Costa Huang
Costa Huang
-- UPDATE 7/7/2024: after chatting with @lewtun, we'd like to see if vLLM is willing to support https://github.com/vllm-project/vllm/issues/6189 officially before merging this PR as it may cause confusion for the...
The current PPO has a nice IMDB example that teaches the model to output more positive texts. However, we don't have something similar for PPOv2 yet. This PR introduces something...
This PR attempts to refactor and pull **all tokenization logic** out of the Trainer class. Having a separate tokenization process gives us higher visibility into what's being used in training,...
### 🚀 The feature, motivation and pitch Hi all, I was wondering if it's possible to do precise model device placement. For example, I would like to place the vLLM...
I get the following error despite installation flash attention was successful via `uv add flash-attn --no-build-isolation`. See https://github.com/astral-sh/uv/issues/6402. Is there anyway to honor the `no-build-isolation` when doing `uv sync`? I...
Would it be possible to install `uv` to a different location other than the default one? We are using a shared filesystem which lives at a difference place than `/home`.
I put the code here. To reproduce my work, `pip install ai2_olmo` and run ``` for beta in 0.05 do for lr in 3e-7 do python mason.py \ --cluster ai2/augusta-google-1...
Ehh something to put in the experimental folder scripts? Contains little snippets for generating
This PR removes the `--experiment_name` in favor of grabbing info from `--exp_name`. It's a bit less organized but I think it's fine because we should not rely on beaker search...