Skip to content
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

allocate embed norm only on pp0 #261

Merged
merged 3 commits into from
Mar 7, 2022
Merged

allocate embed norm only on pp0 #261

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
9e5ffba
allocate embed norm only on pp0
stas00 Mar 7, 2022
caca39a
text
stas00 Mar 7, 2022
bbee453
Merge remote-tracking branch 'origin/main' into embed-norm-pp0-only
stas00 Mar 7, 2022
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
6 changes: 4 additions & 2 deletions megatron/mpu/layers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,7 +36,7 @@
from .utils import divide
from .utils import split_tensor_along_last_dim
from .utils import VocabUtility
from megatron import get_args
from megatron import get_args, mpu
import deepspeed.runtime.activation_checkpointing.checkpointing as ds_checkpointing


Expand Down Expand Up @@ -188,7 +188,9 @@ def __init__(self, num_embeddings, embedding_dim,
# Allocate weights and initialize.
args = get_args()

if args.use_bnb_optimizer or args.embed_layernorm:
# only the first stage embedding runs this class' forward. The head's embedding does its own
# thing, so don't waste memory allocating LN weights.
if mpu.is_pipeline_first_stage() and (args.use_bnb_optimizer or args.embed_layernorm):
self.norm = torch.nn.LayerNorm(embedding_dim)

if args.use_bnb_optimizer:
Expand Down