Merge branch 'master' into cli-ckpt-hparams-subclass-mode

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 = ...