Merge branch 'master' into wkong/perf-opt

Author: SkafteNicki

.github/checkgroup.yml

@@ -47,9 +47,9 @@ subprojects:
       - "!*.md"
       - "!**/*.md"
     checks:
-      - "pytorch.yml / Lit Job (nvidia/cuda:12.1.1-runtime-ubuntu22.04, pytorch, 3.10, L4_X_2)"
-      - "pytorch.yml / Lit Job (nvidia/cuda:12.6.3-runtime-ubuntu22.04, lightning, 3.12, L4_X_2)"
-      - "pytorch.yml / Lit Job (nvidia/cuda:12.6.3-runtime-ubuntu22.04, pytorch, 3.12, L4_X_2)"
+      - "pytorch.yml / Lit Job (nvidia/cuda:12.1.1-runtime-ubuntu22.04, pytorch, 3.10)"
+      - "pytorch.yml / Lit Job (lightning, 3.12)"
+      - "pytorch.yml / Lit Job (pytorch, 3.12)"
 
   - id: "Benchmarks"
     paths:
@@ -148,9 +148,9 @@ subprojects:
       - "!*.md"
       - "!**/*.md"
     checks:
-      - "fabric.yml / Lit Job (nvidia/cuda:12.1.1-runtime-ubuntu22.04, fabric, 3.10, L4_X_2)"
-      - "fabric.yml / Lit Job (nvidia/cuda:12.6.3-runtime-ubuntu22.04, fabric, 3.12, L4_X_2)"
-      - "fabric.yml / Lit Job (nvidia/cuda:12.6.3-runtime-ubuntu22.04, lightning, 3.12, L4_X_2)"
+      - "fabric.yml / Lit Job (nvidia/cuda:12.1.1-runtime-ubuntu22.04, fabric, 3.10)"
+      - "fabric.yml / Lit Job (fabric, 3.12)"
+      - "fabric.yml / Lit Job (lightning, 3.12)"
 
   # Temporarily disabled
   #  - id: "lightning_fabric: TPU workflow"

.lightning/workflows/fabric.yml

@@ -5,24 +5,21 @@ trigger:
     branches: ["master", "release/stable"]
 
 timeout: "60" # minutes
+machine: "L4_X_2"
+image: "nvidia/cuda:12.6.3-runtime-ubuntu22.04"
 parametrize:
   matrix: {}
   include:
     # note that this is setting also all oldest requirements which is linked to python == 3.10
     - image: "nvidia/cuda:12.1.1-runtime-ubuntu22.04"
       PACKAGE_NAME: "fabric"
       python_version: "3.10"
-      machine: "L4_X_2"
-    - image: "nvidia/cuda:12.6.3-runtime-ubuntu22.04"
-      PACKAGE_NAME: "fabric"
+    - PACKAGE_NAME: "fabric"
       python_version: "3.12"
-      machine: "L4_X_2"
     # - image: "nvidia/cuda:12.6-runtime-ubuntu22.04"
     #   PACKAGE_NAME: "fabric"
-    - image: "nvidia/cuda:12.6.3-runtime-ubuntu22.04"
-      PACKAGE_NAME: "lightning"
+    - PACKAGE_NAME: "lightning"
       python_version: "3.12"
-      machine: "L4_X_2"
   exclude: []
 
 env:

.lightning/workflows/pytorch.yml

@@ -5,24 +5,21 @@ trigger:
     branches: ["master", "release/stable"]
 
 timeout: "60" # minutes
+machine: "L4_X_2"
+image: "nvidia/cuda:12.6.3-runtime-ubuntu22.04"
 parametrize:
   matrix: {}
   include:
     # note that this also sets oldest requirements which are linked to Python == 3.10
     - image: "nvidia/cuda:12.1.1-runtime-ubuntu22.04"
       PACKAGE_NAME: "pytorch"
       python_version: "3.10"
-      machine: "L4_X_2"
-    - image: "nvidia/cuda:12.6.3-runtime-ubuntu22.04"
-      PACKAGE_NAME: "pytorch"
+    - PACKAGE_NAME: "pytorch"
       python_version: "3.12"
-      machine: "L4_X_2"
     # - image: "nvidia/cuda:12.6.3-runtime-ubuntu22.04"
     #   PACKAGE_NAME: "pytorch"
-    - image: "nvidia/cuda:12.6.3-runtime-ubuntu22.04"
-      PACKAGE_NAME: "lightning"
+    - PACKAGE_NAME: "lightning"
       python_version: "3.12"
-      machine: "L4_X_2"
   exclude: []
 
 env:

docs/source-pytorch/conf.py

@@ -645,6 +645,7 @@ def package_list_from_file(file):
     r"installation.html$",
     r"starter/installation.html$",
     r"^../common/trainer.html#trainer-flags$",
+    "https://medium.com/pytorch-lightning/quick-contribution-guide-86d977171b3a",
     "https://deepgenerativemodels.github.io/assets/slides/cs236_lecture11.pdf",
     "https://developer.habana.ai", # returns 403 error but redirects to intel.com documentation
     "https://www.intel.com/content/www/us/en/products/docs/processors/what-is-a-gpu.html",

docs/source-pytorch/data/alternatives.rst

@@ -99,7 +99,12 @@ The webdataset library contains a small wrapper (``WebLoader``) that adds a flui
     import lightning as L
     import webdataset as wds
 
-    dataset = wds.WebDataset(urls)
+    dataset = wds.WebDataset(
+        urls,
+        # needed for multi-gpu or multi-node training
+        workersplitter=wds.shardlists.split_by_worker,
+        nodesplitter=wds.shardlists.split_by_node,
+    )
     train_dataloader = wds.WebLoader(dataset)
 
     model = ...

requirements/fabric/test.txt

@@ -6,5 +6,5 @@ pytest-timeout ==2.4.0
 pytest-rerunfailures ==16.0.1
 pytest-random-order ==1.2.0
 click ==8.1.8; python_version < "3.11"
-click ==8.2.1; python_version > "3.10"
+click ==8.3.0; python_version > "3.10"
 tensorboardX >=2.6, <2.7.0  # todo: relax it back to `>=2.2` after fixing tests

src/lightning/pytorch/CHANGELOG.md

@@ -52,8 +52,12 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed preventing recursive symlink creation iwhen `save_last='link'` and `save_top_k=-1` ([#21186](https://github.com/Lightning-AI/pytorch-lightning/pull/21186))
 
 
+- Fixed `LightningCLI` loading of hyperparameters from `ckpt_path` failing for subclass model mode ([#21246](https://github.com/Lightning-AI/pytorch-lightning/pull/21246))
+
+
 - Fixed redundant host-device sync in progressbar printing ([#21233](https://github.com/Lightning-AI/pytorch-lightning/pull/21233))
 
+
 ---
 
 ## [2.5.5] - 2025-09-05

src/lightning/pytorch/cli.py

@@ -564,6 +564,11 @@ def _parse_ckpt_path(self) -> None:
             hparams.pop("_instantiator", None)
             if not hparams:
                 return
+            if "_class_path" in hparams:
+                hparams = {
+                    "class_path": hparams.pop("_class_path"),
+                    "dict_kwargs": hparams,
+                }
             hparams = {self.config.subcommand: {"model": hparams}}
             try:
                 self.config = self.parser.parse_object(hparams, self.config)

tests/tests_pytorch/test_cli.py

@@ -491,6 +491,7 @@ class BoringCkptPathModel(BoringModel):
     def __init__(self, out_dim: int = 2, hidden_dim: int = 2) -> None:
         super().__init__()
         self.save_hyperparameters()
+        self.hidden_dim = hidden_dim
         self.layer = torch.nn.Linear(32, out_dim)
 
 
@@ -526,6 +527,41 @@ def add_arguments_to_parser(self, parser):
     assert "Parsing of ckpt_path hyperparameters failed" in err.getvalue()
 
 
+class BoringCkptPathSubclass(BoringCkptPathModel):
+    def __init__(self, extra: bool = True, **kwargs) -> None:
+        super().__init__(**kwargs)
+        self.extra = extra
+
+
+def test_lightning_cli_ckpt_path_argument_hparams_subclass_mode(cleandir):
+    class CkptPathCLI(LightningCLI):
+        def add_arguments_to_parser(self, parser):
+            parser.link_arguments("model.init_args.out_dim", "model.init_args.hidden_dim", compute_fn=lambda x: x * 2)
+
+    cli_args = ["fit", "--model=BoringCkptPathSubclass", "--model.out_dim=4", "--trainer.max_epochs=1"]
+    with mock.patch("sys.argv", ["any.py"] + cli_args):
+        cli = CkptPathCLI(BoringCkptPathModel, subclass_mode_model=True)
+
+    assert cli.config.fit.model.class_path.endswith(".BoringCkptPathSubclass")
+    assert cli.config.fit.model.init_args == Namespace(out_dim=4, hidden_dim=8, extra=True)
+    hparams_path = Path(cli.trainer.log_dir) / "hparams.yaml"
+    assert hparams_path.is_file()
+    hparams = yaml.safe_load(hparams_path.read_text())
+    assert hparams["out_dim"] == 4
+    assert hparams["hidden_dim"] == 8
+    assert hparams["extra"] is True
+
+    checkpoint_path = next(Path(cli.trainer.log_dir, "checkpoints").glob("*.ckpt"))
+    cli_args = ["predict", "--model=BoringCkptPathModel", f"--ckpt_path={checkpoint_path}"]
+    with mock.patch("sys.argv", ["any.py"] + cli_args):
+        cli = CkptPathCLI(BoringCkptPathModel, subclass_mode_model=True)
+
+    assert isinstance(cli.model, BoringCkptPathSubclass)
+    assert cli.model.hidden_dim == 8
+    assert cli.model.extra is True
+    assert cli.model.layer.out_features == 4
+
+
 def test_lightning_cli_submodules(cleandir):
     class MainModule(BoringModel):
         def __init__(self, submodule1: LightningModule, submodule2: LightningModule, main_param: int = 1):