Fix wrong tensor shape leading to words duplication

#11

The initial context tokens are 3 but the tensor shape is set to [1,2] - after the inference step the next token is set to the correct position but the shape is again with 1 less than it should be.

It leads to words doubling and seemingly hallucinations even with the provided audio sample.

Ready to merge
This branch is ready to get merged automatically.

Sign up or log in to comment