Skip to content

minor change to align with paper for better readliness #3

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

minor change to align with paper for better readliness #3

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
fbb9b4b
align with paper
chivee Jul 26, 2022
6ecc99f
delete unesscary change
chivee Jul 26, 2022
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
15 changes: 7 additions & 8 deletions flash_pytorch/flash_pytorch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -264,7 +264,7 @@ def forward(
j - sequence dimension (target)
"""

b, n, device, g = x.shape[0], x.shape[-2], x.device, self.group_size
b, n, device, c = x.shape[0], x.shape[-2], x.device, self.group_size

# prenorm

Expand Down Expand Up @@ -299,24 +299,23 @@ def forward(

# padding for groups

padding = padding_to_multiple_of(n, g)
padding = padding_to_multiple_of(n, c)

if padding > 0:
quad_q, quad_k, lin_q, lin_k, v = map(lambda t: F.pad(t, (0, 0, 0, padding), value = 0.), (quad_q, quad_k, lin_q, lin_k, v))

mask = default(mask, torch.ones((b, n), device = device, dtype = torch.bool))
mask = F.pad(mask, (0, padding), value = False)

# group along sequence

quad_q, quad_k, lin_q, lin_k, v = map(lambda t: rearrange(t, 'b (g n) d -> b g n d', n = self.group_size), (quad_q, quad_k, lin_q, lin_k, v))
quad_q, quad_k, lin_q, lin_k, v = map(lambda t: rearrange(t, 'b (g c) d -> b g c d', c = c), (quad_q, quad_k, lin_q, lin_k, v))

if exists(mask):
mask = rearrange(mask, 'b (g j) -> b g 1 j', j = g)
mask = rearrange(mask, 'b (g c) -> b g 1 c', c = c)

# calculate quadratic attention output

sim = einsum('... i d, ... j d -> ... i j', quad_q, quad_k) / g
sim = einsum('... i d, ... j d -> ... i j', quad_q, quad_k) / c

sim = sim + self.rel_pos_bias(sim)

Expand All @@ -327,15 +326,15 @@ def forward(
attn = attn.masked_fill(~mask, 0.)

if self.causal:
causal_mask = torch.ones((g, g), dtype = torch.bool, device = device).triu(1)
causal_mask = torch.ones((c,c), dtype = torch.bool, device = device).triu(1)
attn = attn.masked_fill(causal_mask, 0.)

quad_out = einsum('... i j, ... j d -> ... i d', attn, v)

# calculate linear attention output

if self.causal:
lin_kv = einsum('b g n d, b g n e -> b g d e', lin_k, v) / g
lin_kv = einsum('b g n d, b g n e -> b g d e', lin_k, v) / c

# exclusive cumulative sum along group dimension

Expand Down