THUDM / THUDM/slime

system blocking when do megatron update load weights

Open
#674 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
8.5k
Forks
1.3k
Avg merge
5h 36m
Merged PRs (30d)
22

Description

https://github.com/THUDM/slime/blob/38824116e6fd3989029ca30582b9c323f5296238/slime/backends/megatron_utils/update_weight_utils.py#L475

During the training of GLM-AIR with colocate mode, we encountered a system freeze, and after a period of time, a watchdog timeout error would occur.
I'm not sure if this code snippet is the cause, since it could potentially lead to processes in this group using different src_ranks during toch.distribute.broacast?

Contributor guide

Open the contributing guide

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

Start with slime/backends/megatron_utils/update_weight_utils.py at line 475 and inspect the surrounding weight-loading path and torch.distributed broadcast usage. Reproduce the GLM-AIR colocate-mode training freeze if possible, then determine whether inconsistent source ranks cause the watchdog timeout and verify the resolved behavior under that workload.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
distributed-systems, machine-learning
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.