replace flashatten experts by feedforward experts

internlm/model/moe.py

@@ -2,10 +2,10 @@ import typing
 from typing import Dict, Tuple
 
 import torch
-from flash_attn.modules.mlp import ParallelFusedMLP
 
 from internlm.core.context import ParallelMode
 from internlm.core.context import global_context as gpc
+from internlm.model.linear import FeedForward
 from internlm.moe.experts import Experts
 from internlm.moe.sharded_moe import MOELayer, TopKGate
 from internlm.utils.logger import get_logger
@@ -102,17 +102,12 @@ class MoE(torch.nn.Module):
         experts = torch.nn.ModuleList(
             [
                 # TODO have trouble when use internlm.model.linear.FeedForward
-                ParallelFusedMLP(
+                FeedForward(
                     hidden_size,
                     int(hidden_size * gpc.config.model.mlp_ratio),
                     out_features=hidden_size,
-                    activation="gelu_approx",
                     process_group=gpc.get_group(ParallelMode.TENSOR),
-                    bias1=False,
-                    bias2=False,
-                    sequence_parallel=gpc.config.model.sequence_parallel,
-                    checkpoint_lvl=0,
-                    heuristic="auto",
+                    bias=False,
                     device=device,
                     dtype=dtype,
                 )
@@ -143,17 +138,12 @@ class MoE(torch.nn.Module):
         # residual network, see https://arxiv.org/pdf/2201.05596.pdf, seems useful for convergence
         self.use_residual = use_residual
         if use_residual:
-            self.residual_mlp = ParallelFusedMLP(
+            self.residual_mlp = FeedForward(
                 hidden_size,
                 int(hidden_size * gpc.config.model.mlp_ratio),
                 out_features=hidden_size,
-                activation="gelu_approx",
                 process_group=gpc.get_group(ParallelMode.TENSOR),
-                bias1=False,
-                bias2=False,
-                sequence_parallel=gpc.config.model.sequence_parallel,
-                checkpoint_lvl=0,
-                heuristic="auto",
+                bias=False,
                 device=device,
                 dtype=dtype,
             )
@@ -188,9 +178,7 @@ class MoE(torch.nn.Module):
         return output, self.moe_layer.l_aux, self.moe_layer.exp_counts
 
 
-def split_params_into_different_moe_groups_for_optimizer(
-    param_groups: Tuple[Dict], max_group_size=178956971
-) -> Tuple[Dict]:
+def split_params_into_different_moe_groups_for_optimizer(param_groups: Tuple[Dict], max_group_size=None) -> Tuple[Dict]:
     """Split parameters into different MoE groups for optimizer
     Compatiable with muiltiple param groups, each should have a name
 

internlm/moe/experts.py

@@ -37,7 +37,7 @@ class Experts(torch.nn.Module):
         for expert in self.experts:
             # TODO: Create param groups to handle expert + data case (e.g. param.group = moe_group)
             for _, param in expert.named_parameters():
-                param.belong_expert = True
+                param.is_expert = True
                 param.group_name = expert_group_name
 
     def forward(self, inputs):

internlm/train/training_internlm.py

@@ -382,7 +382,7 @@ def record_current_batch_training_metrics(
         infos = {
             "tflops": tflops,
             "step": batch_count,
-            "loss": loss.item(),
+            "loss": loss.item() - moe_loss.item(),
             "moe_loss": moe_loss.item(),
             "tgs (tokens/gpu/second)": tk_per_gpu,
             "lr": lr,