[shardformer/sequence parallel] not support opt of seq-parallel, add warning...

[shardformer/sequence parallel] not support opt of seq-parallel, add warning and fix a bug in gpt2 pp (#4488)

colossalai/shardformer/policies/opt.py

+import warnings
 from functools import partial
 from typing import Callable, Dict, List
 
@@ -39,6 +40,9 @@ class OPTPolicy(Policy):
         from transformers.models.opt.modeling_opt import OPTAttention, OPTDecoder, OPTDecoderLayer
 
         policy = {}
+        if self.shard_config.enable_sequence_parallelism:
+            self.shard_config.enable_sequence_parallelism = False
+            warnings.warn("OPT dosen't support sequence parallelism now, will ignore the sequence parallelism flag.")
 
         if self.shard_config.enable_tensor_parallelism:
             policy[OPTDecoder] = ModulePolicyDescription(sub_module_replacement=[