Add new Dino | Update voir (#238)

* update voir

* dinov2 working

* dinov2 working

* dino-giant working

* Always set OMP_NUM_THREADS

* Add some batch size

* Update tests

* update template

* update lock file

* Allow to override the number of CPU seen by milabench

---------

Co-authored-by: pierre.delaunay <[email protected]>

.gitignore

@@ -55,3 +55,4 @@ benchmarks/voir
 benchmarks/*/base/
 benchmarks/lightning/lightning_logs/
 
+benchmarks/*/src/

benchmarks/_templates/simple/requirements.in

@@ -1,2 +1,2 @@
-voir>=0.2.9,<0.3
+voir>=0.2.17,<0.3
 torch

benchmarks/_templates/stdout/main.py

@@ -15,32 +15,45 @@ def criterion(*args, **kwargs):
     return random.normalvariate(0, 1)
 
 
-def main():
-    device = accelerator.fetch_device(0) # <= This is your cuda device
+def prepare_voir():
+    from benchmate.observer import BenchObserver
+    from benchmate.monitor import bench_monitor
 
     observer = BenchObserver(
-        batch_size_fn=lambda batch: 1,
+        accelerator.Event, 
+        earlystop=65,
+        batch_size_fn=lambda x: len(x[0]),
+        raise_stop_program=False,
         stdout=True,
     )
+
+    return observer, bench_monitor
+
+
+def main():
+    device = accelerator.fetch_device(0) # <= This is your cuda device
+
+    observer, monitor = prepare_voir()
+
     # optimizer = observer.optimizer(optimizer)
-    # criterion = observer.criterion(criterion)
 
     dataloader = list(range(6000))
 
-    for epoch in range(10000):
-        for i in observer.iterate(dataloader):
-            # avoid .item()
-            # avoid torch.cuda; use accelerator from torchcompat instead
-            # avoid torch.cuda.synchronize or accelerator.synchronize
-
-            # y = model(i)
-            loss = criterion()
-            # loss.backward()
-            # optimizer.step()
-
-            observer.record_loss(loss)
-
-            time.sleep(0.1)
+    with monitor():
+        for epoch in range(10000):
+            for i in observer.iterate(dataloader):
+                # avoid .item()
+                # avoid torch.cuda; use accelerator from torchcompat instead
+                # avoid torch.cuda.synchronize or accelerator.synchronize
+
+                # y = model(i)
+                loss = criterion()
+                # loss.backward()
+                # optimizer.step()
+
+                observer.record_loss(loss)
+
+                time.sleep(0.1)
 
     assert epoch < 2, "milabench stopped the train script before the end of training"
     assert i < 72, "milabench stopped the train script before the end of training"

benchmarks/_templates/stdout/requirements.in

@@ -1,2 +1,2 @@
-voir>=0.2.9,<0.3
+voir>=0.2.17,<0.3
 torch

benchmarks/_templates/voir/requirements.in

@@ -1,2 +1,2 @@
-voir>=0.2.9,<0.3
+voir>=0.2.17,<0.3
 torch

benchmarks/accelerate_opt/benchfile.py

@@ -12,12 +12,6 @@
 class AccelerateBenchmark(Package):
     base_requirements = "requirements.in"
 
-    def make_env(self):
-        env = super().make_env()
-        value = self.resolve_argument("--cpus_per_gpu", 8)
-        env["OMP_NUM_THREADS"] = str(value)
-        return env
-
     def build_prepare_plan(self):
         return CmdCommand(
             self,