fix(moe): remove norm&gate force sync (#448)

* add zero broadcast_sync

* delete old sync logic

* fix merged error

* refactor code

* remove some unused function (is norm/gate group)

internlm/initialize/launch.py

@@ -293,10 +293,6 @@ def args_sanity_check():
             model._add_item("moe_use_residual", False)
         if "moe_gate_k" not in model:
             model._add_item("moe_gate_k", 2)
-        assert not (
-            gpc.config.model.num_experts > 1 and gpc.config.parallel.zero1.fsdp
-        ), "FSDP does not support num_experts > 1"
-
     # process the parallel config
     if "sequence_parallel" not in gpc.config.parallel:
         gpc.config.parallel._add_item("sequence_parallel", False)
@@ -345,11 +341,13 @@ def args_sanity_check():
         gpc.config.loss._add_item("moe_loss_coeff", 1.0)
 
     # moe not support overlap and zero1.5 for now
-    if hasattr(gpc.config.model, "num_experts"):
+    if gpc.config.model.get("num_experts", 1) > 1:
+        assert not gpc.config.parallel.zero1.fsdp, "FSDP does not support num_experts > 1"
         assert (
             not optim_ckpt.overlap_sync_grad & optim_ckpt.overlap_sync_param
         ), "not support overlap and moe at the same time"
         assert gpc.config.parallel.zero1.size == -1, "moe only support zero1, set zero1=dict(size=-1,...) can fix this"
+        assert not gpc.config.parallel.sequence_parallel, "moe not support sequence parallel for now"
 
 
 def launch(
@@ -413,7 +411,7 @@ def launch(
             f"data parallel size: {gpc.data_parallel_size}, pipeline parallel size: {gpc.pipeline_parallel_size}, "
             f"tensor parallel size: {gpc.tensor_parallel_size}",
         )
-        if hasattr(gpc.config.model, "num_experts") and gpc.config.model.num_experts > 1:
+        if gpc.config.model.get("num_experts", 1) > 1:
             logger.info(
                 f"Creating MoE with num_experts: {gpc.config.model.num_experts} | "
                 f"expert parallel size: {gpc.expert_parallel_size} | "

internlm/model/modeling_moe.py

@@ -127,11 +127,6 @@ class PackedFlashBaseLayer1D(nn.Module):
         else:
             self.norm1 = nn.LayerNorm(hidden_size, eps=layer_norm_epsilon)
             self.norm2 = nn.LayerNorm(hidden_size, eps=layer_norm_epsilon)
-
-        for param in self.norm1.parameters():
-            param.is_norm = True
-        for param in self.norm2.parameters():
-            param.is_norm = True
         set_fp32_attr_to_module(self.norm1)
         set_fp32_attr_to_module(self.norm2)
 

internlm/model/utils.py

@@ -215,18 +215,6 @@ def is_moe_param(param: torch.Tensor) -> bool:
     return False
 
 
-def is_gate_param(param: torch.Tensor) -> bool:
-    if hasattr(param, "is_gate") and param.is_gate:
-        return True
-    return False
-
-
-def is_norm_param(param: torch.Tensor) -> bool:
-    if hasattr(param, "is_norm") and param.is_norm:
-        return True
-    return False
-
-
 def Silu(w1_o, w2_o):
     return F.silu(w1_o) * w2_o
 

internlm/moe/sharded_moe.py

@@ -352,9 +352,6 @@ class TopKGate(Module):
         self.drop_tokens = drop_tokens
         self.use_rts = use_rts
 
-        for param in self.wg.parameters():
-            param.is_gate = True
-
     def forward(
         self, inputs: torch.Tensor, used_token: torch.Tensor = None
     ) -> Tuple[Tensor, Tensor, Tensor]:  # type: ignore

internlm/solver/optimizer/hybrid_zero_optim.py

@@ -274,12 +274,6 @@ class HybridZeroOptimizer(BaseOptimizer):
     def _is_moe_group(self, param_group):
         return "moe" in param_group.keys() and param_group["moe"]
 
-    def _is_norm_group(self, param_group):
-        return "norm" in param_group.keys() and param_group["norm"]
-
-    def _is_gate_group(self, param_group):
-        return "gate" in param_group.keys() and param_group["gate"]
-
     # TODO check expert dp is correct when enable moe and overlap both
     def _attach_reduction_hook(self):
         # we iterate over the fp16 params
@@ -566,7 +560,6 @@ class HybridZeroOptimizer(BaseOptimizer):
                 last_stage=last_stage,
                 previous_param_norms=previous_param_norms,
                 zero_mode=self._broadcast_parallel_mode[group_id],
-                is_moe_group=self._is_moe_group(self.optim.param_groups[group_id]),
             )
         return total_param_norms
 
@@ -589,7 +582,6 @@ class HybridZeroOptimizer(BaseOptimizer):
                 last_stage=last_stage,
                 previous_zero_grad_count=previous_zero_grad_count,
                 zero_mode=self._broadcast_parallel_mode[group_id],
-                is_moe_group=self._is_moe_group(self.optim.param_groups[group_id]),
             )
         return total_zero_grad_count
 
@@ -675,16 +667,6 @@ class HybridZeroOptimizer(BaseOptimizer):
                     total_param_zero_grad_count[group_name],
                 ) = compute_layer_zero_grad_count(zero_grad_count)
 
-            # Need to allreduce(avg) the norms across different ranks because moe params will not be synced
-            # during allreduce
-            if self._is_moe_group(self.optim.param_groups[group_id]):
-                # model and zero have been reduced!!!
-                pg = gpc.get_group(ParallelMode.EXPERT)
-                scaled_norm = total_norms[group_name] * 1.0 / float(gpc.get_world_size(ParallelMode.EXPERT))
-                scaled_norm_tensor = torch.tensor(scaled_norm, device=get_current_device(), dtype=torch.float)
-                dist.all_reduce(scaled_norm_tensor, group=pg)
-                total_norms[group_name] = scaled_norm_tensor.item()
-
         timer("sync_grad").start()
         self._sync_grad()
         timer("sync_grad").stop()
@@ -767,19 +749,6 @@ class HybridZeroOptimizer(BaseOptimizer):
                 param_shape == flat_fp32_avg_grads.shape
             ), f"fp32 param and grad have different shape {param_shape} vs {flat_fp32_avg_grads.shape}"
 
-            # Parameters shared within a TP group, such as norm and moe gate, have precision inconsistency in gradients.
-            # Therefore, it is recommended to synchronize gradients within the TP group to eliminate accumulated errors.
-            is_tp_sync_groups = (
-                self._is_norm_group(self.optim.param_groups[group_id]),
-                self._is_gate_group(self.optim.param_groups[group_id]),
-            )
-            if any(is_tp_sync_groups):
-                dist.all_reduce(
-                    flat_fp32_avg_grads,
-                    op=dist.ReduceOp.AVG,
-                    group=gpc.get_group(ParallelMode.TENSOR),
-                )
-
             single_grad_partition_groups.append(flat_fp32_avg_grads)
             device = self._fp32_flat_param_groups_of_current_rank[group_id].device
             self._fp32_flat_param_groups_of_current_rank[group_id].grad = flat_fp32_avg_grads.to(device)

internlm/solver/optimizer/utils.py

@@ -262,7 +262,12 @@ def reduce_grads(gradients, parameters, fine_grained=False):
 
 
 def compute_norm(
-    gradients, parameters, last_stage=False, previous_norm=None, norm_type=2, zero_mode=ParallelMode.ZERO1
+    gradients,
+    parameters,
+    last_stage=False,
+    previous_norm=None,
+    norm_type=2,
+    zero_mode=ParallelMode.ZERO1,
 ):
     """Get the norm
     Arguments:
@@ -335,6 +340,15 @@ def compute_norm(
         if torch.is_tensor(total_norm):
             total_norm = total_norm.item()
 
+    # Need to allreduce(avg) the norms across different ranks because moe params will not be synced during allreduce
+    # model and zero have been reduced!!!
+    if zero_mode == ParallelMode.EXPERT_DATA:
+        pg = gpc.get_group(ParallelMode.EXPERT)
+        scaled_norm = total_norm * 1.0 / float(gpc.get_world_size(ParallelMode.DATA))
+        scaled_norm_tensor = torch.tensor(scaled_norm, device=get_current_device(), dtype=torch.float)
+        dist.all_reduce(scaled_norm_tensor, group=pg)
+        total_norm = scaled_norm_tensor.item()
+
     # Scale.
     if total_norm == float("inf") or total_norm == -float("inf"):
         total_norm = -1
@@ -353,7 +367,6 @@ def compute_param_metric(
     previous_param_metrics=None,
     norm_type=2,
     zero_mode=ParallelMode.ZERO1,
-    is_moe_group=False,
 ):
     """Get the metrics of params
     Argumemts:
@@ -424,7 +437,7 @@ def compute_param_metric(
                 total_metrics[param_name] += param_metric
 
     # moe
-    if is_moe_group:
+    if zero_mode == ParallelMode.EXPERT_DATA:
         pg = gpc.get_group(ParallelMode.EXPERT)
         total_metric_values = list(total_metrics.values())
         if isinstance(total_metric_values[0], torch.Tensor):
@@ -463,7 +476,6 @@ def compute_param_norm(
     previous_param_norms=None,
     norm_type=2,
     zero_mode=ParallelMode.ZERO1,
-    is_moe_group=False,
 ):
     """Get the norm of params
     Arguments:
@@ -484,7 +496,6 @@ def compute_param_norm(
         previous_param_metrics=previous_param_norms,
         norm_type=norm_type,
         zero_mode=zero_mode,
-        is_moe_group=is_moe_group,
     )
 
 
@@ -494,7 +505,6 @@ def compute_zero_grad_count(
     last_stage=False,
     previous_zero_grad_count=None,
     zero_mode=ParallelMode.ZERO1,
-    is_moe_group=False,
 ):
     """Get the count of zero gradient for each parameters
     Arguments:
@@ -512,7 +522,6 @@ def compute_zero_grad_count(
         last_stage=last_stage,
         previous_param_metrics=previous_zero_grad_count,
         zero_mode=zero_mode,
-        is_moe_group=is_moe_group,
     )
 
 

internlm/train/utils.py

@@ -4,7 +4,7 @@ import torch
 
 from internlm.core.context.parallel_context import ParallelMode
 from internlm.core.context.parallel_context import global_context as gpc
-from internlm.model.utils import is_gate_param, is_moe_param, is_norm_param
+from internlm.model.utils import is_moe_param
 
 
 def split_params_into_different_groups_for_optimizer(param_groups: Tuple[Dict]) -> Tuple[Dict]:
@@ -39,10 +39,7 @@ def split_params_into_different_groups_for_optimizer(param_groups: Tuple[Dict])
     # create new groups for fp32, norm, moe gate and moe expert
     new_groups = {}
     new_groups["fp32"] = {"name": "fp32", "params": [], "dp_mode": ParallelMode.DATA}
-    if gpc.config.model.get("num_experts", 0) > 1:
+    if gpc.config.model.get("num_experts", 1) > 1:
-        # norm and gate are special group to force sync (when enable MoE).
-        for key in ["gate", "norm"]:
-            new_groups[key] = {"name": key, key: True, "params": [], "dp_mode": ParallelMode.DATA}
         for key in gpc.expert_parallel_group_names:
             new_groups[key] = {"name": key, "moe": True, "params": [], "dp_mode": ParallelMode.EXPERT_DATA}
 
@@ -58,12 +55,7 @@ def split_params_into_different_groups_for_optimizer(param_groups: Tuple[Dict])
         # first split the norm and gate groups, which are special case to force sync (when enable MoE),
         # then fp32 group and the moe group.
         for param in pgroup["params"]:
-            if gpc.config.model.get("num_experts", 0) > 1 and is_norm_param(param):
+            if param.dtype == torch.float32:
-                new_groups["norm"]["params"].append(param)
-            # gate param means MoE is enabled
-            elif is_gate_param(param):
-                new_groups["gate"]["params"].append(param)
-            elif param.dtype == torch.float32:
                 new_groups["fp32"]["params"].append(param)
             # moe param means MoE is enabled
             elif is_moe_param(param):