[deepseek] replace attn (a workaround for bug in transformers)

colossalai/shardformer/policies/deepseek.py

@@ -1,4 +1,3 @@
-import warnings
 from functools import partial
 from typing import Callable, Dict, List, Union
 
@@ -195,11 +194,36 @@ class DeepseekPolicy(Policy):
             )
 
         if self.shard_config.enable_flash_attention:
-            warnings.warn(
-                "Flash attention has already been replaced in deepseek, and now set enable_flash_attention = False."
-            )
-            self.shard_config.enable_flash_attention = False
+            # NOTE: there is a bug for toggling flash attention in AutoModel, which has to be used for deepseek right now
+            from transformers.dynamic_module_utils import get_class_from_dynamic_module
 
+            flash_attn_cls = get_class_from_dynamic_module(
+                "deepseek-ai/deepseek-moe-16b-base--modeling_deepseek.DeepseekFlashAttention2",
+                "deepseek-ai/deepseek-moe-16b-base",
+            )
+
+            class TargetFlashAttn:
+                def __init__(self):
+                    raise RuntimeError("This class should not be instantiated")
+
+                @staticmethod
+                def from_native_module(original_attn: nn.Module, *args, **kwargs) -> nn.Module:
+                    flash_attn_module = flash_attn_cls(original_attn.config, original_attn.layer_idx)
+                    flash_attn_module.q_proj = original_attn.q_proj
+                    flash_attn_module.k_proj = original_attn.k_proj
+                    flash_attn_module.v_proj = original_attn.v_proj
+                    flash_attn_module.o_proj = original_attn.o_proj
+                    flash_attn_module.rotary_emb = original_attn.rotary_emb
+                    return flash_attn_module
+
+            self.append_or_create_submodule_replacement(
+                description=SubModuleReplacementDescription(
+                    suffix="self_attn",
+                    target_module=TargetFlashAttn,
+                ),
+                policy=policy,
+                target_key="DeepseekDecoderLayer",
+            )
         return policy
 
     def postprocess(self):

tests/test_shardformer/test_model/test_shard_deepseek_ghz.py

@@ -220,6 +220,7 @@ def check_deepseek(rank, world_size, port):
     run_deepseek_test()
 
 
+@pytest.mark.skip("redundant")
 @pytest.mark.dist
 @rerun_if_address_is_in_use()
 @clear_cache_before_run()