Skip to content

[moe] support optimizer checkpoint #5015

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
oahzxl merged 11 commits into from
Nov 8, 2023
Merged

[moe] support optimizer checkpoint #5015

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
a77ee71
Refactor MoE Manager setup method
oahzxl Nov 6, 2023
046eb12
unshard optim ckpt
oahzxl Nov 6, 2023
d85e0d1
optim io
oahzxl Nov 6, 2023
890b910
update transformer version
oahzxl Nov 7, 2023
ace359f
update requirements
oahzxl Nov 7, 2023
dcfc0af
update ckpt
oahzxl Nov 7, 2023
0bc0f07
update ckpt
oahzxl Nov 7, 2023
b3f9865
update ckpt
oahzxl Nov 7, 2023
93306e5
fix engine
oahzxl Nov 7, 2023
81562a7
Merge branch 'main' of https://github.com/hpcaitech/ColossalAI into ckpt
oahzxl Nov 7, 2023
3b63c97
fix engine
oahzxl Nov 8, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 3 additions & 6 deletions colossalai/booster/plugin/moe_hybrid_parallel_plugin.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,7 @@
)
from colossalai.cluster import ProcessGroupMesh
from colossalai.interface import ModelWrapper, OptimizerWrapper
from colossalai.moe import MoeCheckpintIO
from colossalai.moe import MoECheckpintIO
from colossalai.pipeline.schedule import OneForwardOneBackwardSchedule
from colossalai.pipeline.stage_manager import PipelineStageManager
from colossalai.shardformer import ShardConfig
Expand Down Expand Up @@ -322,8 +322,8 @@ def seed_worker(worker_id):
**_kwargs,
)

def get_checkpoint_io(self) -> MoeCheckpintIO:
self.checkpoint_io = MoeCheckpintIO(self.dp_group, self.pp_group, self.tp_group, self.zero_stage)
def get_checkpoint_io(self) -> MoECheckpintIO:
self.checkpoint_io = MoECheckpintIO(self.dp_group, self.pp_group, self.tp_group, self.zero_stage)
return self.checkpoint_io

def configure(
Expand Down Expand Up @@ -359,9 +359,6 @@ def configure(
max_norm=self.max_norm,
**self.amp_config,
)
self.checkpoint_io.link_master_and_working_param(
Comment thread
ver217 marked this conversation as resolved.
optimizer.working_to_master_map, optimizer.master_to_working_map
)
else:
optimizer = HybridParallelNaiveOptimizer(
optimizer, model, use_pipeline=self.enable_pipeline_parallelism, param_info=param_info
Expand Down
10 changes: 6 additions & 4 deletions colossalai/inference/tensor_parallel/engine.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,13 +79,15 @@ def __init__(

self.multi_query_group_num = model.config.num_attention_heads
# default to attention_heads
self.multi_query_attention = model.config.multi_query_attention
if hasattr(model.config, "multi_query_attention"):
self.multi_query_attention = getattr(model.config, "multi_query_attention")

if hasattr(model.config, "multi_query_group_num"):
self.multi_query_group_num = model.config.multi_query_group_num
self.multi_query_group_num = getattr(model.config, "multi_query_group_num")

if hasattr(model.config, "num_key_value_heads"):
self.multi_query_group_num = model.config.num_key_value_heads
self.multi_query_group_num = getattr(model.config, "num_key_value_heads")

self.tp_size = -1 # to be set with given shard config in self.prepare_shard_config
self.cache_manager = None

Expand All @@ -108,7 +110,7 @@ def _init_manager(self) -> None:
assert self.head_num % self.tp_size == 0, f"Cannot shard {self.head_num} heads with tp size {self.tp_size}"
self.head_num //= self.tp_size # update sharded number of heads

if self.multi_query_attention:
if hasattr(self, "multi_query_attention"):
# NOTE the logic of MQA tensor parallelism should be specified.
assert (
self.multi_query_group_num % self.tp_size == 0
Expand Down
4 changes: 2 additions & 2 deletions colossalai/moe/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
from .checkpoint import MoeCheckpintIO
from .checkpoint import MoECheckpintIO
from .experts import MLPExperts
from .layers import SparseMLP
from .routers import MoeRouter, Top1Router, Top2Router, TopKRouter
Expand All @@ -13,5 +13,5 @@
"NormalNoiseGenerator",
"UniformNoiseGenerator",
"SparseMLP",
"MoeCheckpintIO",
"MoECheckpintIO",
]
Loading