trl VS sparsegpt

I am using TRL for RLHF fine-tuning the Llama-2-7B model and getting an OOM error (even with batch_size=1). If anyone used TRL for RLHF can please tell me what I am doing wrong? Code details can be found in the GitHub issue.

SFTtrainer from huggingface

For lora - PEFT seems to work. I don't have patience to wait 5 hours, but modifying this example seems to work. You don't even need to modify that much, as their model just as neo-x uses query_key_value name for self-attention.

They have examples of making GPT output more positive (code) by using a sentiment model as reward. There are other examples about reducing toxicity, summarization here: https://github.com/lvwerra/trl/tree/main/examples . Should be fairly simple to modify the sentiment example and try the calculator reward you mentioned above.

You can use this -> https://github.com/lvwerra/trl/blob/main/examples/sentiment/scripts/gpt-neox-20b_peft/merge_peft_adapter.py

Just the hh directly. From the results it seems like it might possibly be enough but I might also try instruction tuning then running the whole process from that base. I will also be running the reinforcement learning by using a Lora using this as an example https://github.com/lvwerra/trl/tree/main/examples/sentiment/scripts/gpt-neox-20b_peft

This package is pretty simple to use! https://github.com/lvwerra/trl

SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot (https://arxiv.org/abs/2301.00774)

There is this : https://github.com/IST-DASLab/sparsegpt

Check the paper here, it's intersting: https://arxiv.org/abs/2301.00774

There's no chance that we've peeked from a bang for buck sense - we still haven't adequately investigated sparse networks.

Relevantish: https://arxiv.org/abs/2301.00774

The fact that we can reach those levels of sparseness with pruning also indicates that we're not doing a very good job of generating the initial network conditions.

Being able to come up with trainable initial settings for sparse networks across different topologies is hard, but given that we've had a degree of success with pre-trained networks, pre-training and pre-pruning might also allow for sparse networks with minimally compromised learning capabilities.

If it's possible to pre-train composable network modules, it might also be feasible to define trainable sparse networks with significantly relaxed topological constraints.

Training uses gradient descent, so you want to have good precision during that process. But once you have the overall structure of the network, https://arxiv.org/abs/2210.17323 (GPTQ) showed that you can cut down the precision quite a bit without losing a lot of accuracy. It seems you can cut down further for larger models. For the 13B Llama-based ones, going below 5 bit per parameter is noticeably worse, but for 30B models you can do 4 bits.

The same group did another paper https://arxiv.org/abs/2301.00774 which shows that in addition to reducing the precision of each parameter, you can also prune out a bunch of parameters entirely. It's harder to apply this optimization because models are usually loaded into RAM densely, but I hope someone figures out how to do it for popular models.