Multi node check

.github/workflows/integration.yml → .github/workflows/tests_integration.yml

@@ -47,25 +47,22 @@ jobs:
     steps:
       - uses: actions/setup-python@v4
         with:
-          python-version: '3.10'
+          python-version: '3.11'
 
       - name: Check out repository code
         uses: actions/checkout@v3
 
       - name: dependencies
         run: |
-          if [[ ! -d "~/.cargo/bin" ]]; then
-            wget --no-check-certificate --secure-protocol=TLSv1_2 -qO- https://sh.rustup.rs | sh -s -- -y 
-          fi
           export PATH="~/.cargo/bin:${PATH}"
           python -m pip install -U pip
           python -m pip install -U poetry
 
       - name: install
         run: |
-          pip install pytest
-          poetry lock --no-update
-          pip install -e .
+          poetry install --with dev
+          source $(poetry env info -p)/bin/activate
+          pip install psycopg2-binary
 
       - name: tests
         env:
@@ -74,4 +71,4 @@ jobs:
           POSTGRES_DB: milabench
           POSTGRES_HOST: localhost
           POSTGRES_PORT: 5432
-        run: pytest tests/integration
+        run: poetry run pytest tests/integration

.github/workflows/tests.yml → .github/workflows/tests_run.yml

@@ -1,4 +1,4 @@
-name: tests
+name: run
 
 on:
   # Runs every sunday
@@ -55,7 +55,7 @@ jobs:
       - uses: conda-incubator/setup-miniconda@v2
         with:
           auto-activate-base: false
-          python-version: 3.10
+          python-version: 3.11
           miniconda-version: "latest"
           activate-environment: test
 

.github/workflows/tests_unit.yml

@@ -0,0 +1,67 @@
+name: unit
+
+on:
+  push:
+
+  # Runs for pull requests
+  pull_request:
+    branches:
+      - master
+
+  # Runs on publish
+  release:
+    types:
+      [published]
+
+  # Allow manual triggers
+  workflow_dispatch:
+
+
+jobs:
+  tests:
+    runs-on: ubuntu-latest
+
+    # Cancel previous jobs if a new version was pushed
+    concurrency:
+      group: "${{ github.ref }}-${{ matrix.arch }}"
+      cancel-in-progress: true
+
+    steps:
+      - uses: actions/checkout@v3
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+
+      - name: dependencies
+        run: |
+          pip install -U pip
+          pip install poetry
+          poetry env use python3.11
+          source $(poetry env info -p)/bin/activate
+          #
+          # poetry doesnot work when installing those !?
+          #
+          pip install antlr4-python3-runtime==4.9.3
+          pip install -e .
+          pip install -e benchmate
+          #
+          #
+          #
+          poetry install --with dev
+
+      - name: tests
+        run: |
+          source $(poetry env info -p)/bin/activate
+          coverage run --source=milabench -m pytest --ignore=tests/integration tests/ 
+          coverage report -m
+          coverage xml
+
+      - name: Upload coverage to Codecov
+        uses: codecov/codecov-action@v3
+        with:
+          file: ./coverage.xml
+          flags: unittests
+          env_vars: PLATFORM,PYTHON
+          name: codecov-umbrella
+          fail_ci_if_error: false

README.md

@@ -83,7 +83,7 @@ The benchmark suite has been validated on the following configurations:
 
 | Python version |          GPU                   |   Configuration file |
 |       -        |        -                       |           -          |
-| 3.9.12 (conda) | 2 node x 8xNVIDIA A100 80GB    | config/standard.yaml |
+| 3.11   (conda) | 2 node x 8xNVIDIA A100 80GB    | config/standard.yaml |
 | 3.9.12 (conda) | 8x NVIDIA RTX8000 48GB         | config/standard.yaml |
 | 3.9.16 (conda) | 2x NVIDIA K80                  | config/ci.yaml       |
 | 3.9.16 (conda) | 2x AMD MI100                   | config/ci.yaml       |

benchmarks/accelerate_opt/main.py

@@ -126,16 +126,6 @@ class CustomInitProcessGroupKwargs(InitProcessGroupKwargs):
             world_size=int(os.environ["WORLD_SIZE"]),
         )
 
-        # Accelerator SUCK, it is impossible to make it use hccl
-        # We can bypass Accelerator logic by initializing the group ourselves
-        if acc.device_type == "hpu":
-            acc.init_process_group(
-                init_method=f"tcp://{MASTER_ADDR}:{MASTER_PORT}",
-                timeout=timedelta(seconds=60),
-                rank=int(os.environ["RANK"]),
-                world_size=int(os.environ["WORLD_SIZE"]),
-            )
-
         accelerator = Accelerator(kwargs_handlers=[init_process_group_kwargs])
     else:
         accelerator = Accelerator()

benchmate/benchmate/datagen.py

@@ -7,6 +7,7 @@
 from collections import defaultdict
 from pathlib import Path
 
+import torchcompat.core as acc
 import torch
 from tqdm import tqdm
 
@@ -79,24 +80,40 @@ def generate_sets(root, sets, shape):
         json.dump(sets, fp)
 
 
+def device_count():
+    try:
+        return acc.device_count()
+    except:
+        return 1
+
 def generate_fakeimagenet():
+    # config = json.loads(os.environ["MILABENCH_CONFIG"])
+
     parser = argparse.ArgumentParser()
     parser.add_argument("--batch-size", default=512, type=int)
     parser.add_argument("--batch-count", default=60, type=int)
+    parser.add_argument("--device-count", default=device_count(), type=int)
+    parser.add_argument("--device", default=None, type=str)
     parser.add_argument("--image-size", default=[3, 384, 384], type=int, nargs="+")
     parser.add_argument("--val", default=0.1, type=float, nargs="+")
     parser.add_argument("--test", default=0.1, type=float, nargs="+")
+
     args, _ = parser.parse_known_args()
 
+    if overrides := os.getenv("MILABENCH_TESTING_PREPARE"):
+        bs, bc = overrides.split(",")
+        args.batch_size, args.batch_count = int(bs), int(bc)
+
     data_directory = os.environ["MILABENCH_DIR_DATA"]
-    dest = os.path.join(data_directory, "FakeImageNet")
+
+    dest = os.path.join(data_directory, f"FakeImageNet")
     print(f"Generating fake data into {dest}...")
 
-    total_images = args.batch_size * args.batch_count
+    total_images = args.batch_size * args.batch_count * args.device_count
     size_spec = {
-        "train": total_images,
-        "val": int(total_images * args.val),
-        "test": int(total_images * args.test),
+        f"train": total_images,
+        f"val": int(total_images * args.val),
+        f"test": int(total_images * args.test),
     }
 
     generate_sets(dest, size_spec, args.image_size)

benchmate/benchmate/dataset.py

@@ -1,5 +1,9 @@
 import os
 from collections import defaultdict
+import math
+
+import torch
+from torch.utils.data.distributed import DistributedSampler
 
 
 def no_transform(args):
@@ -48,3 +52,16 @@ def __getitem__(self, item):
 
     def __len__(self):
         return len(self.clip)
+
+
+class ExclusiveSetSampler(DistributedSampler):
+    def __init__(self, dataset, num_sets: int, set_id: int, shuffle: bool = True, seed: int = 0, drop_last: bool = False) -> None:
+        super().__init__(
+            dataset, 
+            num_replicas=num_sets,
+            rank=set_id,
+            shuffle=shuffle,
+            seed=seed,
+            drop_last=drop_last        
+        )
+

benchmate/benchmate/warden.py

@@ -233,12 +233,12 @@ def __exit__(self, *args):
 def destroy(*processes, step=1, timeout=30):
     processes = list(processes)
 
-    def kill(proc, signal):
+    def kill(proc, sig):
         try:
             if getattr(proc, "did_setsid", False):
-                os.killpg(os.getpgid(proc.pid), signal.SIGTERM)
+                os.killpg(os.getpgid(proc.pid), sig)
             else:
-                os.kill(proc.pid, signal.SIGTERM)
+                os.kill(proc.pid, sig)
         except ProcessLookupError:
             pass
 
@@ -249,11 +249,9 @@ def kill(proc, signal):
     elapsed = 0
     def wait(proc):
         nonlocal elapsed
-
         while (ret := proc.poll()) is None and elapsed < timeout:
             time.sleep(step)
             elapsed += step
-
         return ret is None
 
     k = 0
@@ -280,25 +278,34 @@ def wait(proc):
 
 
 @contextmanager
-def process_cleaner():
+def process_cleaner(timeout=30):
     """Delay signal handling until all the processes have been killed"""
 
-    with Protected():
+    def kill_everything(processes):
+        def _():
+            destroy(*processes, timeout=timeout)
+
+        return _
+
+    with Protected() as signalhandler:
         with GPUProcessWarden() as warden:  # => SIGTERM all processes using GPUs
             processes = []
-            try:                            # NOTE: we have not waited much between both signals
+
+            # when a signal is received kill the known processes first
+            # then handle the signal
+            signalhandler.stop = kill_everything(processes)
 
+            try:                            # NOTE: we have not waited much between both signals
                 warden.kill()               # => SIGKILL all processes using GPUs
 
                 yield processes             # => Run milabench, spawning processes for the benches
 
             finally:
                 warden.terminate()          # => SIGTERM all processes using GPUs
 
-                destroy(*processes)         # => SIGTERM+SIGKILL milabench processes
+                destroy(*processes, timeout=timeout) # => SIGTERM+SIGKILL milabench processes
 
                                             # destroy waited 30s
 
                 # warden.__exit__           # => SIGKILL all  processes still using GPUs
 
-

config/slurm.yaml

@@ -6,12 +6,12 @@ multi-node-full:
   # DGX run: 2 nodes x 8 A100 80Go SXM4
   - --partition=staff-idt
   - -w cn-d[003-004]
-  - --ntasks=1
+  - --ntasks=2
   - --gpus-per-task=a100l:8
   - --exclusive
   - --nodes=2
   - --cpus-per-task=128
-  - --time=1:30:00
+  - --time=2:00:00
   - --ntasks-per-node=1
   - --mem=0
 

milabench/_version.py

@@ -1,5 +1,5 @@
 """This file is generated, do not modify"""
 
-__tag__ = "v0.1.0-32-ge9e52501"
-__commit__ = "e9e52501ad92d2ee2dac97e66f601a0458404986"
-__date__ = "2024-06-26 02:37:50 -0400"
+__tag__ = "v0.1.0-24-gdd7f3888"
+__commit__ = "dd7f3888ac0524b3b587e415d1de0e2019cd751f"
+__date__ = "2024-07-03 16:07:47 -0400"