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

[transformer] try to fix mga in onnxruntime #2519

Merged
merged 2 commits into from
May 8, 2024
Merged

[transformer] try to fix mga in onnxruntime #2519

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
054c149
[transformer] try to fix mga in onnxruntime
Mddct May 8, 2024
a5a7326
fix v shape
Mddct May 8, 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
35 changes: 25 additions & 10 deletions wenet/transformer/attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -204,16 +204,31 @@ def _update_kv_and_cache(
new_cache = (k, v)
# for multi query or multi group attention
if self.h_kv != self.h and self.h_kv != 1:
k = torch.repeat_interleave(
k,
self.h // self.h_kv,
dim=-3,
)
v = torch.repeat_interleave(
v,
self.h // self.h_kv,
dim=-3,
)
# NOTE: onnxruntime issues:
# https://github.com/wenet-e2e/wenet/issues/2517
# k = torch.repeat_interleave(
# k,
# self.h // self.h_kv,
# dim=-3,
# )
# v = torch.repeat_interleave(
# v,
# self.h // self.h_kv,
# dim=-3,
# )
n_repeat = self.h // self.h_kv
k_shape = k.size()
k = k.unsqueeze(-3).expand(
k_shape[:-2] + torch.Size([n_repeat]) +
k_shape[-2:]).reshape(k_shape[:-3] +
torch.Size([self.h_kv * n_repeat]) +
k_shape[-2:])
v_shape = v.size()
v = v.unsqueeze(-3).expand(
v_shape[:-2] + torch.Size([n_repeat]) +
v_shape[-2:]).reshape(v_shape[:-3] +
torch.Size([self.h_kv * n_repeat]) +
v_shape[-2:])
return k, v, new_cache

def forward(
Expand Down
Loading