Skip to content

Drop stale param refs in _update_param_with_optimizer to cut VRAM spike #931

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
Evggamgam wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Drop stale param refs in _update_param_with_optimizer to cut VRAM spike #931

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
d1b4c65
Drop stale param refs in _update_param_with_optimizer to cut VRAM spike
Evggamgam Apr 22, 2026
512c46a
Merge branch 'main' into fix/densify-vram-transient
Evggamgam Apr 25, 2026
b4757fe
Merge branch 'main' into fix/densify-vram-transient
Evggamgam May 3, 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
20 changes: 13 additions & 7 deletions gsplat/strategy/ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,16 +92,22 @@ def _update_param_with_optimizer(
f"Got requires_grad={param.requires_grad}"
)
continue

optimizer = optimizers[name]
# Rebind refs to `new_param` before allocating new Adam state, so
# the old parameter tensor is released during optimizer_fn calls.
param_state = optimizer.state.pop(param, {})
for i in range(len(optimizer.param_groups)):
param_state = optimizer.state[param]
del optimizer.state[param]
for key in param_state.keys():
if key != "step":
v = param_state[key]
param_state[key] = optimizer_fn(key, v)
optimizer.param_groups[i]["params"] = [new_param]
optimizer.state[new_param] = param_state
optimizer.state[new_param] = param_state
del param

for key in list(param_state.keys()):
if key == "step":
continue
old = param_state[key]
param_state[key] = optimizer_fn(key, old)
del old


@torch.no_grad()
Expand Down