github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-27 12:43:02 +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

83
colossalai/shardformer/policies/chatglm2.py
View File

@@ -45,6 +45,16 @@ class ChatGLMPolicy(Policy):
policy = {}
if self.shard_config.enable_fused_normalization:
if self.model.config.rmsnorm:
norm_cls = col_nn.FusedRMSNorm
else:
norm_cls = col_nn.FusedLayerNorm
else:
if self.model.config.rmsnorm:
norm_cls = col_nn.RMSNorm
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:
@@ -96,52 +106,34 @@ class ChatGLMPolicy(Policy):
)
# optimization configuration
if self.shard_config.enable_fused_normalization:
if not self.model.config.rmsnorm:
self.append_or_create_submodule_replacement(
description=[
SubModuleReplacementDescription(suffix="input_layernorm", target_module=col_nn.FusedLayerNorm),
SubModuleReplacementDescription(
suffix="post_attention_layernorm", target_module=col_nn.FusedLayerNorm
),
],
policy=policy,
target_key=GLMBlock,
)
self.append_or_create_submodule_replacement(
description=[
SubModuleReplacementDescription(
suffix="input_layernorm",
target_module=norm_cls,
kwargs={"sp_partial_derived": use_sequence_parallel},
),
SubModuleReplacementDescription(
suffix="post_attention_layernorm",
target_module=norm_cls,
kwargs={"sp_partial_derived": use_sequence_parallel},
),
],
policy=policy,
target_key=GLMBlock,
)
if self.model.config.post_layer_norm:
self.append_or_create_submodule_replacement(
description=[
SubModuleReplacementDescription(
suffix="encoder.final_layernorm", target_module=col_nn.FusedLayerNorm
)
],
policy=policy,
target_key=ChatGLMModel,
)
else:
self.append_or_create_submodule_replacement(
description=[
SubModuleReplacementDescription(suffix="input_layernorm", target_module=col_nn.FusedRMSNorm),
SubModuleReplacementDescription(
suffix="post_attention_layernorm", target_module=col_nn.FusedRMSNorm
),
],
policy=policy,
target_key=GLMBlock,
)
if self.model.config.post_layer_norm:
self.append_or_create_submodule_replacement(
description=[
SubModuleReplacementDescription(
suffix="encoder.final_layernorm", target_module=col_nn.FusedRMSNorm
)
],
policy=policy,
target_key=ChatGLMModel,
if self.model.config.post_layer_norm:
self.append_or_create_submodule_replacement(
description=[
SubModuleReplacementDescription(
suffix="encoder.final_layernorm",
target_module=norm_cls,
)
],
policy=policy,
target_key=ChatGLMModel,
)
# use flash attention
if self.shard_config.enable_flash_attention:
@@ -224,9 +216,6 @@ class ChatGLMPolicy(Policy):
class ChatGLMModelPolicy(ChatGLMPolicy):
def __init__(self) -> None:
super().__init__()
def module_policy(self):
pass