Repository navigation
[Fix] Resolve deprecated pipeline modules in from_pretrained - #14026
Open
Nas01010101 wants to merge 3 commits into
Open
Nas01010101 wants to merge 3 commits into
Nas01010101 wants to merge 3 commits into
Conversation
Contributor
|
Claude: Thanks for this — the One thing to flag: with this PR applied, loading a deprecated pipeline still doesn't complete. ( Drafted by Claude. |
Nas01010101
pushed a commit
to Nas01010101/diffusers
that referenced
this pull request
Jun 28, 2026
Follow-up to the review feedback on huggingface#14026: the earlier fix gets past the spurious ValueError in _get_custom_components_and_folders, but loading a deprecated pipeline still failed end-to-end. get_class_obj_and_candidates received is_pipeline_module=False for relocated modules (hasattr(pipelines, library_name) is False), so it fell to the bare importlib.import_module(library_name) branch and raised ModuleNotFoundError: No module named 'wuerstchen'. Even resolving the import there is not enough: that branch builds class_candidates via getattr(library, c, None), and a relocated module does not expose the base classes in importable_classes, so no load method would resolve and load_sub_model would then raise 'cannot be loaded ...'. Deprecated modules must be handled with pipeline-module semantics. Add an explicit branch that imports diffusers.pipelines.deprecated.<library_name> and resolves the class via dict.fromkeys, matching the is_pipeline_module path. Genuinely missing modules still raise ModuleNotFoundError. Adds GetClassObjAndCandidatesTests covering the wuerstchen prior component (class + from_pretrained load method) and the missing-module case.
|
For me the code changes by the 2 contributors of this PR fix the problem as tested on my side yesterday. I have not tried to run the related patched automated test collections. |
added 3 commits
August 19, 2026 20:47
Pipelines relocated to `diffusers.pipelines.deprecated` (e.g. Wuerstchen via `hasattr(pipelines, module_candidate)` check in `_get_custom_components_and_folders` mistook them for missing custom modules and raised a spurious ValueError when loading repos such as `warp-ai/wuerstchen-prior`. Also check the deprecated namespace (via `importlib.util.find_spec`, no import side effects) before treating a component as a missing custom module. Genuinely missing modules still raise. Fixes huggingface#14008
Follow-up to the review feedback on huggingface#14026: the earlier fix gets past the spurious ValueError in _get_custom_components_and_folders, but loading a deprecated pipeline still failed end-to-end. get_class_obj_and_candidates received is_pipeline_module=False for relocated modules (hasattr(pipelines, library_name) is False), so it fell to the bare importlib.import_module(library_name) branch and raised ModuleNotFoundError: No module named 'wuerstchen'. Even resolving the import there is not enough: that branch builds class_candidates via getattr(library, c, None), and a relocated module does not expose the base classes in importable_classes, so no load method would resolve and load_sub_model would then raise 'cannot be loaded ...'. Deprecated modules must be handled with pipeline-module semantics. Add an explicit branch that imports diffusers.pipelines.deprecated.<library_name> and resolves the class via dict.fromkeys, matching the is_pipeline_module path. Genuinely missing modules still raise ModuleNotFoundError. Adds GetClassObjAndCandidatesTests covering the wuerstchen prior component (class + from_pretrained load method) and the missing-module case.
Gap 1 — _fetch_class_library_tuple: deprecated modules (e.g. Wuerstchen)
live under diffusers.pipelines.deprecated, so hasattr(pipelines, pipeline_dir)
returns False and the function fell back to returning ("diffusers", class_name).
That breaks save_pretrained because the written model_index.json entry ("diffusers",
"WuerstchenPrior") is not resolvable on reload. Fix: promote the pipeline_dir to
is_pipeline_module=True when _is_deprecated_pipeline_module(pipeline_dir) is True,
so the returned library is the subpackage name (e.g. "wuerstchen").
Gap 2 — maybe_raise_or_warn: the guard was `if not is_pipeline_module`, which caused
deprecated-library callers to enter the type-check path, import the non-existent
top-level module, and eventually hit issubclass(model_cls, None) (TypeError) once
expected_class_obj stayed None. Fix: extend the guard to also skip the type-check
path for deprecated modules (`not _is_deprecated_pipeline_module(library_name)`),
mirroring the existing pipeline-module behaviour.
Gap 3 — simple_get_class_obj: the same is_pipeline_module=False path caused
importlib.import_module("wuerstchen") to raise ModuleNotFoundError. Fix: add an
elif branch that imports from diffusers.pipelines.deprecated.<library_name>,
mirroring the fix already present in get_class_obj_and_candidates.
Tests added for all three gaps (FetchClassLibraryTupleTests,
MaybeRaiseOrWarnTests, SimpleGetClassObjTests), including a roundtrip assertion
that proves the tuple written by _fetch_class_library_tuple is resolvable through
get_class_obj_and_candidates.
Nas01010101
force-pushed
the
fix/load-deprecated-pipelines
branch
from
August 20, 2026 01:05
aadc755 to
8a0866e
Compare
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Fixes #14008
The bug
DiffusionPipeline.from_pretrainedraises a spuriousValueErrorwhen loading any pipeline that was relocated underdiffusers.pipelines.deprecated(for example Wuerstchen, moved in #13157):Root cause
This is a bug in the deprecation mechanism itself, not in the deprecated pipeline.
_get_custom_components_and_foldersdecides whether a component refers to a built-in module viahasattr(diffusers.pipelines, module_candidate). When a pipeline is moved todiffusers.pipelines.deprecated, its module name is no longer a top-level attribute ofdiffusers.pipelines(deprecated modules are nested under thedeprecatedkey in_import_structure). SohasattrreturnsFalse, the component is mistaken for a missing custom module, and loading fails.The fix suggested in the issue (
hasattr(diffusers.pipelines.deprecated, module_candidate)) does not work:diffusers.pipelines.deprecatedis a_LazyModule, so the module name (wuerstchen) is not an attribute until it is imported; only the lazily exposed class names are.hasattrtherefore still returnsFalse.Fix
Resolve the module spec under the deprecated namespace without importing it, via
importlib.util.find_spec("diffusers.pipelines.deprecated.<module>"). This is import side-effect free (no heavy import is triggered) and correctly recognizes relocated modules. Genuinely missing modules still raise, sincefind_specreturnsNonefor them.Tests
Added
GetCustomComponentsAndFoldersTestsintests/pipelines/test_pipeline_utils.py(the existing home forpipeline_loading_utilsunit tests). All are function level and require no network or GPU.Reverting only the production change reproduces the issue (
from_pretrainedraises theValueErrorabove) and makes the new tests fail, confirming they guard the bug.ruff checkandruff format --checkpass on the changed files.Before submitting
.ai/review-rules.md, and I understand and take responsibility for the change.