Skip to content

[core] respect local_files_only=True when using sharded checkpoints #12005

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
merged 18 commits into from
Aug 14, 2025
Merged

[core] respect local_files_only=True when using sharded checkpoints #12005

Changes from 7 commits
Commits
Show all changes
18 commits
Select commit Hold shift + click to select a range
8d431dc
tighten compilation tests for quantization
sayakpaul Jul 28, 2025
69920ef
feat: model_info but local.
sayakpaul Jul 28, 2025
d5c1772
up
sayakpaul Jul 28, 2025
f38a644
Revert "tighten compilation tests for quantization"
sayakpaul Jul 28, 2025
2d993b7
Merge branch 'main' into local-model-info
sayakpaul Jul 29, 2025
85279df
Merge branch 'main' into local-model-info
sayakpaul Aug 1, 2025
d117474
Merge branch 'main' into local-model-info
sayakpaul Aug 7, 2025
71843a0
Merge branch 'main' into local-model-info
sayakpaul Aug 12, 2025
fb2397f
up
sayakpaul Aug 12, 2025
832de66
Merge branch 'main' into local-model-info
sayakpaul Aug 13, 2025
01784c3
Merge branch 'main' into local-model-info
sayakpaul Aug 13, 2025
b7af511
reviewer feedback.
sayakpaul Aug 13, 2025
04cd2dc
reviewer feedback.
sayakpaul Aug 13, 2025
1c528a4
up
sayakpaul Aug 13, 2025
1b939e5
up
sayakpaul Aug 13, 2025
2a9734f
empty
sayakpaul Aug 13, 2025
09e063c
Merge branch 'main' into local-model-info
sayakpaul Aug 13, 2025
7fd1a82
update
DN6 Aug 14, 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
32 changes: 30 additions & 2 deletions src/diffusers/utils/hub_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,6 +45,7 @@
)
from packaging import version
from requests import HTTPError
from requests.exceptions import ConnectionError

from .. import __version__
from .constants import (
Expand Down Expand Up @@ -403,9 +404,26 @@ def _get_checkpoint_shard_files(

ignore_patterns = ["*.json", "*.md"]
# `model_info` call must guarded with the above condition.
model_files_info = model_info(pretrained_model_name_or_path, revision=revision, token=token)
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So the purpose of this check is to verify if the necessary sharded files are present in the model repo before attempting a download, presumably to avoid a large download if all files aren't present. If we cannot connect to the hub, we just have to assume the necessary shard files are already present locally.

I think we can just skip this check if local_files_only=True and then check if all the shard filenames are present in the cached_folder

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

How about now?

Copy link
Collaborator

@DN6 DN6 Aug 13, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think just this is sufficient

if not local_files_only:
    # run model_info check

Run snapshot download

Then after the cached_filenames is created, iterate over the files to verify they exist

for filename in cached_filename:
      if not if not os.path.exists(filename):
           raise EnvironmentError("expected file not present in {cached_folder}")

Copy link
Member Author

@sayakpaul sayakpaul Aug 13, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

  1. We don't have to run snapshot_download() when local_files_only=False, that might be unnecessary.
  2. Why run snapshot_download() after also running model_info()?
  3. Even if we run snapshot_download() regardless of local_files_only var, I think we should have it inside try-except in case the endpoint cannot be pinged for some reason and raise the ConnectionError as before.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

See if b7af511 resolves this.

Copy link
Member Author

@sayakpaul sayakpaul Aug 13, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah I see what you mean. Let me update. Sorry about the back and forth.

local = False
try:
model_files_info = model_info(pretrained_model_name_or_path, revision=revision, token=token)
except ConnectionError as e:
if local_files_only:
temp_dir = snapshot_download(
repo_id=pretrained_model_name_or_path, cache_dir=cache_dir, local_files_only=local_files_only
)
model_files_info = _get_filepaths_for_folder(temp_dir)
local = True
else:
raise EnvironmentError(
f"We couldn't connect to '{HUGGINGFACE_CO_RESOLVE_ENDPOINT}' to load {pretrained_model_name_or_path}. You should try"
" again after checking your internet connection."
) from e
for shard_file in original_shard_filenames:
shard_file_present = any(shard_file in k.rfilename for k in model_files_info.siblings)
if local:
shard_file_present = any(shard_file in k for k in model_files_info)
else:
shard_file_present = any(shard_file in k.rfilename for k in model_files_info.siblings)
if not shard_file_present:
raise EnvironmentError(
f"{shards_path} does not appear to have a file named {shard_file} which is "
Expand Down Expand Up @@ -441,6 +459,16 @@ def _get_checkpoint_shard_files(
return cached_filenames, sharded_metadata


def _get_filepaths_for_folder(folder):
relative_paths = []
for root, dirs, files in os.walk(folder):
for fname in files:
abs_path = os.path.join(root, fname)
rel_path = os.path.relpath(abs_path, start=folder)
relative_paths.append(rel_path)
return relative_paths


def _check_legacy_sharding_variant_format(folder: str = None, filenames: List[str] = None, variant: str = None):
if filenames and folder:
raise ValueError("Both `filenames` and `folder` cannot be provided.")
Expand Down
Loading