Skip to content

Solve checkpoint and validation bugs in training corrdiff unet #1000

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
wants to merge 11 commits into
base: main
Choose a base branch
from
Open

Solve checkpoint and validation bugs in training corrdiff unet #1000

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
0492d33
make dimensions consistent for checkpointing
albertocarpentieri Jul 3, 2025
00e6013
add use_reentrant=False to checkpoint in songuent for torch.compile s…
albertocarpentieri Jul 3, 2025
3f35501
removed use_patch_grad_acc from loss_valid_kwargs in corrdiff train.p…
albertocarpentieri Jul 3, 2025
489e895
set graph static for corrdiff training to enable checkpoint
albertocarpentieri Jul 3, 2025
b1c0033
Merge branch 'main' into main
CharlelieLrt Jul 11, 2025
7bfce5a
Merge branch 'main' into main
CharlelieLrt Jul 12, 2025
a74aa67
change the checkpoint reference dimension from x to y as it is the sa…
albertocarpentieri Jul 14, 2025
8dcec67
correct positional embedding in song unet
albertocarpentieri Jul 14, 2025
3cd2587
correct embedding for gridtype==test and N_grid_channels==2
albertocarpentieri Jul 14, 2025
c925df1
Change single dimension shape with geometric mean to use checkpointing
albertocarpentieri Jul 14, 2025
7b1ec5d
Merge remote-tracking branch 'upstream/main'
albertocarpentieri Jul 17, 2025
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
2 changes: 1 addition & 1 deletion examples/weather/corrdiff/train.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -325,6 +325,7 @@ def main(cfg: DictConfig) -> None:
find_unused_parameters=True, # dist.find_unused_parameters,
bucket_cap_mb=35,
gradient_as_bucket_view=True,
static_graph=True
)
if cfg.wandb.watch_model and dist.rank == 0:
wandb.watch(model)
Expand Down Expand Up @@ -681,7 +682,6 @@ def main(cfg: DictConfig) -> None:
"img_clean": img_clean_valid,
"img_lr": img_lr_valid,
"augment_pipe": None,
"use_patch_grad_acc": use_patch_grad_acc,
}
if use_patch_grad_acc is not None:
loss_valid_kwargs[
Expand Down
49 changes: 22 additions & 27 deletions physicsnemo/models/diffusion/song_unet.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,7 @@

import numpy as np
import nvtx
import math
import torch
from torch.nn.functional import silu
from torch.utils.checkpoint import checkpoint
Expand Down Expand Up @@ -336,7 +337,7 @@ def __init__(
self.img_shape_x = img_resolution[1]

# set the threshold for checkpointing based on image resolution
self.checkpoint_threshold = (self.img_shape_y >> checkpoint_level) + 1
self.checkpoint_threshold = (math.floor(math.sqrt(self.img_shape_x * self.img_shape_y)) >> checkpoint_level) + 1

# Optional additive learned positition embed after the first conv
self.additive_pos_embed = additive_pos_embed
Expand Down Expand Up @@ -552,10 +553,10 @@ def forward(self, x, noise_labels, class_labels, augment_labels=None):
else:
# For UNetBlocks check if we should use gradient checkpointing
if isinstance(block, UNetBlock):
if x.shape[-1] > self.checkpoint_threshold:
if math.floor(math.sqrt(x.shape[-2] * x.shape[-1])) > self.checkpoint_threshold:
# self.checkpoint = checkpoint?
# else: self.checkpoint = lambda(block,x,emb:block(x,emb))
x = checkpoint(block, x, emb)
x = checkpoint(block, x, emb, use_reentrant=False)
else:
# AssertionError: Only support NHWC layout.
x = block(x, emb)
Expand Down Expand Up @@ -584,12 +585,12 @@ def forward(self, x, noise_labels, class_labels, augment_labels=None):
x = torch.cat([x, skips.pop()], dim=1)
# check for checkpointing on decoder blocks and up sampling blocks
if (
x.shape[-1] > self.checkpoint_threshold and "_block" in name
math.floor(math.sqrt(x.shape[-2] * x.shape[-1])) > self.checkpoint_threshold and "_block" in name
) or (
x.shape[-1] > (self.checkpoint_threshold / 2)
math.floor(math.sqrt(x.shape[-2] * x.shape[-1])) > (self.checkpoint_threshold / 2)
and "_up" in name
):
x = checkpoint(block, x, emb)
x = checkpoint(block, x, emb, use_reentrant=False)
else:
x = block(x, emb)
return aux
Expand Down Expand Up @@ -1111,28 +1112,22 @@ def _get_positional_embedding(self):
elif self.gridtype == "linear":
if self.N_grid_channels != 2:
raise ValueError("N_grid_channels must be set to 2 for gridtype linear")
x = np.meshgrid(np.linspace(-1, 1, self.img_shape_y))
y = np.meshgrid(np.linspace(-1, 1, self.img_shape_x))
grid_x, grid_y = np.meshgrid(y, x)
y = np.meshgrid(np.linspace(-1, 1, self.img_shape_y))
x = np.meshgrid(np.linspace(-1, 1, self.img_shape_x))
grid_y, grid_x = np.meshgrid(x, y)
grid = torch.from_numpy(
np.stack((grid_x, grid_y), axis=0)
np.stack((grid_y, grid_x), axis=0)
) # (2, img_shape_y, img_shape_x)
grid.requires_grad = False
elif self.gridtype == "sinusoidal" and self.N_grid_channels == 4:
# print('sinusuidal grid added ......')
x1 = np.meshgrid(np.sin(np.linspace(0, 2 * np.pi, self.img_shape_y)))
x2 = np.meshgrid(np.cos(np.linspace(0, 2 * np.pi, self.img_shape_y)))
y1 = np.meshgrid(np.sin(np.linspace(0, 2 * np.pi, self.img_shape_x)))
y2 = np.meshgrid(np.cos(np.linspace(0, 2 * np.pi, self.img_shape_x)))
grid_x1, grid_y1 = np.meshgrid(y1, x1)
grid_x2, grid_y2 = np.meshgrid(y2, x2)
grid = torch.squeeze(
torch.from_numpy(
np.expand_dims(
np.stack((grid_x1, grid_y1, grid_x2, grid_y2), axis=0), axis=0
)
)
) # (4, img_shape_y, img_shape_x)
x1 = np.meshgrid(np.sin(np.linspace(0, 2 * np.pi, self.img_shape_x)))
x2 = np.meshgrid(np.cos(np.linspace(0, 2 * np.pi, self.img_shape_x)))
y1 = np.meshgrid(np.sin(np.linspace(0, 2 * np.pi, self.img_shape_y)))
y2 = np.meshgrid(np.cos(np.linspace(0, 2 * np.pi, self.img_shape_y)))
grid_y1, grid_x1 = np.meshgrid(x1, y1)
grid_y2, grid_x2 = np.meshgrid(x2, y2)
grid = torch.from_numpy(np.stack((grid_x1, grid_y1, grid_x2, grid_y2), axis=0))
grid.requires_grad = False
elif self.gridtype == "sinusoidal" and self.N_grid_channels != 4:
if self.N_grid_channels % 4 != 0:
Expand All @@ -1153,10 +1148,10 @@ def _get_positional_embedding(self):
) # (N_grid_channels, img_shape_y, img_shape_x)
grid.requires_grad = False
elif self.gridtype == "test" and self.N_grid_channels == 2:
idx_x = torch.arange(self.img_shape_y)
idx_y = torch.arange(self.img_shape_x)
mesh_x, mesh_y = torch.meshgrid(idx_x, idx_y)
grid = torch.stack((mesh_x, mesh_y), dim=0) # (2, img_shape_y, img_shape_x)
idx_x = torch.arange(self.img_shape_x)
idx_y = torch.arange(self.img_shape_y)
mesh_y, mesh_x = torch.meshgrid(idx_y, idx_x)
grid = torch.stack((mesh_y, mesh_x), dim=0) # (2, img_shape_y, img_shape_x)
else:
raise ValueError("Gridtype not supported.")
return grid
Expand Down