Skip to content
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

Jump to bottom

Fix (weight_eq): fix for llm equalization #638

Merged
merged 2 commits into from
Jun 22, 2023
Merged

Fix (weight_eq): fix for llm equalization #638

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e5dc94b
Fix (weight_eq): fix for llm equalization
Giuseppe5 Jun 22, 2023
5976626
rename new_module to new_value
Giuseppe5 Jun 22, 2023
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
31 changes: 25 additions & 6 deletions src/brevitas/graph/equalize.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -444,11 +444,17 @@ def _no_equalize():
if len(src_axes) > 0:
for module, axis in src_axes.items():
if hasattr(module, 'bias') and module.bias is not None:
module.bias.data = module.bias.data * inverse_scaling_factors.view_as(module.bias)
_update_weights(
module,
module.bias.clone() * inverse_scaling_factors.view_as(module.bias),
attr='bias')
src_broadcast_size = [1] * module.weight.ndim
src_broadcast_size[axis] = module.weight.size(axis)
module.weight.data = module.weight.data * torch.reshape(
inverse_scaling_factors, src_broadcast_size)
_update_weights(
module, (
module.weight.clone() *
torch.reshape(inverse_scaling_factors, src_broadcast_size)),
attr='weight')
for module, axis in sink_axes.items():
src_broadcast_size = [1] * module.weight.ndim
src_broadcast_size[axis] = module.weight.size(axis)
Expand All @@ -457,12 +463,23 @@ def _no_equalize():
# additive factor for equalization.
additive_factor = module.running_mean.data * module.weight.data / torch.sqrt(
module.running_var.data + module.eps)
module.bias.data = module.bias.data + additive_factor * (scaling_factors - 1)
module.weight.data = module.weight.data * torch.reshape(scaling_factors, src_broadcast_size)
_update_weights(
module, module.bias.clone() + additive_factor * (scaling_factors - 1), attr='bias')
_update_weights(
module,
module.weight.clone() * torch.reshape(scaling_factors, src_broadcast_size),
attr='weight')

return scaling_factors


def _update_weights(original_module, new_value, attr='weight'):
if isinstance(original_module, WeightBiasTuple):
setattr(getattr(original_module, attr), 'data', new_value)
else:
setattr(original_module, attr, nn.Parameter(new_value))


def _equalize(
model: GraphModule,
regions: Set[Tuple[str]],
Expand Down Expand Up @@ -509,7 +526,9 @@ def _is_supported_module(graph_model: GraphModule, node: Node) -> bool:
if isinstance(module, _supported_layers):
# We support only self-attention
if isinstance(module, nn.MultiheadAttention):
return all([node.all_input_nodes[0].name == n.name for n in node.all_input_nodes])
kwargs = dict(node.kwargs)
kwargs.update(zip(module.forward.__code__.co_varnames[1:], node.args))
return kwargs['query'].name == kwargs['key'].name == kwargs['value'].name
return True
return False

Expand Down