Skip to content

fix unused parameters error for distributed training with multihead #1311

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

Open
bernstei wants to merge 2 commits into
base: main
Choose a base branch
from
Open

fix unused parameters error for distributed training with multihead #1311

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5b4809d
when copying readouts from foundations models, don't copy if tensor h…
Dec 18, 2025
70052dd
Merge branch 'main' into unused_parameters_readout_bias
bernstei May 7, 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
40 changes: 17 additions & 23 deletions mace/tools/finetuning_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -195,11 +195,10 @@ def load_foundations_elements(

if load_readout:
# Transferring readouts
for i, readout in enumerate(model.readouts):
for readout, foundation_readout in zip(model.readouts, model_foundations.readouts):
if readout.__class__.__name__ == "LinearReadoutBlock":
model_readouts_zero_linear_weight = readout.linear.weight.clone()
model_readouts_zero_linear_weight = (
model_foundations.readouts[i]
foundation_readout
.linear.weight.view(num_channels_foundation, -1)
.repeat(1, len(model_heads))
.flatten()
Expand All @@ -217,17 +216,16 @@ def load_foundations_elements(
), "Readout block must have linear_1 or linear_mid"
if hasattr(readout, "linear_1"):
shape_input_1 = (
model_foundations.readouts[i]
foundation_readout
.linear_1.__dict__["irreps_out"]
.num_irreps
)
shape_output_1 = readout.linear_1.__dict__["irreps_out"].num_irreps
else:
raise ValueError("Readout block must have linear_1")
if hasattr(readout, "linear_1"):
model_readouts_one_linear_1_weight = readout.linear_1.weight.clone()
model_readouts_one_linear_1_weight = (
model_foundations.readouts[i]
foundation_readout
.linear_1.weight.view(num_channels_foundation, -1)
.repeat(1, len(model_heads))
.flatten()
Expand All @@ -236,10 +234,9 @@ def load_foundations_elements(
readout.linear_1.weight = torch.nn.Parameter(
model_readouts_one_linear_1_weight
)
if readout.linear_1.bias is not None:
model_readouts_one_linear_1_bias = readout.linear_1.bias.clone()
if readout.linear_1.bias is not None and readout.linear_1.bias.nelement() != 0:
model_readouts_one_linear_1_bias = (
model_foundations.readouts[i]
foundation_readout
.linear_1.bias.view(-1)
.repeat(len(model_heads))
.clone()
Expand All @@ -249,7 +246,7 @@ def load_foundations_elements(
)
if hasattr(readout, "linear_mid"):
readout.linear_mid.weight = torch.nn.Parameter(
model_foundations.readouts[i]
foundation_readout
.linear_mid.weight.view(
shape_input_1,
shape_input_1,
Expand All @@ -260,28 +257,25 @@ def load_foundations_elements(
/ ((shape_input_1) / (shape_output_1)) ** 0.5
)
# if it has biases transfer them too
if readout.linear_mid.bias is not None:
if readout.linear_mid.bias is not None and readout.linear_mid.bias.nelement() != 0:
readout.linear_mid.bias = torch.nn.Parameter(
model_foundations.readouts[i]
foundation_readout
.linear_mid.bias.repeat(len(model_heads))
.clone()
)
if hasattr(readout, "linear_2"):
model_readouts_one_linear_2_weight = readout.linear_2.weight.clone()
model_readouts_one_linear_2_weight = model_foundations.readouts[
i
].linear_2.weight.view(shape_input_1, -1).repeat(
len(model_heads), len(model_heads)
).flatten().clone() / (
((shape_input_1) / (shape_output_1)) ** 0.5
)
model_readouts_one_linear_2_weight = (foundation_readout
.linear_2.weight.view(shape_input_1, -1).repeat(
len(model_heads), len(model_heads)
).flatten().clone() / (
((shape_input_1) / (shape_output_1)) ** 0.5
))
readout.linear_2.weight = torch.nn.Parameter(
model_readouts_one_linear_2_weight
)
if readout.linear_2.bias is not None:
model_readouts_one_linear_2_bias = readout.linear_2.bias.clone()
if readout.linear_2.bias is not None and readout.linear_2.bias.nelement() != 0:
model_readouts_one_linear_2_bias = (
model_foundations.readouts[i]
foundation_readout
.linear_2.bias.view(-1)
.repeat(len(model_heads))
.flatten()
Expand Down
Loading