[hotfix]: modify create_ep_hierarchical_group and add test (#5032)

* feat: modify create_ep_hierarchical_group args

* test: add ep tests

* fix: remove get_process_group_ranks

* fix: fix src_rank

colossalai/moe/_operation.py

@@ -150,7 +150,8 @@ class HierarchicalAllToAll(torch.autograd.Function):
     def forward(
         ctx: Any,
         inputs: Tensor,
-        groups: Tuple[ProcessGroup],
+        groups: Tuple[ProcessGroup, ProcessGroup],
+        src_rank: int
     ) -> Tensor:
         """
         Returns:
@@ -159,12 +160,12 @@ class HierarchicalAllToAll(torch.autograd.Function):
         # TODO: we can reduce comm volume by removing empty capacity
         if ctx is not None:
             ctx.comm_grps = groups
+            ctx.src_rank = src_rank
         intra_node_group, inter_node_group = groups
 
         local_world_size = dist.get_world_size(intra_node_group)
         num_group = dist.get_world_size(inter_node_group) if inter_node_group is not None else 1
         world_size = local_world_size * num_group
-        src_rank = dist.get_process_group_ranks(intra_node_group)[0]
         outputs = torch.empty_like(inputs)
 
         if dist.get_rank() == src_rank:
@@ -196,9 +197,10 @@ class HierarchicalAllToAll(torch.autograd.Function):
         return outputs
 
     @staticmethod
-    def backward(ctx: Any, *grad_outputs) -> Tuple[Tensor, None]:
+    def backward(ctx: Any, *grad_outputs) -> Tuple[Tensor, None, None]:
         return (
-            HierarchicalAllToAll.forward(None, grad_outputs[0], ctx.comm_grps),
+            HierarchicalAllToAll.forward(None, grad_outputs[0], ctx.comm_grps, ctx.src_rank),
+            None,
             None,
         )
 

colossalai/moe/layers.py

@@ -13,7 +13,7 @@ from colossalai.moe.load_balance import LoadBalancer
 from colossalai.moe.manager import MOE_MANAGER
 from colossalai.moe.routers import MoeRouter, get_router_cls
 from colossalai.moe.utils import create_ep_hierarchical_group, get_noise_generator
-from colossalai.tensor.moe_tensor.api import get_dp_group, get_ep_group, get_ep_size
+from colossalai.tensor.moe_tensor.api import get_dp_group, get_ep_group, get_ep_group_ranks, get_ep_size
 
 
 class SparseMLP(nn.Module):
@@ -105,8 +105,11 @@ class SparseMLP(nn.Module):
         if self.expert_parallel is not None:
             self.ep_group = get_ep_group(self.experts)
             self.ep_size = get_ep_size(self.experts)
-            self.ep_hierarchical_group = create_ep_hierarchical_group(
+            self.ep_hierarchical_group = None
-                self.ep_group) if enable_hierarchical_comm else None
+            if enable_hierarchical_comm:
+                self.ep_intra_src_rank, *self.ep_hierarchical_group = create_ep_hierarchical_group(
+                    get_ep_group_ranks(self.experts)
+                )
             self.dp_group = get_dp_group(self.experts)
         else:
             self.ep_group = None
@@ -225,10 +228,10 @@ class SparseMLP(nn.Module):
         """
         if not overlap or dist.get_world_size(self.ep_group) == 1:
             if self.ep_hierarchical_group is not None:
-                expert_input = HierarchicalAllToAll.apply(dispatch_data, self.ep_hierarchical_group)
+                expert_input = HierarchicalAllToAll.apply(dispatch_data, self.ep_hierarchical_group, self.ep_intra_src_rank)
                 expert_input = expert_input.reshape(self.ep_size, self.num_local_experts, -1, self.hidden_size)
                 expert_output = self.experts(expert_input)
-                expert_output = HierarchicalAllToAll.apply(expert_output, self.ep_hierarchical_group)
+                expert_output = HierarchicalAllToAll.apply(expert_output, self.ep_hierarchical_group, self.ep_intra_src_rank)
                 return expert_output
             else:
                 expert_input = AllToAll.apply(dispatch_data, self.ep_group, False)[0]

colossalai/moe/utils.py

@@ -179,15 +179,15 @@ def set_moe_args(config: Any, args: dict):
 
 
 def create_ep_hierarchical_group(
-    ep_group: dist.ProcessGroup,
+    ep_group_ranks: List[int],
     nproc_per_node: Optional[int] = None,
-) -> Tuple[Optional[dist.ProcessGroup],
+) -> Tuple[int, dist.ProcessGroup, Optional[dist.ProcessGroup]]:
-           Optional[dist.ProcessGroup]]:
     """
     e.g., If ep_group = [1, 2, 5, 6], and nproc_per_node = 4
         Then, ep_intra_group = [1, 2] & [5, 6], ep_inter_group = [1, 5] & None
     """
     assert dist.is_initialized(), "Please initialize torch.distributed first."
+    rank = dist.get_rank()
     if nproc_per_node is None:
         nproc_per_node = os.environ.get("LOCAL_WORLD_SIZE")
         assert nproc_per_node is not None, "Please use torchrun to launch the job, or specify nproc_per_node manually."
@@ -197,24 +197,23 @@ def create_ep_hierarchical_group(
             "nproc_per_node should be a divisor of world_size."
     num_node = dist.get_world_size() // nproc_per_node
 
-    rank = dist.get_rank()
+    intra_src_rank = None
-    ep_ranks = dist.get_process_group_ranks(ep_group)
-
     ep_intra_node_group = None
     for i in range(num_node):
         ep_intra_ranks = [
             i * nproc_per_node + j
             for j in range(nproc_per_node)
-            if j in ep_ranks
+            if j in ep_group_ranks
         ]
         group = dist.new_group(ep_intra_ranks)
         if rank in ep_intra_ranks:
             assert ep_intra_node_group is None
             ep_intra_node_group = group
+            intra_src_rank = ep_intra_ranks[0]
 
     ep_inter_node_group = None
     ep_inter_ranks = [
-        ep_ranks[0] + i * nproc_per_node
+        ep_group_ranks[0] + i * nproc_per_node
         for i in range(num_node)
     ]
     if len(ep_inter_ranks) > 1:
@@ -222,4 +221,4 @@ def create_ep_hierarchical_group(
         if rank in ep_inter_ranks:
             ep_inter_node_group = group
 
-    return ep_intra_node_group, ep_inter_node_group
+    return intra_src_rank, ep_intra_node_group, ep_inter_node_group

colossalai/tensor/moe_tensor/api.py

@@ -1,3 +1,5 @@
+from typing import List
+
 import torch
 import torch.distributed as dist
 from torch.distributed import ProcessGroup
@@ -124,7 +126,7 @@ def get_dp_rank(tensor: torch.Tensor) -> int:
     return dist.get_rank(get_dp_group(tensor))
 
 
-def get_ep_group_ranks(tensor: torch.Tensor) -> int:
+def get_ep_group_ranks(tensor: torch.Tensor) -> List[int]:
     """
     Get the expert parallel group ranks of the given tensor.
 
@@ -137,7 +139,7 @@ def get_ep_group_ranks(tensor: torch.Tensor) -> int:
     return tensor.moe_info.ep_group_ranks
 
 
-def get_dp_group_ranks(tensor: torch.Tensor) -> int:
+def get_dp_group_ranks(tensor: torch.Tensor) -> List[int]:
     """
     Get the data parallel group ranks of the given tensor.
 

tests/test_moe/test_moe_ep_tp.py

@@ -1,5 +1,6 @@
 import os
 import warnings
+from typing import Dict
 
 import pytest
 import torch
@@ -123,7 +124,7 @@ def sync_local_from_ep(local_model: SparseMLP, ep_model: SparseMLP, assert_grad_
             local_param.data.copy_(all_param.data)
 
 
-def run_test(rank: int, world_size: int, port: int, num_experts: int, batch_size: int, dim: int, seed: int):
+def run_test(rank: int, world_size: int, port: int, num_experts: int, batch_size: int, dim: int, config: Dict):
     assert batch_size % world_size == 0
 
     colossalai.launch(config=dict(), rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl")
@@ -133,8 +134,9 @@ def run_test(rank: int, world_size: int, port: int, num_experts: int, batch_size
     local_model = SparseMLP(num_experts=num_experts, hidden_size=dim, intermediate_size=dim * 2)
     MOE_MANAGER.__init__()
     MOE_MANAGER.setup(parallel="EP")
-    os.environ["LOCAL_WORLD_SIZE"] = str(world_size)
+    enable_hierarchical_comm = config.get("enable_hierarchical_comm", False)
-    enable_hierarchical_comm = torch.__version__ >= "1.13.1"
+    if enable_hierarchical_comm:
+        os.environ["LOCAL_WORLD_SIZE"] = str(world_size)
     ep_model = SparseMLP(
         num_experts=num_experts,
         hidden_size=dim,
@@ -161,7 +163,6 @@ def run_test(rank: int, world_size: int, port: int, num_experts: int, batch_size
     tp_grad_handler = MoeGradientHandler(tp_model)
 
     rank = dist.get_rank()
-    torch.cuda.manual_seed(seed)
     input_data = torch.randn(batch_size, dim, device=get_current_device())
     micro_batch_size = batch_size // world_size
     index = rank * micro_batch_size
@@ -218,11 +219,14 @@ def run_test(rank: int, world_size: int, port: int, num_experts: int, batch_size
 @pytest.mark.parametrize("num_experts", [4, 64])
 @pytest.mark.parametrize("batch_size", [16])
 @pytest.mark.parametrize("dim", [64])
-@pytest.mark.parametrize("seed", [42, 127])
+@pytest.mark.parametrize("config", [
+    {"enable_hierarchical_comm": False},
+    {"enable_hierarchical_comm": True},
+])
 @rerun_if_address_is_in_use()
-def test_moe_ep_tp(num_experts: int, batch_size: int, dim: int, seed: int):
+def test_moe_ep_tp(num_experts: int, batch_size: int, dim: int, config: Dict):
-    spawn(run_test, 2, num_experts=num_experts, batch_size=batch_size, dim=dim, seed=seed)
+    spawn(run_test, 2, num_experts=num_experts, batch_size=batch_size, dim=dim, config=config)
 
 
 if __name__ == '__main__':
-    test_moe_ep_tp(num_experts=8, batch_size=32, dim=32, seed=42)
+    test_moe_ep_tp(num_experts=8, batch_size=32, dim=32)