triton::getCvtOrder(mlir::Attribute, mlir::Attribute): Assertion `!(srcMmaLayout && dstMmaLayout && !srcMmaLayout.isAmpere()) && "mma -> mma layout conversion is only supported on Ampere"' failed.

As stated I have Volta not Ampere, so I just thought "oh well, I'm out of luck" but per Reddit user yoracale "Mmm odd it should definitely work. I'm unsure why there is an error, would you be kind enough to make an issue on Github? Thanks "

[danielhanchen]

So we do support Tesla T4 (Turing CUDA 7.5) 100%.

I think V100s work (CUDA 7.0?) I think Titan V is also 7.0, so I thought it should work.