[fp8] fix missing fp8_comm flag in mixtral (#6057)

colossalai/shardformer/modeling/mixtral.py

@@ -31,6 +31,7 @@ from colossalai.moe._operation import (
     all_to_all_uneven,
 )
 from colossalai.pipeline.stage_manager import PipelineStageManager
+from colossalai.quantization.fp8 import all_reduce_fp8
 from colossalai.shardformer.layer._operation import (
     all_to_all_comm,
     gather_forward_split_backward,
@@ -142,6 +143,10 @@ class EPMixtralSparseMoeBlock(ParallelModule):
             for i in range(1, self.ep_size):
                 activate_experts += output_split_sizes[i * self.num_experts_per_ep : (i + 1) * self.num_experts_per_ep]
             activate_experts = (activate_experts > 0).float()
+
+        if self.fp8_communication:
+            all_reduce_fp8(activate_experts, group=self.moe_dp_group)
+        else:
             dist.all_reduce(activate_experts, group=self.moe_dp_group)
 
         input_split_list = input_split_sizes.view(self.ep_size, self.num_experts_per_ep).sum(dim=-1).tolist()

colossalai/shardformer/policies/mixtral.py

@@ -178,6 +178,7 @@ class MixtralPolicy(Policy):
                             "ep_group": self.shard_config.ep_group,
                             "tp_group": self.shard_config.tensor_parallel_process_group,
                             "moe_dp_group": self.shard_config.moe_dp_group,
+                            "fp8_communication": self.shard_config.fp8_communication,
                         },
                     )
                 ],