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

Jump to bottom

Add intermediate_size to GPT-NeoX models #1212

Merged
merged 11 commits into from
Sep 7, 2024
Merged

Add intermediate_size to GPT-NeoX models #1212

Changes from 1 commit
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
6c6a46b
Update transformer.py -> Add `intermediate_size`
dtamayo-nlp May 10, 2024
a80e302
add support for rwkv and mamba and add todos about swiglu
jahatef Jun 19, 2024
22149f0
refactor activations and mlps
jahatef Jul 25, 2024
148a86f
change llama config to swiglu
jahatef Jul 25, 2024
63e8677
fixes gelu fusion
jahatef Jul 26, 2024
ae544cf
pre-commit run
jahatef Aug 5, 2024
1170078
add assert message to mamba linear
Quentin-Anthony Aug 6, 2024
048826f
Update 1-3B.yml
jahatef Sep 4, 2024
8f5d152
Update 1-3B.yml
jahatef Sep 4, 2024
0da5ac8
fixes various issues
jahatef Sep 4, 2024
3400c43
add back swiglu check
jahatef Sep 7, 2024
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
Next Next commit
Update transformer.py -> Add intermediate_size
  • Loading branch information
@dtamayo-nlp
dtamayo-nlp committed May 10, 2024
commit 6c6a46bbc84f3f66e5019a3f4c3cce69c27ec80c
20 changes: 13 additions & 7 deletions megatron/model/transformer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,13 +102,19 @@ def __init__(
self.activation_type = neox_args.activation
self.bias_gelu_fusion = neox_args.bias_gelu_fusion

# auto scale so geglu has equal parameters
ff_mult = int(4 * 2 / 3) if self.activation_type == "geglu" else 4
ff_dim = (
int(ff_mult * neox_args.hidden_size) * 2
if self.activation_type == "geglu"
else ff_mult * neox_args.hidden_size
)

if neox_args.intermediate_size:
ff_dim = neox_args.intermediate_size

else:
# auto scale so geglu has equal parameters
ff_mult = int(4 * 2 / 3) if self.activation_type == "geglu" else 4
ff_dim = (
int(ff_mult * neox_args.hidden_size) * 2
if self.activation_type == "geglu"
else int(ff_mult * neox_args.hidden_size)
)

self.dense_h_to_4h = mpu.ColumnParallelLinear(
neox_args=neox_args,
input_size=neox_args.hidden_size,
Expand Down