[wip] Distributed Scion/Muon #1630
Open
+3,049
−4
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
rakkit
requested review from
tianyu-l,
fegin,
wwwjn and
wconstab
as code owners
meta-cla
bot
added
the
CLA Signed
tianyu-l
requested changes
Aug 26, 2025
There was a problem hiding this comment.
Thanks for the PR on cutting-edge features!
I didn't read the papers so please forgive me if what I comment doesn't make sense.
I guess for "core" changes such as this one on optimizers, the recommended path is to first land in pytorch/pytorch, and then expose minimal interfaces to torchtitan. torchtitan shouldn't be a place to host core features.
cc @janeyx99 on interesting optimizer work
|
update: the init refactor is done, you can check the diff here that this optimizer is not aggressive at all, and not too much hack on 'components.optimizer'. Though we need add the configs. i added the debug configs, so can try it now. CONFIG_FILE="./torchtitan/experiments/distributed_scion/train_configs/debug_model.toml" NGPU=4 ./run_train.sh there is a "clean" version where I removed the code for logging, which can make the code easier to read and understand. a random test (be aware scion here using a higher LR, muon/scion allows us to train a model with high LR) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This is a distributed version of Scion or Modular Norm, muon is considered to be a variant of this by using explicit AdamW for LLM's embedding/output.
Works:
Missing
Need some extra work to adjust the EP changes for EP-[shard(1)] and ETP? And it's not working for multiple shared_experts.
CC @janEbert @ofivite