[shardformer] fix opt test hanging (#4521)

* [shardformer] fix opt test hanging

* fix

* test

* test

* test

* fix test

* fix test

* remove print

* add fix
This commit is contained in:
flybird11111
2023-08-30 14:50:34 +08:00
committed by GitHub
parent e241b74f24
commit d367b88785
13 changed files with 472 additions and 249 deletions

View File

@@ -184,24 +184,33 @@ class T5BasePolicy(Policy):
# use flash attention
if self.shard_config.enable_flash_attention:
policy[T5Attention] = ModulePolicyDescription(method_replacement={
self.append_or_create_method_replacement(description={
'forward': get_t5_flash_attention_forward(),
})
},
policy=policy,
target_key=T5Attention)
# use jit operator
if self.shard_config.enable_jit_fused:
policy[T5LayerFF] = ModulePolicyDescription(method_replacement={
self.append_or_create_method_replacement(description={
'forward': get_jit_fused_T5_layer_ff_forward(),
'dropout_add': get_jit_fused_dropout_add_func(),
})
policy[T5LayerSelfAttention] = ModulePolicyDescription(method_replacement={
},
policy=policy,
target_key=T5LayerFF)
self.append_or_create_method_replacement(description={
'forward': get_T5_layer_self_attention_forward(),
'dropout_add': get_jit_fused_dropout_add_func(),
})
policy[T5LayerCrossAttention] = ModulePolicyDescription(method_replacement={
},
policy=policy,
target_key=T5LayerSelfAttention)
self.append_or_create_method_replacement(description={
'forward': get_T5_layer_cross_attention_forward(),
'dropout_add': get_jit_fused_dropout_add_func(),
})
},
policy=policy,
target_key=T5LayerCrossAttention)
return policy
def postprocess(self):