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

Add encdec_attention cache to transformer.py to speed up inference. #116

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
Open

Add encdec_attention cache to transformer.py to speed up inference. #116

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
43 changes: 30 additions & 13 deletions thumt/models/transformer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,15 +15,21 @@

class AttentionSubLayer(modules.Module):

def __init__(self, params, name="attention"):
def __init__(self,
params,
name="attention"):
super(AttentionSubLayer, self).__init__(name=name)

self.dropout = params.residual_dropout
self.normalization = params.normalization

with utils.scope(name):
self.attention = modules.MultiHeadAttention(
params.hidden_size, params.num_heads, params.attention_dropout)
params.hidden_size,
params.num_heads,
params.attention_dropout,
)

self.layer_norm = modules.LayerNorm(params.hidden_size)

def forward(self, x, bias, memory=None, state=None):
Expand All @@ -33,11 +39,11 @@ def forward(self, x, bias, memory=None, state=None):
y = x

if self.training or state is None:
y = self.attention(y, bias, memory, None)
else:
y = self.attention(y, bias, memory, kv=None)
else:
kv = [state["k"], state["v"]]
y, k, v = self.attention(y, bias, memory, kv)
state["k"], state["v"] = k, v
state["k"], state["v"] = k, v # update state here

y = nn.functional.dropout(y, self.dropout, self.training)

Expand Down Expand Up @@ -82,11 +88,11 @@ def __init__(self, params, name="layer"):
super(TransformerEncoderLayer, self).__init__(name=name)

with utils.scope(name):
self.self_attention = AttentionSubLayer(params)
self.self_attention = AttentionSubLayer(params)
self.feed_forward = FFNSubLayer(params)

def forward(self, x, bias):
x = self.self_attention(x, bias)
x = self.self_attention(x, bias) # memory=None, kv=None
x = self.feed_forward(x)
return x

Expand All @@ -100,12 +106,15 @@ def __init__(self, params, name="layer"):
self.self_attention = AttentionSubLayer(params,
name="self_attention")
self.encdec_attention = AttentionSubLayer(params,
name="encdec_attention")
name="encdec_attention")
self.feed_forward = FFNSubLayer(params)

def __call__(self, x, attn_bias, encdec_bias, memory, state=None):
x = self.self_attention(x, attn_bias, state=state)
x = self.encdec_attention(x, encdec_bias, memory)
def __call__(self, x, attn_bias, encdec_bias, memory,
self_attention_state=None,
encdec_attention_state=None
):
x = self.self_attention(x, attn_bias, state=self_attention_state) # memory=None
x = self.encdec_attention(x, encdec_bias, memory, state=encdec_attention_state) # add state, but not used in training
x = self.feed_forward(x)
return x

Expand Down Expand Up @@ -157,7 +166,9 @@ def forward(self, x, attn_bias, encdec_bias, memory, state=None):
for i, layer in enumerate(self.layers):
if state is not None:
x = layer(x, attn_bias, encdec_bias, memory,
state["decoder"]["layer_%d" % i])
state["self_attention_kv"]["layer_%d" % i], # self_attention
state["encdec_attention_kv"]["layer_%d" % i] # encdec_attention
)
else:
x = layer(x, attn_bias, encdec_bias, memory, None)

Expand Down Expand Up @@ -320,13 +331,19 @@ def forward(self, features, labels, mode="train", level="sentence"):

def empty_state(self, batch_size, device):
state = {
"decoder": {
"self_attention_kv": {
"layer_%d" % i: {
"k": torch.zeros([batch_size, 0, self.hidden_size],
device=device),
"v": torch.zeros([batch_size, 0, self.hidden_size],
device=device)
} for i in range(self.num_decoder_layers)
},
"encdec_attention_kv": {
"layer_%d" % i: {
"k": None,
"v": None
} for i in range(self.num_decoder_layers)
}
}

Expand Down