Skip to content

Transformer Decoder: extend docs on input embedding scale #91

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

Open
NeoLegends wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Transformer Decoder: extend docs on input embedding scale #91

Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c8a0d33
Transformer Decoder: extend docs on input embedding scale
NeoLegends Dec 22, 2025
0d38e01
Remove LM mention
NeoLegends Jan 20, 2026
0981626
describe scale
NeoLegends Jan 20, 2026
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
3 changes: 2 additions & 1 deletion i6_models/assemblies/transformer/transformer_decoder_v1.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -128,7 +128,8 @@ class TransformerDecoderV1Config(ModelConfiguration):
block_cfg: Configuration for TransformerDecoderV1.
input_dropout: Dropout applied to the input embedding.
input_embedding_scale: Scale applied to the input embedding.
Set to `None` to apply a (tuned) default.
Set to `None` to apply a default that is suitable for ASR AED decoder models.
Copy link
Member

@albertz albertz Jan 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I would not mention any specific model at all here. I think this is just confusing. I would instead just say what default you use.

When training a pure LM, scale 1.0 may be a better choice.
num_blocks: Number of transformer blocks in the decoder.
num_output: Number of output labels/vocab dim.
logits_bias: Whether to add a bias to the output logits.
Expand Down