Add reproducible legacy container environment

.dockerignore

@@ -0,0 +1,17 @@
+.git
+.venv
+**/__pycache__
+**/*.py[cod]
+**/*.so
+**/*.pt
+**/*.csv
+**/*.log
+**/checkpoints
+**/ckpts
+**/sweeps
+**/tensorboard
+**/test_results
+build
+dist
+exps
+wandb

Dockerfile

@@ -0,0 +1,81 @@
+FROM python:3.9.19-slim-bookworm@sha256:69e712dbe4c4a166527cbf69374533125cfb6ee93a5e39031a0191c741d386d7 AS builder
+
+ENV DEBIAN_FRONTEND=noninteractive \
+    DGLBACKEND=pytorch \
+    PIP_DISABLE_PIP_VERSION_CHECK=1 \
+    PIP_NO_CACHE_DIR=1 \
+    PYTHONPATH=/opt/graphormer \
+    PYTHONUNBUFFERED=1
+
+RUN apt-get update \
+    && apt-get install --yes --no-install-recommends \
+        build-essential \
+        ca-certificates \
+        git \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /opt/graphormer
+COPY . .
+
+RUN test -f fairseq/setup.py \
+    || (echo "Initialize submodules before building: git submodule update --init --recursive" >&2; exit 1)
+
+RUN python -m pip install \
+        pip==23.3.2 \
+        setuptools==69.5.1 \
+        wheel==0.48.0 \
+    && python -m pip install \
+        torch==1.9.1+cu111 \
+        torchaudio==0.9.1 \
+        --find-links https://download.pytorch.org/whl/cu111/torch_stable.html \
+    && python -m pip install --requirement docker/requirements.txt \
+    && python -m pip install \
+        dgl==0.7.2 \
+        --find-links https://data.dgl.ai/wheels/repo.html
+
+COPY docker/sitecustomize.py /usr/local/lib/python3.9/site-packages/sitecustomize.py
+
+RUN cd fairseq \
+    && CFLAGS="-include cstdint" CXXFLAGS="-include cstdint" \
+        python -m pip install --no-build-isolation .
+
+RUN python - <<'PY'
+import numpy
+import pyximport
+
+pyximport.install(
+    setup_args={"include_dirs": numpy.get_include()},
+    inplace=True,
+    language_level=3,
+)
+import graphormer.data.algos
+PY
+
+RUN python docker/verify_environment.py
+
+
+FROM python:3.9.19-slim-bookworm@sha256:69e712dbe4c4a166527cbf69374533125cfb6ee93a5e39031a0191c741d386d7 AS runtime
+
+LABEL org.opencontainers.image.title="Graphormer legacy runtime" \
+      org.opencontainers.image.description="Pinned Python 3.9, PyTorch 1.9, and Fairseq environment for Graphormer" \
+      org.opencontainers.image.source="https://github.com/microsoft/Graphormer" \
+      org.opencontainers.image.version="legacy-py39-torch1.9"
+
+ENV PIP_DISABLE_PIP_VERSION_CHECK=1 \
+    DGLBACKEND=pytorch \
+    PYTHONPATH=/opt/graphormer \
+    PYTHONUNBUFFERED=1
+
+RUN apt-get update \
+    && apt-get install --yes --no-install-recommends \
+        ca-certificates \
+        g++ \
+        libgomp1 \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY --from=builder /usr/local /usr/local
+COPY --from=builder /opt/graphormer /opt/graphormer
+
+WORKDIR /opt/graphormer
+
+CMD ["bash"]

README.md

@@ -47,7 +47,57 @@ Next you may want to read:
 
 ## Requirements and Installation
 
-#### Setup with Conda
+### Containerized legacy environment
+
+Graphormer extends a pinned Fairseq revision and uses a legacy Python and
+PyTorch stack. The container is the recommended way to run this environment
+without changing packages in the host Python installation.
+
+Initialize the Fairseq submodule and build the image:
+
+```bash
+git submodule update --init --recursive
+docker build --tag graphormer:legacy .
+```
+
+Verify CPU imports and graph preprocessing:
+
+```bash
+docker run --rm graphormer:legacy \
+    python docker/verify_environment.py
+```
+
+GPU execution requires the NVIDIA Container Toolkit:
+
+```bash
+docker run --rm --gpus all graphormer:legacy \
+    python docker/verify_environment.py --require-cuda
+```
+
+Run an interactive shell and mount data or checkpoints separately from the
+repository:
+
+```bash
+docker run --rm --gpus all --interactive --tty \
+    --volume /path/to/data:/data \
+    graphormer:legacy
+```
+
+For single-host distributed training, provide an explicit process-group URL:
+
+```bash
+docker run --rm --gpus all graphormer:legacy \
+    fairseq-train ... \
+    --distributed-world-size 2 \
+    --distributed-init-method tcp://localhost:29500
+```
+
+The image intentionally pins Python 3.9, PyTorch 1.9.1, CUDA 11.1 user-space
+libraries, and pip 23.3.2. It is a reproducibility environment for the existing
+Fairseq-based code, not a claim of compatibility with current Python or
+PyTorch releases. Multi-node training is not covered by this setup.
+
+### Local installation
 
 ```
 bash install.sh

docker/requirements.txt

@@ -0,0 +1,10 @@
+Cython==0.29.36
+googledrivedownloader==0.4
+lmdb==2.3.0
+numpy==1.23.5
+ogb==1.3.2
+protobuf==3.20.3
+rdkit-pypi==2021.9.3
+tensorboard==2.11.2
+tensorboardX==2.4.1
+torch-geometric==2.3.1

docker/sitecustomize.py

@@ -0,0 +1,3 @@
+"""Compatibility imports for the pinned PyTorch 1.9 runtime."""
+
+import distutils.version

docker/verify_environment.py

@@ -0,0 +1,77 @@
+#!/usr/bin/env python
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+import argparse
+import json
+
+import numpy as np
+import torch
+from ogb.utils.mol import smiles2graph
+from torch_geometric.data import Data
+
+import fairseq
+import graphormer
+import graphormer.models
+import graphormer.tasks.graph_prediction
+from graphormer.data import algos
+from graphormer.data.wrapper import preprocess_item
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--require-cuda", action="store_true")
+    parser.add_argument("--expected-gpus", type=int)
+    return parser.parse_args()
+
+
+def main():
+    args = parse_args()
+
+    adjacency = np.array(
+        [
+            [False, True, False],
+            [True, False, True],
+            [False, True, False],
+        ]
+    )
+    distances, _ = algos.floyd_warshall(adjacency)
+    if distances.tolist() != [[0, 1, 2], [1, 0, 1], [2, 1, 0]]:
+        raise RuntimeError("Graphormer shortest-path preprocessing returned bad data")
+
+    graph = smiles2graph("CCO")
+    item = preprocess_item(
+        Data(
+            x=torch.from_numpy(graph["node_feat"]).long(),
+            edge_index=torch.from_numpy(graph["edge_index"]).long(),
+            edge_attr=torch.from_numpy(graph["edge_feat"]).long(),
+            y=torch.tensor([0.0]),
+        )
+    )
+    if tuple(item.spatial_pos.shape) != (3, 3):
+        raise RuntimeError("Graphormer molecular preprocessing returned a bad shape")
+
+    cuda_available = torch.cuda.is_available()
+    gpu_count = torch.cuda.device_count()
+    if args.require_cuda and not cuda_available:
+        raise RuntimeError("CUDA was required but is not available")
+    if args.expected_gpus is not None and gpu_count != args.expected_gpus:
+        raise RuntimeError(f"Expected {args.expected_gpus} GPUs, found {gpu_count}")
+
+    print(
+        json.dumps(
+            {
+                "cuda_available": cuda_available,
+                "fairseq": fairseq.__version__,
+                "gpu_count": gpu_count,
+                "graphormer": graphormer.__file__,
+                "preprocessed_nodes": item.spatial_pos.size(0),
+                "torch": torch.__version__,
+            },
+            sort_keys=True,
+        )
+    )
+
+
+if __name__ == "__main__":
+    main()

graphormer/data/dgl_datasets/dgl_dataset.py

@@ -148,7 +148,7 @@ def __preprocess_dgl_graph(
 
         return pyg_graph
 
-    def __getitem__(self, idx):
+    def get(self, idx):
         if isinstance(idx, int):
             if self.__indices__ is not None:
                 idx = self.__indices__[idx]
@@ -157,5 +157,11 @@ def __getitem__(self, idx):
         else:
             raise TypeError("index to a GraphormerDGLDataset can only be an integer.")
 
-    def __len__(self) -> int:
+    def len(self) -> int:
         return len(self.dataset) if self.__indices__ is None else len(self.__indices__)
+
+    def __getitem__(self, idx):
+        return self.get(idx)
+
+    def __len__(self) -> int:
+        return self.len()

graphormer/data/pyg_datasets/pyg_dataset.py

@@ -93,7 +93,7 @@ def create_subset(self, subset):
         return dataset
 
     @lru_cache(maxsize=16)
-    def __getitem__(self, idx):
+    def get(self, idx):
         if isinstance(idx, int):
             item = self.dataset[idx]
             item.idx = idx
@@ -102,5 +102,11 @@ def __getitem__(self, idx):
         else:
             raise TypeError("index to a GraphormerPYGDataset can only be an integer.")
 
-    def __len__(self):
+    def len(self):
         return self.num_data
+
+    def __getitem__(self, idx):
+        return self.get(idx)
+
+    def __len__(self):
+        return self.len()