Skip to content

Fix layerwise SGDM matrix decay grouping #175

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

Merged
charlesmartin14 merged 2 commits into from
Oct 18, 2026
Merged

Fix layerwise SGDM matrix decay grouping #175

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
171a907
Separate SGDM matrix decay groups
charlesmartin14 Oct 18, 2026
cd25d22
Test separate SGDM decay groups
charlesmartin14 Oct 18, 2026
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
5 changes: 4 additions & 1 deletion baseline/gpt2_small/muon_speedrun/optim.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -140,8 +140,11 @@ def make_stock_optimizers(model, rt, kind):
else:
role = 'other'
multiplier = SGD_OTHER_MULTIPLIER
# Keep matrix decay and bias/LayerNorm no-decay groups separate even
# when they share a blockwise LR multiplier.
decay = SGD_WEIGHT_DECAY if p.ndim >= 2 else 0.0
grouped[(role, multiplier, decay)].append(p)
decay_role = role + ('_matrices' if p.ndim >= 2 else '_bias_and_layernorm')
grouped[(decay_role, multiplier, decay)].append(p)
groups = []
for (role, multiplier, decay), params in grouped.items():
peak = SGD_PEAK_LR * multiplier
Expand Down
7 changes: 4 additions & 3 deletions baseline/gpt2_small/muon_speedrun/test_sgd_momentum.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,9 +38,10 @@ def test_tied_parameter_once_and_classical_momentum(self):
self.assertEqual(len(params), len(set(map(id, params))))
self.assertEqual(set(map(id, params)), set(map(id, model.parameters())))
groups = {g['role']:g for g in opt.param_groups}
self.assertAlmostEqual(groups['block_0']['lr_multiplier'], SGD_BLOCK_MULTIPLIERS[0])
self.assertEqual(groups['block_0']['weight_decay'], SGD_WEIGHT_DECAY)
self.assertEqual(groups['other']['weight_decay'], 0.0)
self.assertAlmostEqual(groups['block_0_matrices']['lr_multiplier'], SGD_BLOCK_MULTIPLIERS[0])
self.assertEqual(groups['block_0_matrices']['weight_decay'], SGD_WEIGHT_DECAY)
self.assertEqual(groups['block_0_bias_and_layernorm']['weight_decay'], 0.0)
self.assertEqual(groups['other_bias_and_layernorm']['weight_decay'], 0.0)
self.assertTrue(all(g["momentum"] == SGD_MOMENTUM for g in opt.param_groups))
self.assertTrue(all(not g["nesterov"] for g in opt.param_groups))
self.assertEqual(optimizer_metadata(opt)["class"], "torch.optim.SGD")
Expand Down
Loading