Merge pull request #17 from olipinski/logdir

1. **`fts.py`**:
   - Introduced a `log_dir` parameter to allow specifying a custom directory for artifacts, defaulting to `trainer.log_dir` or `trainer.default_root_dir`.
   - Added a `log_dir` property to determine the directory dynamically.
   - Updated various methods to use the new `log_dir` property instead of directly accessing `trainer.log_dir`.

2. **`fts_supporters.py`**:
   - Adjusted methods to utilize the `log_dir` property for saving schedules and validating configurations.

3. **`test_finetuning_scheduler_callback.py`**:
   - Added a test to validate the behavior of `FinetuningScheduler` with and without a specified `log_dir` when using a logger without a `save_dir`. I used `MLFlowLogger` as one such test case of the more general pattern. 

Arguably, Lightning should handle this scenario by updating its `trainer.log_dir` resolution logic to accommodate artifact persistence via `trainer.log_dir` for loggers that do not have a `save_dir` set (using `trainer.default_root_dir`) but your PR was an excellent opportunity to refactor and enhance FTS `log_dir` handling.

Author: speediedan

.azure-pipelines/gpu-tests.yml

@@ -92,10 +92,10 @@ jobs:
         bash ./tests/special_tests.sh --mark_type=standalone --filter_pattern='test_f'
       displayName: 'Testing: standalone multi-gpu'
 
-    - bash: |
-        . /tmp/venvs/fts_dev/bin/activate
-        bash ./tests/special_tests.sh --mark_type=exp_patch --filter_pattern='test_f' --experiment_patch_mask="1 0 0 1"
-      displayName: 'Testing: experimental einsum patch'
+    # - bash: |
+    #     . /tmp/venvs/fts_dev/bin/activate
+    #     bash ./tests/special_tests.sh --mark_type=exp_patch --filter_pattern='test_f' --experiment_patch_mask="1 0 0 1"
+    #   displayName: 'Testing: Experimental Multi-GPU'
 
     - bash: |
         . /tmp/venvs/fts_dev/bin/activate

.gitignore

@@ -9,6 +9,9 @@ lightning_logs/
 # ignore ipynb files themselves since we want to only store the source *.py
 *.ipynb
 
+# SQLite database files
+*.db
+
 # Test-tube
 test_tube_*/
 

CHANGELOG.md

@@ -268,7 +268,7 @@ name/pattern-based configuration instead of manually inspecting modules and appl
 
 ### Added
 
-- **FSDP Scheduled Fine-Tuning** is now supported! [See the tutorial here.](https://finetuning-scheduler.readthedocs.io/en/stable/advanced/fsdp_scheduled_fine_tuning.html)
+- **FSDP Scheduled Fine-Tuning** is now supported! [See the tutorial here.](https://finetuning-scheduler.readthedocs.io/en/stable/distributed/fsdp_scheduled_fine_tuning.html)
 - Introduced [``StrategyAdapter``](https://finetuning-scheduler.readthedocs.io/en/stable/api/finetuning_scheduler.strategy_adapters.html#finetuning_scheduler.strategy_adapters.StrategyAdapter)s. If you want to extend Fine-Tuning Scheduler (FTS) to use a custom, currently unsupported strategy or override current FTS behavior in the context of a given training strategy, subclassing ``StrategyAdapter`` is now a way to do so. See [``FSDPStrategyAdapter``](https://finetuning-scheduler.readthedocs.io/en/stable/api/finetuning_scheduler.strategy_adapters.html#finetuning_scheduler.strategy_adapters.FSDPStrategyAdapter) for an example implementation.
 - support for `pytorch-lightning` 1.9.0
 

CITATION.cff

@@ -125,6 +125,9 @@ identifiers:
   - description: "Fine-Tuning Scheduler (v2.5.0)"
     type: doi
     value: 10.5281/zenodo.14537830
+  - description: "Fine-Tuning Scheduler (v2.5.1)"
+    type: doi
+    value: 10.5281/zenodo.15099039
 license: "Apache-2.0"
 url: "https://finetuning-scheduler.readthedocs.io/"
 repository-code: "https://github.com/speediedan/finetuning-scheduler"

dockers/base-cuda/Dockerfile

@@ -90,9 +90,9 @@ RUN \
         # ... pytorch nightly dev version
         #pip install --pre torch==2.7.0.dev20250201 torchvision==0.22.0.dev20250201 --index-url https://download.pytorch.org/whl/nightly/cu128; \
         # temporarily remove torchvision from the nightly build until it supports cu128 in nightlies
-        pip install --pre torch==2.7.0.dev20250201 --index-url https://download.pytorch.org/whl/nightly/cu128; \
+        #pip install --pre torch==2.7.0.dev20250201 --index-url https://download.pytorch.org/whl/nightly/cu128; \
         # ... test channel
-        #pip install --pre torch==2.7.0 torchvision==0.22.0 --index-url https://download.pytorch.org/whl/test/cu128; \
+        pip install --pre torch==2.7.0 torchvision==0.22.0 --index-url https://download.pytorch.org/whl/test/cu128; \
     fi && \
     # Install all requirements
     pip install -r requirements/devel.txt --no-cache-dir && \

docs/source/conf.py

@@ -96,6 +96,8 @@ def _transform_changelog(path_in: str, path_out: str) -> None:
     "sphinx_togglebutton",
 ]
 
+autodoc_typehints = "none"
+
 # Suppress warnings about duplicate labels (needed for PL tutorials)
 suppress_warnings = [
     "autosectionlabel.*",

docs/source/index.rst

@@ -100,9 +100,9 @@ thawed/unfrozen parameter groups associated with each fine-tuning phase as desir
 and executed in ascending order. In addition to being zero-indexed, fine-tuning phase keys should be contiguous and
 either integers or convertible to integers via ``int()``.
 
-1. First, generate the default schedule to ``Trainer.log_dir``. It will be named after your
-   :external+pl:class:`~lightning.pytorch.core.module.LightningModule` subclass with the suffix
-   ``_ft_schedule.yaml``.
+1. First, generate the default schedule (output to :paramref:`~finetuning_scheduler.fts.FinetuningScheduler.log_dir`,
+   defaults to ``Trainer.log_dir``). It will be named after your
+   :external+pl:class:`~lightning.pytorch.core.module.LightningModule` subclass with the suffix ``_ft_schedule.yaml``.
 
 .. code-block:: python
 

pyproject.toml

@@ -7,11 +7,11 @@ requires = [
 [tool.ruff]
 line-length = 120
 # Enable Pyflakes `E` and `F` codes by default.
-select = [
+lint.select = [
     "E", "W",  # see: https://pypi.org/project/pycodestyle
     "F",  # see: https://pypi.org/project/pyflakes
 ]
-ignore = [
+lint.ignore = [
     "E731", # Do not assign a lambda expression, use a def
 ]
 # Exclude a variety of commonly ignored directories.
@@ -23,7 +23,7 @@ exclude = [
     "build",
     "temp",
 ]
-ignore-init-module-imports = true
+lint.ignore-init-module-imports = true
 output-format = "pylint"
 
 [tool.ruff.per-file-ignores]

requirements/base.txt

@@ -1,4 +1,4 @@
 #lightning>=2.7.0,<2.7.1
 # the below is uncommented when master is targeting a specific pl dev master commit
-git+https://github.com/Lightning-AI/lightning.git@efe311cd46a372aeb5912ea5adfeef573a5d64ca#egg=lightning
+git+https://github.com/Lightning-AI/lightning.git@669486afdd524fb66c1afc36bf93955384ac1224#egg=lightning
 torch>=2.4.0

requirements/docs.txt

@@ -3,8 +3,8 @@ myst-parser==0.18.1
 nbsphinx>=0.8.5
 pandoc>=1.0
 docutils>=0.16
-sphinxcontrib-fulltoc>=1.0
-sphinxcontrib-mockautodoc
+#sphinxcontrib-fulltoc>=1.0
+#sphinxcontrib-mockautodoc
 sphinx-autodoc-typehints>=1.16
 sphinx-paramlinks>=0.5.1
 sphinx-togglebutton>=0.2