Integration test failing

[SimJeg]

One of the integration test is now failing

FAILED tests/integration/test_ruler.py::TestRuler::test_ruler_is_correct[0.1-quantized-press_dict0] - assert '8973915' in 'The special magic number mentioned in the text for **"troubled-midnight"** (which appears to be a typo or mishearing of **"troubled-midnight"**) is:\n\n**897391'

Also the CI/CD sometimes takes forever to build flash-attn.

[SimJeg]

Breaking changes are coming in v5: huggingface/transformers@cb39692

[maxjeblick]

Thanks!
So this issue is really about transformers v5 compatiblity. I can work on this item.

[SimJeg]

v5 has been released today, so it's another issue

[SimJeg]

Removed quantization tests in #177, needs to be refactored

[SimJeg]

I could successfully run the integration tests with quantization locally. The issue was with the installation of optimum-quanto. Given that this installation requires nvcc, and that installing it in the CI/CD can require up to 5min, we will now skip these tests.

For the record, it can be useful to set TORCH_CUDA_ARCH_LIST to "8.0 8.6 9.0" to avoid compilation errors with optimum-quanto for recent GPU architectures.