github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-27 04:33:04 +00:00
Code Issues Packages Projects Releases Wiki Activity

[hotfix] Add layer norm gradients all-reduce for sequence parallel (#4926)

Browse Source 
* [hotfix] Add layer norm gradients all-reduce for sequence parallel. (#4915)

* Add layer norm gradients all-reduce for sequence parallel.

* skip pipeline inference test

* [hotfix] fixing polices of sequence parallel (#4922)

* Add layer norm gradients all-reduce for sequence parallel.

* fix parameter passing when calling get_autopolicy

---------

Co-authored-by: littsk <1214689160@qq.com>

* Hotfix/add grad all reduce for sequence parallel (#4927)

* Add layer norm gradients all-reduce for sequence parallel.


* fix parameter passing when calling get_autopolicy

* fix bug using wrong variables

---------

Co-authored-by: littsk <1214689160@qq.com>

* fix policy initialization

* fix bloom and chatglm policices

* polish code of handling layernorm

* fix moe module

* polish code of class initializing

---------

Co-authored-by: Zhongkai Zhao <kanezz620@gmail.com>
This commit is contained in:
littsk
2023-11-03 13:32:43 +08:00
committed by GitHub
parent d99b2c961a
commit 1a3315e336
30 changed files with 1120 additions and 552 deletions
Download Patch File Download Diff File Expand all files Collapse all files

79
colossalai/shardformer/policies/gpt2.py
View File

@@ -39,6 +39,11 @@ class GPT2Policy(Policy):
from transformers.models.gpt2.modeling_gpt2 import GPT2Attention, GPT2Block, GPT2Model
policy = {}
if self.shard_config.enable_fused_normalization:
norm_cls = col_nn.FusedLayerNorm
else:
norm_cls = col_nn.LayerNorm
use_sequence_parallel = self.shard_config.enable_sequence_parallelism
overlap = self.shard_config.enable_sequence_overlap
if self.shard_config.enable_tensor_parallelism:
@@ -102,33 +107,37 @@ class GPT2Policy(Policy):
)
# optimization configuration
if self.shard_config.enable_fused_normalization:
self.append_or_create_submodule_replacement(
description=SubModuleReplacementDescription(
suffix="ln_f",
target_module=col_nn.FusedLayerNorm,
),
policy=policy,
target_key=GPT2Model,
)
self.append_or_create_submodule_replacement(
description=SubModuleReplacementDescription(
suffix="ln_f",
target_module=norm_cls,
),
policy=policy,
target_key=GPT2Model,
)
self.append_or_create_submodule_replacement(
description=[
SubModuleReplacementDescription(
suffix="ln_1",
target_module=col_nn.FusedLayerNorm,
),
SubModuleReplacementDescription(
suffix="ln_2",
target_module=col_nn.FusedLayerNorm,
),
SubModuleReplacementDescription(
suffix="ln_cross_attn", target_module=col_nn.FusedLayerNorm, ignore_if_not_exist=True
),
],
policy=policy,
target_key=GPT2Block,
)
self.append_or_create_submodule_replacement(
description=[
SubModuleReplacementDescription(
suffix="ln_1",
target_module=norm_cls,
kwargs={"sp_partial_derived": use_sequence_parallel},
),
SubModuleReplacementDescription(
suffix="ln_2",
target_module=norm_cls,
kwargs={"sp_partial_derived": use_sequence_parallel},
),
SubModuleReplacementDescription(
suffix="ln_cross_attn",
target_module=norm_cls,
ignore_if_not_exist=True,
kwargs={"sp_partial_derived": use_sequence_parallel},
),
],
policy=policy,
target_key=GPT2Block,
)
if self.shard_config.enable_flash_attention:
self.append_or_create_method_replacement(
@@ -192,9 +201,6 @@ class GPT2Policy(Policy):
# GPT2Model
class GPT2ModelPolicy(GPT2Policy):
def __init__(self) -> None:
super().__init__()
def module_policy(self):
from transformers.models.gpt2.modeling_gpt2 import GPT2Model
@@ -216,9 +222,6 @@ class GPT2ModelPolicy(GPT2Policy):
# GPT2LMHeadModel
class GPT2LMHeadModelPolicy(GPT2Policy):
def __init__(self) -> None:
super().__init__()
def module_policy(self):
from transformers.models.gpt2.modeling_gpt2 import GPT2LMHeadModel
@@ -263,9 +266,6 @@ class GPT2LMHeadModelPolicy(GPT2Policy):
# GPT2DoubleHeadsModel
class GPT2DoubleHeadsModelPolicy(GPT2Policy):
def __init__(self) -> None:
super().__init__()
def module_policy(self):
from transformers.models.gpt2.modeling_gpt2 import GPT2DoubleHeadsModel
@@ -317,9 +317,6 @@ class GPT2DoubleHeadsModelPolicy(GPT2Policy):
# GPT2ForQuestionAnswering
class GPT2ForQuestionAnsweringPolicy(GPT2Policy):
def __init__(self) -> None:
super().__init__()
def module_policy(self):
from transformers.models.gpt2.modeling_gpt2 import GPT2ForQuestionAnswering
@@ -347,9 +344,6 @@ class GPT2ForQuestionAnsweringPolicy(GPT2Policy):
# GPT2ForTokenClassification
class GPT2ForTokenClassificationPolicy(GPT2Policy):
def __init__(self) -> None:
super().__init__()
def module_policy(self):
from transformers.models.gpt2.modeling_gpt2 import GPT2ForTokenClassification
@@ -387,9 +381,6 @@ class GPT2ForTokenClassificationPolicy(GPT2Policy):
# GPT2ForSequenceClassification
class GPT2ForSequenceClassificationPolicy(GPT2Policy):
def __init__(self) -> None:
super().__init__()
def module_policy(self):
from transformers.models.gpt2.modeling_gpt2 import GPT2ForSequenceClassification