huggingface / huggingface/lighteval
[BUG] batch_size = auto:1 issue
- Dominant language
- Python
- Stars
- 2.5k
- Forks
- 555
- Avg merge
- 1d 6h
- Merged PRs (30d)
- 1
Description
## Disclaimer
It's not a Lighteval bug but an lm-evaluation-harness one. I'm just letting you know so you can investigate the batch size issue
## Describe the bug
I got the bug with setting the `batch_size = auto`. It originally occurred with the `glaiveai/Reflection-Llama-3.1-70B` evaluation, but it has repeated itself now with `rombodawg/Rombos-LLM-V2.5-Qwen-32b` (I ran a manual evaluation for this model to check for an another bug)
The error occurs when `batch_size = auto:1` is passed to the evaluation process, so it hangs the evaluation process. The longest I detected was two days with no "vital signs" in the log file. According to `htop` and `nvidia-smi` the process was indeed running.
This issue can be fixed by removing the `--batch_size auto` flag from the slurm file to set "batch_size=1" for the whole evaluation process.
## To Reproduce
1. I use [lm-evaluation-harness](https://github.com/huggingface/lm-evaluation-harness/tree/adding_all_changess), with `adding_all_changess` branch. My accelerate setup is this:
```
accelerate launch --multi_gpu --num_processes 4 \
-m lm_eval \
--model_args pretrained=${model},revision=${revision},trust_remote_code=False,dtype=bfloat16,parallelize=True \
--tasks leaderboard \
--batch_size auto \
--output_path eval_results \
--log_samples \
--hf_hub_log_args hub_results_org=open-llm-leaderboard,hub_repo_name=${model////__}-details,push_results_to_hub=True,push_samples_to_hub=True,public_repo=True
```
## Expected behavior
I expect that `batch_size=auto` will not block the evaluation process when it is automatically set to 1
p.s. Wow, what a cool automatic bug report template!
p.p.s. Please, ping me if something is unclear or I need to change the type of this issue. I decided to describe it as a bug
Contributor guide
No contributing guide indexed for this repository
Assessment
This issue has not been assessed yet.