[tensor] impl ColoDDP for ColoTensor (#1009)

* impl ColoDDP for ColoTensor

* polish code

colossalai/nn/parallel.py

@@ -0,0 +1,78 @@
+import torch
+import torch.distributed as dist
+from colossalai.core import global_context as gpc
+from colossalai.context import ParallelMode
+from functools import partial
+
+__all__ = ['ColoDDP']
+
+
+def free_storage(data: torch.Tensor) -> None:
+    """Free underlying storage of a Tensor."""
+    if data.storage().size() > 0:
+        # Since we're modifying the Tensor's Storage directly, make sure the Tensor
+        # is the sole occupant of the Storage.
+        assert data.storage_offset() == 0
+        data.storage().resize_(0)
+
+
+class ColoDDP(torch.nn.Module):
+
+    def __init__(self, module: torch.nn.Module) -> None:
+        super().__init__()
+        self.module = module
+        self.comm_stream: torch.cuda.Stream = torch.cuda.Stream()
+        self.dp_world_size = gpc.get_world_size(ParallelMode.DATA)
+        for p in module.parameters():
+            if p.requires_grad:
+                p.register_hook(partial(self.grad_handle, p))
+
+    def parameters(self, recurse: bool = True):
+        return self.module.parameters(recurse)
+
+    def named_parameters(self, prefix: str = '', recurse: bool = True):
+        return self.module.named_parameters(prefix, recurse)
+
+    def forward(self, *args, **kwargs):
+        self.module.zero_grad(set_to_none=True)
+        return self.module(*args, **kwargs)
+
+    def backward(self, loss: torch.Tensor):
+        loss.backward()
+        torch.cuda.current_stream().wait_stream(self.comm_stream)
+        for p in self.module.parameters():
+            p.grad = p._saved_grad
+
+    def grad_handle(self, p, grad):
+        empty_grad = torch.empty_like(grad)
+        free_storage(empty_grad)
+        if self.dp_world_size > 1:
+            grad = grad / self.dp_world_size
+            self.comm_stream.wait_stream(torch.cuda.current_stream())
+            with torch.cuda.stream(self.comm_stream):
+                dist.all_reduce(grad, group=gpc.get_group(ParallelMode.DATA))
+                ColoDDP._save_grad(p, grad)
+            grad.record_stream(self.comm_stream)
+        else:
+            ColoDDP._save_grad(p, grad)
+        return empty_grad
+
+    @staticmethod
+    def _save_grad(p, grad):
+        if hasattr(p, '_saved_grad'):
+            p._saved_grad.add_(grad)
+        else:
+            p._saved_grad = grad
+
+    def zero_grad(self, set_to_none: bool = False) -> None:
+        self.module.zero_grad(set_to_none=True)
+        for p in self.module.parameters():
+            if getattr(p, '_saved_grad', None) is not None:
+                if set_to_none:
+                    p._saved_grad = None
+                else:
+                    if p._saved_grad.grad_fn is not None:
+                        p._saved_grad.detach_()
+                    else:
+                        p._saved_grad.requires_grad_(False)
+                    p._saved_grad.zero_()

tests/test_tensor/test_gpt.py

@@ -9,8 +9,10 @@ from colossalai.utils import ColoInitContext
 from colossalai.tensor import TensorSpec, ComputePattern, ParallelAction, DistSpecManager, distspec
 from colossalai.core import global_context as gpc
 from functools import partial
-from _utils import tensor_equal, tensor_shard_equal
+from _utils import tensor_equal, tensor_shard_equal, set_seed
 from tests.components_to_test.registry import non_distributed_component_funcs
+from torch.nn.parallel import DistributedDataParallel as DDP
+from colossalai.nn.parallel import ColoDDP
 
 
 def init_1d_row_spec(model):
@@ -43,7 +45,7 @@ def check_grad_equal(model, torch_model):
         assert tensor_shard_equal(torch_p.grad, p.grad)
 
 
-def run_gpt(init_spec_func):
+def run_gpt(init_spec_func, use_ddp):
     get_components_func = non_distributed_component_funcs.get_callable('gpt2')
     model_builder, train_dataloader, test_dataloader, optimizer_class, criterion = get_components_func()
 
@@ -51,37 +53,50 @@ def run_gpt(init_spec_func):
         model = model_builder()
     model = model.cuda()
     torch_model = model_builder().cuda()
+    if use_ddp:
+        model = ColoDDP(model)
+        torch_model = DDP(torch_model,
+                          device_ids=[gpc.get_global_rank()],
+                          process_group=gpc.get_group(ParallelMode.DATA))
     for torch_p, p in zip(torch_model.parameters(), model.parameters()):
         torch_p.data.copy_(p)
     init_spec_func(model)
     check_param_equal(model, torch_model)
     model.train()
     torch_model.train()
+    set_seed(gpc.get_local_rank(ParallelMode.DATA))
     for i, (input_ids, attn_mask) in enumerate(train_dataloader):
         logits = model(input_ids, attn_mask)
         torch_logits = torch_model(input_ids, attn_mask)
         assert tensor_equal(torch_logits, logits)
         loss = criterion(logits, input_ids)
         torch_loss = criterion(torch_logits, input_ids)
-        loss.backward()
+        if use_ddp:
+            model.backward(loss)
+        else:
+            loss.backward()
         torch_loss.backward()
         check_grad_equal(model, torch_model)
         if i > 0:
             break
 
 
-def run_dist(rank, world_size, port):
-    config = dict(parallel=dict(tensor=dict(mode="1d", size=world_size),))
+def run_dist(rank, world_size, port, use_ddp):
+    if use_ddp and world_size == 1:
+        return
+    tp_world_size = world_size // 2 if use_ddp else world_size
+    config = dict(parallel=dict(tensor=dict(mode="1d", size=tp_world_size),))
     colossalai.launch(config=config, rank=rank, world_size=world_size, host='localhost', port=port, backend='nccl')
-    run_gpt(init_1d_row_spec)
-    run_gpt(init_1d_col_spec)
+    run_gpt(init_1d_row_spec, use_ddp)
+    run_gpt(init_1d_col_spec, use_ddp)
 
 
 @pytest.mark.dist
 @pytest.mark.parametrize('world_size', [1, 4])
+@pytest.mark.parametrize('use_ddp', [False, True])
 @rerun_if_address_is_in_use()
-def test_gpt(world_size):
-    run_func = partial(run_dist, world_size=world_size, port=free_port())
+def test_gpt(world_size, use_ddp):
+    run_func = partial(run_dist, world_size=world_size, port=free_port(), use_ddp=use_ddp)
     mp.spawn(run_func, nprocs=world_size)