merge master

dmlc · Jul 8, 2024 · 4e0ce22 · 4e0ce22
2 parents 8669ce4 + 5b320e2
commit 4e0ce22
Show file tree

Hide file tree

Showing 135 changed files with 6,223 additions and 1,009 deletions.
diff --git a/.gitmodules b/.gitmodules
@@ -31,3 +31,6 @@
 [submodule "third_party/cuco"]
 	path = third_party/cuco
 	url = https://github.com/NVIDIA/cuCollections.git
+[submodule "third_party/GKlib"]
+	path = third_party/GKlib
+	url = https://github.com/KarypisLab/GKlib.git
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -394,13 +394,14 @@ else(EXTERNAL_METIS_PATH)
   target_include_directories(dgl PRIVATE "third_party/METIS/include")
   # Compile METIS
   if(NOT MSVC)
-    set(GKLIB_PATH "${CMAKE_CURRENT_SOURCE_DIR}/third_party/METIS/GKlib")
+    set(GKLIB_PATH "${CMAKE_CURRENT_SOURCE_DIR}/third_party/GKlib")
     include(${GKLIB_PATH}/GKlibSystem.cmake)
     include_directories(${GKLIB_PATH})
+    add_library(GKlib ${GKlib_sources})
     include_directories("third_party/METIS/include/")
     add_subdirectory("third_party/METIS/libmetis/")
-    list(APPEND DGL_LINKER_LIBS metis)
-    endif(NOT MSVC)
+    list(APPEND DGL_LINKER_LIBS metis GKlib)
+  endif(NOT MSVC)
 endif(EXTERNAL_METIS_PATH)
 
 

diff --git a/Jenkinsfile b/Jenkinsfile
@@ -158,7 +158,7 @@ def is_authorized(name) {
     'frozenbugs', 'peizhou001', 'zheng-da', 'czkkkkkk', 'thvasilo',
     // Intern:
     'keli-wen', 'caojy1998', 'RamonZhou', 'xiangyuzhi', 'Skeleton003', 'yxy235',
-    'hutiechuan', 'pyynb',
+    'hutiechuan', 'pyynb', 'az15240', 
     // Friends:
     'nv-dlasalle', 'yaox12', 'chang-l', 'Kh4L', 'VibhuJawa', 'kkranen',
     'TristonC', 'mfbalin',
@@ -575,7 +575,7 @@ pipeline {
                 docker {
                   label "dgl-ci-linux-cpu"
                   image "dgllib/dgl-ci-cpu:v240511_1440"
-                  args "-u root --shm-size=4gb"
+                  args "-u root --shm-size=8gb"
                   alwaysPull true
                 }
               }

diff --git a/conda/dgl/meta.yaml b/conda/dgl/meta.yaml
@@ -1,6 +1,6 @@
 package:
   name: dgl{{ environ.get('DGL_PACKAGE_SUFFIX', '') }}
-  version: 2.3{{ environ.get('DGL_VERSION_SUFFIX', '') }}
+  version: 2.4{{ environ.get('DGL_VERSION_SUFFIX', '') }}
 
 source:
   git_rev: {{ environ.get('DGL_RELEASE_BRANCH', 'master') }}

diff --git a/docs/source/api/python/dgl.graphbolt.rst b/docs/source/api/python/dgl.graphbolt.rst
@@ -82,7 +82,7 @@ An item set is an iterable collection of items.
     :template: graphbolt_classtemplate.rst
 
     ItemSet
-    ItemSetDict
+    HeteroItemSet
 
 
 ItemSampler
@@ -141,6 +141,7 @@ A subgraph sampler is for sampling subgraphs from a graph.
     SubgraphSampler
     SampledSubgraph
     NeighborSampler
+    TemporalNeighborSampler
     LayerNeighborSampler
     SampledSubgraphImpl
     FusedSampledSubgraphImpl

diff --git a/docs/source/guide/minibatch-custom-sampler.rst b/docs/source/guide/minibatch-custom-sampler.rst
@@ -75,7 +75,7 @@ can be used on heterogeneous graphs:
 
     import dgl.graphbolt as gb
     hg = gb.FusedCSCSamplingGraph()
-    train_set = item_set = gb.ItemSetDict(
+    train_set = item_set = gb.HeteroItemSet(
         {
             "user": gb.ItemSet(
                 (torch.arange(0, 5), torch.arange(5, 10)),
@@ -130,4 +130,4 @@ criteria, you could write your own transform function. Please check the method
 for reference.
 
 You could also refer to examples in
-`Link Prediction <https://github.com/dmlc/dgl/blob/master/examples/sampling/graphbolt/link_prediction.py>`__.
+`Link Prediction <https://github.com/dmlc/dgl/blob/master/examples/graphbolt/link_prediction.py>`__.
diff --git a/docs/source/guide/minibatch-edge.rst b/docs/source/guide/minibatch-edge.rst
@@ -248,7 +248,7 @@ over the edge types.
 
 Data loader definition is almost identical to that of homogeneous graph. The
 only difference is that the train_set is now an instance of
-:class:`~dgl.graphbolt.ItemSetDict` instead of :class:`~dgl.graphbolt.ItemSet`.
+:class:`~dgl.graphbolt.HeteroItemSet` instead of :class:`~dgl.graphbolt.ItemSet`.
 
 .. code:: python
 
@@ -266,7 +266,7 @@ only difference is that the train_set is now an instance of
             (seeds, labels), names=("seeds", "labels")
         ),
     }
-    train_set = gb.ItemSetDict(seeds_labels)
+    train_set = gb.HeteroItemSet(seeds_labels)
     datapipe = gb.ItemSampler(train_set, batch_size=128, shuffle=True)
     datapipe = datapipe.sample_neighbor(g, [10, 10]) # 2 layers.
     datapipe = datapipe.fetch_feature(

diff --git a/docs/source/guide/minibatch-inference.rst b/docs/source/guide/minibatch-inference.rst
@@ -122,7 +122,7 @@ of unlabeled data. Neighborhood sampling will work fine for model
 selection and validation.
 
 One can see
-`GraphSAGE <https://github.com/dmlc/dgl/blob/master/examples/sampling/graphbolt/node_classification.py>`__
+`GraphSAGE <https://github.com/dmlc/dgl/blob/master/examples/graphbolt/node_classification.py>`__
 and
-`RGCN <https://github.com/dmlc/dgl/blob/master/examples/sampling/graphbolt/rgcn/hetero_rgcn.py>`__
+`RGCN <https://github.com/dmlc/dgl/blob/master/examples/graphbolt/rgcn/hetero_rgcn.py>`__
 for examples of offline inference.
diff --git a/docs/source/guide/minibatch-link.rst b/docs/source/guide/minibatch-link.rst
@@ -139,7 +139,7 @@ above.
 
 
 DGL provides the
-`unsupervised learning GraphSAGE <https://github.com/dmlc/dgl/blob/master/examples/sampling/graphbolt/link_prediction.py>`__
+`unsupervised learning GraphSAGE <https://github.com/dmlc/dgl/blob/master/examples/graphbolt/link_prediction.py>`__
 that shows an example of link prediction on homogeneous graphs.
 
 For heterogeneous graphs

diff --git a/docs/source/guide/minibatch-node.rst b/docs/source/guide/minibatch-node.rst
@@ -162,7 +162,7 @@ customized batching iterator. During each iteration that yields
 
 
 DGL provides an end-to-end stochastic training example `GraphSAGE
-implementation <https://github.com/dmlc/dgl/blob/master/examples/sampling/graphbolt/node_classification.py>`__.
+implementation <https://github.com/dmlc/dgl/blob/master/examples/graphbolt/node_classification.py>`__.
 
 For heterogeneous graphs
 ~~~~~~~~~~~~~~~~~~~~~~~~
@@ -200,7 +200,7 @@ For example, one can still use the provided
 :class:`~dgl.graphbolt.NeighborSampler` class and
 :class:`~dgl.graphbolt.DataLoader` class for
 stochastic training. The only difference is that the itemset is now an
-instance of :class:`~dgl.graphbolt.ItemSetDict` which is a dictionary
+instance of :class:`~dgl.graphbolt.HeteroItemSet` which is a dictionary
 of node types to node IDs.
 
 .. code:: python
@@ -247,6 +247,6 @@ dictionaries of node types and predictions here.
         opt.step()
 
 DGL provides an end-to-end stochastic training example `RGCN
-implementation <https://github.com/dmlc/dgl/blob/master/examples/sampling/graphbolt/rgcn/hetero_rgcn.py>`__.
+implementation <https://github.com/dmlc/dgl/blob/master/examples/graphbolt/rgcn/hetero_rgcn.py>`__.
 
 
diff --git a/docs/source/guide/minibatch-sparse.rst b/docs/source/guide/minibatch-sparse.rst
@@ -171,4 +171,4 @@ Launch training:
             total_loss += loss.item()
 
 For more details, please refer to the `full example
-<https://github.com/dmlc/dgl/blob/master/examples/sampling/graphbolt/sparse/graphsage.py>`__.
+<https://github.com/dmlc/dgl/blob/master/examples/graphbolt/sparse/graphsage.py>`__.
diff --git a/docs/source/install/index.rst b/docs/source/install/index.rst
@@ -6,7 +6,7 @@ System requirements
 DGL works with the following operating systems:
 
 * Ubuntu 20.04+
-* CentOS 8+
+* CentOS 8+ (Although gcc 9 is needed)
 * RHEL 8+
 * macOS X
 * Windows 10

diff --git a/examples/distributed/graphsage/README.md b/examples/distributed/graphsage/README.md
@@ -139,6 +139,19 @@ python3 ~/workspace/dgl/tools/launch.py \
 "python3 node_classification.py --graph_name ogbn-products --ip_config ip_config.txt --num_epochs 30 --batch_size 1000 --num_gpus 4"
 ```
 
+Unsupervised training(train with link prediction dataloader).
+
+```
+python3 ~/workspace/dgl/tools/launch.py \
+--workspace ~/workspace/dgl/examples/distributed/graphsage/ \
+--num_trainers 1 \
+--num_samplers 0 \
+--num_servers 1 \
+--part_config data/ogbn-products.json \
+--ip_config ip_config.txt \
+"python3 node_classification_unsupervised.py --graph_name ogbn-products --ip_config ip_config.txt --num_epochs 30 --batch_size 1000 --remove_edge"
+```
+
 ### Running with GraphBolt
 
 In order to run with `GraphBolt`, we need to partition graph into `GraphBolt` data formats.Please note that both `DGL` and `GraphBolt` partitions are saved together.