Skip to content

Optimize beam search & add flash attention+xformers support #105

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
xsank wants to merge 9 commits into
base: main
Choose a base branch
from
Open

Optimize beam search & add flash attention+xformers support #105

Show file tree
Hide file tree
Changes from 8 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
4a475bb
opt beam search & add fa ing...
xsank Nov 7, 2025
c052a4f
fix flash attention
xsank Nov 10, 2025
7eaadfc
optimize flash attention
xsank Nov 11, 2025
e0e24b9
for flash attention, casual mask is useless
xsank Nov 11, 2025
277303f
add missing commit
xsank Nov 12, 2025
c2ae62b
add attention meta
xsank Nov 12, 2025
8b8b6a5
only fa init meta
xsank Nov 12, 2025
d9aee69
add xformers support
xsank Nov 14, 2025
66f9512
xformers no need to use seq lens now
xsank Nov 14, 2025
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
2 changes: 2 additions & 0 deletions fireredasr/models/fireredasr_aed.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,7 @@
from fireredasr.models.module.transformer_decoder import TransformerDecoder


@torch.compile(mode="max-autotune")
class FireRedAsrAed(torch.nn.Module):
@classmethod
def from_args(cls, args):
Expand All @@ -28,6 +29,7 @@ def transcribe(self, padded_input, input_lengths,
beam_size=1, nbest=1, decode_max_len=0,
softmax_smoothing=1.0, length_penalty=0.0, eos_penalty=1.0):
enc_outputs, _, enc_mask = self.encoder(padded_input, input_lengths)
self.decoder.clear()
nbest_hyps = self.decoder.batch_beam_search(
enc_outputs, enc_mask,
beam_size, nbest, decode_max_len,
Expand Down
Loading