make AcceleratRunner a subclass of Accelerator

add TorchRunner
add DeepSpeedRunner

.github/workflows/push.yaml

@@ -14,7 +14,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ["3.7", "3.8", "3.9", "3.10", "3.11", "3.12"]
+        python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
     steps:
       - uses: actions/checkout@v3
       - uses: actions/setup-python@v4
@@ -24,7 +24,7 @@ jobs:
       - name: Install dependencies
         run: pip install -r requirements.txt && pip install -e .
       - name: Install dependencies for testing
-        run: pip install pytest pytest-cov torch torcheval torchmetrics torchvision accelerate
+        run: pip install pytest pytest-cov
       - name: pytest
         run: pytest --cov=materialx --cov-report=xml --cov-report=html .
       - name: Upload coverage report for documentation

danling/__init__.py

@@ -17,7 +17,7 @@
 
 from lazy_imports import try_import
 
-from danling import metrics, modules, optim, registry, runner, tensors, typing, utils
+from danling import defaults, metrics, modules, optim, registry, runner, tensors, typing, utils
 
 from .metrics import (
     AverageMeter,
@@ -29,7 +29,7 @@
 )
 from .optim import LRScheduler
 from .registry import GlobalRegistry, Registry
-from .runner import AccelerateRunner, BaseRunner, TorchRunner
+from .runner import AccelerateRunner, BaseRunner, Config, DeepSpeedRunner, TorchRunner
 from .tensors import NestedTensor, PNTensor, tensor
 from .utils import (
     catch,
@@ -47,6 +47,7 @@
     from .metrics import Metrics, MultiTaskMetrics
 
 __all__ = [
+    "defaults",
     "metrics",
     "modules",
     "optim",
@@ -55,9 +56,11 @@
     "tensors",
     "utils",
     "typing",
+    "Config",
     "BaseRunner",
-    "AccelerateRunner",
     "TorchRunner",
+    "AccelerateRunner",
+    "DeepSpeedRunner",
     "LRScheduler",
     "Registry",
     "GlobalRegistry",

danling/runner/defaults.py → danling/defaults.py

@@ -15,14 +15,15 @@
 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
 # See the LICENSE file for more details.
 
-DEFAULT_RUN_NAME = "Run"
-DEFAULT_EXPERIMENT_NAME = "DanLing"
-DEFAULT_EXPERIMENT_ID = "xxxxxxxxxxxxxxxx"
-DEFAULT_IGNORED_KEYS_IN_HASH = {
+RUN_NAME = "Run"
+EXPERIMENT_NAME = "DanLing"
+EXPERIMENT_ID = "xxxxxxxxxxxxxxxx"
+SEED = 1016
+IGNORED_CONFIG_IN_HASH = {
     "timestamp",
-    "iters",
-    "steps",
-    "epochs",
+    "iter",
+    "step",
+    "epoch",
     "results",
     "score_split",
     "score",

danling/runner/README.md

@@ -4,26 +4,26 @@ The Runner of DanLing sets up the basic environment for running neural networks.
 
 ## Components
 
-For cross-platform compatibilities, DanLing features a two-level Runner + RunnerState system.
+For cross-platform compatibilities, DanLing features a two-level Runner + Config system.
 
 ### PlatformRunner
 
 PlatformRunner implements platform-specific features like `step` and `prepare`.
 
-The Runner contains all runtime information that is irrelevant to the checkpoint (e.g. `world_size`, `rank`, etc.). All other information should be saved in `RunnerState`.
+The Runner contains all runtime information that is irrelevant to the checkpoint (e.g. `world_size`, `rank`, etc.). All other information should be saved in `Config`.
 
 Currently, only [`AccelerateRunner`][danling.runner.AccelerateRunner] is supported.
 
 ### [`BaseRunner`][danling.runner.BaseRunner]
 
-[`BaseRunner`](danling.runner.BaseRunner) defines shared attributes and implements platform-agnostic features, including `init_logging`, `results` and `scores`.
+[`BaseRunner`][danling.runner.BaseRunner] defines shared attributes and implements platform-agnostic features, including `init_logging`, `results` and `scores`.
 
-### [`RunnerState`][danling.runner.RunnerState]
+### [`Config`][danling.runner.Config]
 
-[`RunnerState`][danling.runner.RunnerState] stores the state of a run (e.g. `epochs`, `run_id`, `network`, etc.).
+[`Config`][danling.runner.Config] stores the state of a run (e.g. `epoch`, `run_id`, `network`, etc.).
 
-With `RunnerState` and corresponding weights, you can resume a run from any point.
-Therefore, all members in `RunnerState` will be saved in the checkpoint, and thus should be json serialisable.
+With `Config` and corresponding weights, you can resume a run from any point.
+Therefore, all members in `Config` will be saved in the checkpoint, and thus should be json serialisable.
 
 ## Experiments Management
 

danling/runner/__init__.py

@@ -15,19 +15,20 @@
 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
 # See the LICENSE file for more details.
 
-from . import defaults
 from .accelerate_runner import AccelerateRunner
 from .base_runner import BaseRunner
-from .state import RunnerState
+from .config import Config
+from .deepspeed_runner import DeepSpeedRunner
 from .torch_runner import TorchRunner
 from .utils import on_local_main_process, on_main_process
 
 __all__ = [
-    "RunnerState",
+    "Config",
     "BaseRunner",
+    "TorchRunner",
     "AccelerateRunner",
+    "DeepSpeedRunner",
     "TorchRunner",
     "on_main_process",
     "on_local_main_process",
-    "defaults",
 ]