Skip to content

Achieve Over 20% Speedup with PyTorch SDPA #99

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
wxwmd wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Achieve Over 20% Speedup with PyTorch SDPA #99

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d3efabe
accelerate inference with sdpa
Oct 21, 2025
f840f4d
accelerate inference with sdpa
Oct 21, 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
26 changes: 17 additions & 9 deletions fireredasr/models/module/transformer_decoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,6 +40,8 @@ def batch_beam_search(self, encoder_outputs, src_masks,
softmax_smoothing=1.0, length_penalty=0.0, eos_penalty=1.0):
B = beam_size
N, Ti, H = encoder_outputs.size()
flash_attn_enabled = (N == 1)

device = encoder_outputs.device
maxlen = decode_max_len if decode_max_len > 0 else Ti
assert eos_penalty > 0.0 and eos_penalty <= 1.0
Expand Down Expand Up @@ -68,7 +70,9 @@ def batch_beam_search(self, encoder_outputs, src_masks,
dec_output = dec_layer.forward(
dec_output, encoder_outputs,
tgt_mask, src_mask,
cache=caches[i])
cache=caches[i],
flash_attn_enabled=flash_attn_enabled
)
caches[i] = dec_output
i += 1

Expand Down Expand Up @@ -183,7 +187,7 @@ def __init__(self, d_model, n_head, dropout):
self.mlp = PositionwiseFeedForward(d_model, d_model*4, dropout)

def forward(self, dec_input, enc_output, self_attn_mask, cross_attn_mask,
cache=None):
cache=None, flash_attn_enabled=False):
x = dec_input
residual = x
x = self.self_attn_norm(x)
Expand All @@ -193,12 +197,12 @@ def forward(self, dec_input, enc_output, self_attn_mask, cross_attn_mask,
self_attn_mask = self_attn_mask[:, -1:, :]
else:
xq = x
x = self.self_attn(xq, x, x, mask=self_attn_mask)
x = self.self_attn(xq, x, x, mask=self_attn_mask, flash_attn_enabled=flash_attn_enabled)
x = residual + x

residual = x
x = self.cross_attn_norm(x)
x = self.cross_attn(x, enc_output, enc_output, mask=cross_attn_mask)
x = self.cross_attn(x, enc_output, enc_output, mask=cross_attn_mask, flash_attn_enabled=flash_attn_enabled)
x = residual + x

residual = x
Expand Down Expand Up @@ -227,7 +231,7 @@ def __init__(self, d_model, n_head, dropout=0.1):
self.fc = nn.Linear(n_head * self.d_k, d_model)
self.dropout = nn.Dropout(dropout)

def forward(self, q, k, v, mask=None):
def forward(self, q, k, v, mask=None, flash_attn_enabled=False):
bs = q.size(0)

q = self.w_qs(q).view(bs, -1, self.n_head, self.d_k)
Expand All @@ -237,10 +241,14 @@ def forward(self, q, k, v, mask=None):
k = k.transpose(1, 2)
v = v.transpose(1, 2)

if mask is not None:
mask = mask.unsqueeze(1)

output = self.attention(q, k, v, mask=mask)
if flash_attn_enabled:
output = F.scaled_dot_product_attention(
q, k, v, scale=1/(self.d_k ** 0.5)
)
else:
if mask is not None:
mask = mask.unsqueeze(1)
output = self.attention(q, k, v, mask=mask)

output = output.transpose(1, 2).contiguous().view(bs, -1, self.d_model)
output = self.fc(output)
Expand Down