github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-28 13:05:26 +00:00
Code Issues Packages Projects Releases Wiki Activity

[fp8] support hybrid parallel plugin (#5982)

Browse Source 
* support fp8 comm for qwen2 model

* support fp8 comm for qwen2 model

* support fp8 comm for qwen2 model

* fp8

* fix

* bert and bloom

* chatglm and command

* gpt2,gptj,bert, falcon,blip2

* mistral,opy,sam,t5,vit,whisper

* fix

* fix

* fix
This commit is contained in:
Wang Binluo
2024-08-12 18:17:05 +08:00
committed by GitHub
parent f1a3a326c4
commit b2483c8e31
27 changed files with 633 additions and 83 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
colossalai/shardformer/policies/gptj.py
View File

@@ -77,6 +77,7 @@ class GPTJPolicy(Policy):
target_module=col_nn.Linear1D_Col,
kwargs={
"overlap": overlap,
"fp8_communication": self.shard_config.fp8_communication,
},
),
SubModuleReplacementDescription(
@@ -84,6 +85,7 @@ class GPTJPolicy(Policy):
target_module=col_nn.Linear1D_Col,
kwargs={
"overlap": overlap,
"fp8_communication": self.shard_config.fp8_communication,
},
),
SubModuleReplacementDescription(
@@ -91,19 +93,29 @@ class GPTJPolicy(Policy):
target_module=col_nn.Linear1D_Col,
kwargs={
"overlap": overlap,
"fp8_communication": self.shard_config.fp8_communication,
},
),
SubModuleReplacementDescription(
suffix="attn.out_proj",
target_module=col_nn.Linear1D_Row,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
},
),
SubModuleReplacementDescription(
suffix="mlp.fc_in",
target_module=col_nn.Linear1D_Col,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
},
),
SubModuleReplacementDescription(
suffix="mlp.fc_out",
target_module=col_nn.Linear1D_Row,
kwargs={
"fp8_communication": self.shard_config.fp8_communication,
},
),
SubModuleReplacementDescription(
suffix="attn.attn_dropout",
@@ -125,7 +137,14 @@ class GPTJPolicy(Policy):
description=SubModuleReplacementDescription(
suffix="wte",
target_module=embedding_cls,
kwargs={"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by},
kwargs=(
{
"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by,
"fp8_communication": self.shard_config.fp8_communication,
}
if self.shard_config.enable_tensor_parallelism
else {"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by}
),
),
policy=policy,
target_key=GPTJModel,
@@ -264,6 +283,7 @@ class GPTJForCausalLMPolicy(GPTJPolicy):
kwargs={
"gather_output": True,
"make_vocab_size_divisible_by": self.shard_config.make_vocab_size_divisible_by,
"fp8_communication": self.shard_config.fp8_communication,
},
)
]