github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-27 20:46:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

[shardformer] support lazy init (#4202)

Browse Source 
* [shardformer] support lazy init

* [shardformer] linear support lazy init

* [shardformer] embedding support lazy init

* [shardformer] norm support lazy init

* [shardformer] fused linear support lazy init

* [test] update shardformer test layer

* [test] shardformer with lazy init fit ddp

* [lazy] hotfix deepcopy of param

* [shardformer] fix bert policy and update test

* [shardformer] fix bloom policy and update test

* [shardformer] fix opt policy and update test

* [shardformer] fix t5 policy and update test

* [shardformer] fix gpt2 policy and update test

* [shardformer] fix llama policy and update test
This commit is contained in:
Hongxin Liu
2023-07-10 10:48:53 +08:00
parent f3bcc292c8
commit 890774b2fb
25 changed files with 263 additions and 157 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
colossalai/shardformer/policies/bert.py
View File

@@ -46,11 +46,12 @@ class BertPolicy(Policy):
Reshape the Embedding layer to make the embedding dimension divisible by world_size
"""
# TODO:
vocab_size = self.model.config.vocab_size
world_size = self.shard_config.tensor_parallel_size
if vocab_size % world_size != 0:
new_vocab_size = vocab_size + world_size - vocab_size % world_size
self.model.resize_token_embeddings(new_vocab_size)
if self.shard_config.enable_tensor_parallelism:
vocab_size = self.model.config.vocab_size
world_size = self.shard_config.tensor_parallel_size
if vocab_size % world_size != 0:
new_vocab_size = vocab_size + world_size - vocab_size % world_size
self.model.resize_token_embeddings(new_vocab_size)
return self.model
def module_policy(self):
@@ -229,10 +230,11 @@ class BertForPreTrainingPolicy(BertPolicy):
return []
def postprocess(self):
binding_map = {"bert.embeddings.word_embeddings.weight": "cls.predictions.decoder.weight"}
for k, v in binding_map.items():
param = getattr_(self.model, k)
setattr_(self.model, v, param)
if self.shard_config.enable_tensor_parallelism:
binding_map = {"bert.embeddings.word_embeddings.weight": "cls.predictions.decoder.weight"}
for k, v in binding_map.items():
param = getattr_(self.model, k)
setattr_(self.model, v, param)
return self.model
@@ -269,10 +271,11 @@ class BertLMHeadModelPolicy(BertPolicy):
return []
def postprocess(self):
binding_map = {"bert.embeddings.word_embeddings.weight": "cls.predictions.decoder.weight"}
for k, v in binding_map.items():
param = getattr_(self.model, k)
setattr_(self.model, v, param)
if self.shard_config.enable_tensor_parallelism:
binding_map = {"bert.embeddings.word_embeddings.weight": "cls.predictions.decoder.weight"}
for k, v in binding_map.items():
param = getattr_(self.model, k)
setattr_(self.model, v, param)
return self.model
@@ -288,10 +291,11 @@ class BertForMaskedLMPolicy(BertPolicy):
return module_policy
def postprocess(self):
binding_map = {"bert.embeddings.word_embeddings.weight": "cls.predictions.decoder.weight"}
for k, v in binding_map.items():
param = getattr_(self.model, k)
setattr_(self.model, v, param)
if self.shard_config.enable_tensor_parallelism:
binding_map = {"bert.embeddings.word_embeddings.weight": "cls.predictions.decoder.weight"}
for k, v in binding_map.items():
param = getattr_(self.model, k)
setattr_(self.model, v, param)
return self.model