[Tensor] test model check results for a simple net (#887)

tests/test_tensor/test_net_tp.py → tests/test_tensor/test_model.py

@@ -2,6 +2,7 @@ from tests.components_to_test.registry import non_distributed_component_funcs
 
 import colossalai
 import pytest
+import torch
 import torch.multiprocessing as mp
 from colossalai.testing import parameterize, rerun_if_address_is_in_use
 from colossalai.utils.cuda import get_current_device
@@ -9,15 +10,30 @@ from colossalai.utils import free_port
 from colossalai.utils import ColoInitContext
 from colossalai.tensor import named_params_with_colotensor, TensorSpec, ComputePattern, ParallelAction, ColoTensor
 from colossalai.context import ParallelMode
+from colossalai.core import global_context as gpc
 
 from functools import partial
+import random
+import os
+import numpy as np
 
 
-def run_simple_net():
+def set_seed(seed):
+    random.seed(seed)
+    os.environ['PYTHONHASHSEED'] = str(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    torch.backends.cudnn.deterministic = True
+
+
+def run_1d_row_tp():
     # A simple net with two stacked nn.Linear
     get_components_func = non_distributed_component_funcs.get_callable('simple_net')
     model_builder, train_dataloader, test_dataloader, optimizer_class, criterion = get_components_func()
+    rank = gpc.get_local_rank(ParallelMode.PARALLEL_1D)
 
+    set_seed(1)
     with ColoInitContext(device=get_current_device()):
         model = model_builder(checkpoint=True)
 
@@ -26,6 +42,11 @@ def run_simple_net():
     ]
     spec = TensorSpec(parallel_action_list)
 
+    set_seed(1)
+    if rank == 0:
+        model_torch = model_builder(checkpoint=True)
+        model_torch = model_torch.cuda()
+
     # A naive way to set spec for all weights in Linear
     for name, p in named_params_with_colotensor(model):
         if not isinstance(p, ColoTensor):
@@ -33,15 +54,16 @@ def run_simple_net():
         if 'weight' in name and 'LayerNorm' not in name and 'ln' not in name and 'embed' not in name:
             p.set_spec(spec)
 
-    model.cuda()
-
-    for param in named_params_with_colotensor(model):
-        print(param)
+    model = model.cuda()
 
     for i, (data, label) in enumerate(train_dataloader):
         data = data.to(get_current_device())
         label = label.to(get_current_device())
 
+        torch.distributed.broadcast(data, 0, group=gpc.get_group(ParallelMode.PARALLEL_1D))
+        torch.distributed.broadcast(label, 0, group=gpc.get_group(ParallelMode.PARALLEL_1D))
+
+        # Bcast rank0 data to all processes
         if criterion:
             output = model(data)
             loss = criterion(output, label)
@@ -49,22 +71,34 @@ def run_simple_net():
             output = model(data, label)
             loss = output
 
-        print(loss.torch_tensor())
+        # For reference
+        if rank == 0:
+            if criterion:
+                output_torch = model_torch(data)
+                loss_torch = criterion(output_torch, label)
+            else:
+                output_torch = model_torch(data, label)
+                loss_torch = output_torch
+
+        if rank == 0:
+            # print(loss.torch_tensor().item())
+            # print('loss torch', loss_torch.item())
+            assert torch.allclose(loss.torch_tensor(), loss_torch, rtol=1e-2)
+
         loss.backward()
 
+        if rank == 0:
+            loss_torch.backward()
         if i > 5:
             break
 
-    # TODO(jzy) check the results with col.nn.Linear?
-
 
 def run_dist(rank, world_size, port):
     config = dict(parallel=dict(tensor=dict(mode="1d", size=world_size),))
     colossalai.launch(config=config, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
-    run_simple_net()
+    run_1d_row_tp()
 
 
-@pytest.mark.skip
 @pytest.mark.dist
 @parameterize('world_size', [1, 4])
 @rerun_if_address_is_in_use()