Skip to content

[DO NOT MERGE] fix: route muon-indivisible params to adamw (topology-dependent dion assert) #3438

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
hallerite wants to merge 1 commit into from
Closed

[DO NOT MERGE] fix: route muon-indivisible params to adamw (topology-dependent dion assert) #3438

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 11 additions & 0 deletions src/prime_rl/trainer/optim/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -122,13 +122,24 @@ def _create_muon_optimizer(
parallel_dims: ParallelDims,
lr: float | None = None,
) -> Optimizer:
if parallel_dims.dp_shard_enabled or parallel_dims.cp_enabled:
muon_mesh_size = parallel_dims.get_mesh("dp_shard_cp").size()
else:
muon_mesh_size = parallel_dims.world_mesh.size()

def muon_enabled(n, p):
if p.ndim < 2:
return False
if "lm_head" in n:
return False
if "embed_tokens" in n:
return False
# Muon distributes orthogonalization by sharding a parameter's leading dim
# across the mesh; dion asserts divisibility. Route indivisible params (e.g.
# small attention-gate projections) to adamw like the other muon-unsuitable
# params rather than crashing the step.
if p.shape[0] % muon_mesh_size != 0:
return False
return True

muon_params = []
Expand Down