lm-sys / lm-sys/FastChat

LLama 2 finetuning on multi-GPU with long context length

Open
#2,460 0 comments 1 reaction 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
39.5k
Forks
4.8k
PR merge metrics
No merged PRs in 30d

Description

I tried to finetune 13B model on 8A100 80gb GPU. When model is getting loaded It does DDP rather than TP or Zero. And since model copy is loaded on each GPU rather than it's TP. And when I traing on big context length I get OOM error.

What can be something that I am doing wrong? I have setup everything to latest versions and checked up installations.

When I run a 7b model in int4, context length to 4k and with LoRA then all GPU's fill up to half and training runs.

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

The report names no files, tests, or launch command. Start by reproducing the 13B, 8×A100, long-context setup and recording the model-loading and distributed-training configuration; done means the expected parallelism is confirmed and the reported out-of-memory behavior is explained or reproduced.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
distributed-systems, machine-learning
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
20/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.