Alexandre Strube

4.1.4 on GCC-11.3.0 suffers from the same problem. If I use `configopts += '--with-slurm --with-pmi=/usr/include/slurm --with-pmi-libdir=/usr/lib/x86_64-linux-gnu'` as @SethosII mentions, it goes further, but then much later complain again that can't...

@wanbo432503 have you tried @Go0day 's suggestion? This has been a bit, I will close it. If you still need help, please reopen!

@JarringBye instead of setting --worker-address to 0.0.0.0, how about you set it to 127.0.0.1? Do you still have this issue? It is not a bug of fastchat.

@chentao169 this is not an issue of Fastchat, right? Maybe you can have a look at this one: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard Let's close this issue?

@ww-tt hey, any news? Have you solved it? We have been using vicuna for a bit here, works fine in general.

While this is good, it's not really an issue of fastchat, right? I would recommend you have a look at this colab on how to fine-tune llama2 as an example:...

We can't do much about this, I fear. If the gpu has not enough memory, it has not enough memory...

@zhisbug just a reminder that the developer of Flash Attention gave up on the V100s. https://github.com/Dao-AILab/flash-attention/issues/148#issuecomment-1573216640

> The authors of Flash Attention have also developed the triton-based implementation (https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/flash_attn_triton.py). How about replacing their original implementation with the triton-based? That would force everyone to use PyTorch 2.0,...

@jshin49 doesn't seem so, look at the comment before.