Skip to content

Separate concat_ua into concat_u + concat_x for STUNormPostProcConfig #441

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

Closed
melody-lui3 wants to merge 1 commit into from
Closed

Separate concat_ua into concat_u + concat_x for STUNormPostProcConfig #441

Show file tree
Hide file tree
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
6 changes: 4 additions & 2 deletions generative_recommenders/modules/stu.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -346,7 +346,8 @@ def forward(
group_norm=self._use_group_norm,
num_heads=self._num_heads,
linear_dim=self._hidden_dim,
concat_ux=True,
concat_u=True,
concat_x=True,
training=self.training,
kernel=self.hammer_kernel(),
recompute_y_in_backward=self._recompute_y,
Expand Down Expand Up @@ -412,7 +413,8 @@ def cached_forward(
group_norm=self._use_group_norm,
num_heads=self._num_heads,
linear_dim=self._hidden_dim,
concat_ux=True,
concat_u=True,
concat_x=True,
training=self.training,
kernel=self.hammer_kernel(),
recompute_y_in_backward=self._recompute_y,
Expand Down
14 changes: 9 additions & 5 deletions generative_recommenders/ops/hstu_compute.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -101,7 +101,8 @@ def hstu_compute_output(
linear_dim: int,
dropout_ratio: float,
training: bool,
concat_ux: bool,
concat_u: bool,
concat_x: bool,
group_norm: bool,
recompute_y_in_backward: bool,
kernel: HammerKernel = HammerKernel.PYTORCH,
Expand All @@ -117,7 +118,8 @@ def hstu_compute_output(
eps=norm_eps,
dropout_ratio=dropout_ratio,
training=training,
concat_ux=concat_ux,
concat_u=concat_u,
concat_x=concat_x,
group_norm=group_norm,
num_heads=num_heads,
linear_dim=linear_dim,
Expand All @@ -134,7 +136,7 @@ def hstu_compute_output(
eps=norm_eps,
dropout_ratio=dropout_ratio,
training=training,
concat_ux=concat_ux,
concat_ux=concat_u and concat_x,
num_heads=num_heads,
linear_dim=linear_dim,
)
Expand All @@ -147,7 +149,8 @@ def hstu_compute_output(
eps=norm_eps,
dropout_ratio=dropout_ratio,
training=training,
concat_ux=concat_ux,
concat_u=concat_u,
concat_x=concat_x,
)
return triton_cc_addmm(x, y, output_weight)
else:
Expand All @@ -161,7 +164,8 @@ def hstu_compute_output(
eps=norm_eps,
dropout_ratio=dropout_ratio,
training=training,
concat_ux=concat_ux,
concat_u=concat_u,
concat_x=concat_x,
group_norm=group_norm,
num_heads=num_heads,
linear_dim=linear_dim,
Expand Down
19 changes: 14 additions & 5 deletions generative_recommenders/ops/pytorch/pt_hstu_linear.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,8 @@ def pytorch_norm_mul_dropout(
dropout_ratio: float,
training: bool,
silu_u: bool = False,
concat_ux: bool = False,
concat_u: bool = False,
concat_x: bool = False,
group_norm: bool = False,
num_heads: int = 1,
linear_dim: int = -1,
Expand All @@ -47,6 +48,8 @@ def pytorch_norm_mul_dropout(
bias=bias.to(torch.float32),
eps=eps,
).view(-1, num_heads * linear_dim)
if concat_u and concat_x:
y = torch.cat([u, x, y], dim=1)
else:
y = u * F.layer_norm(
x,
Expand All @@ -55,8 +58,12 @@ def pytorch_norm_mul_dropout(
bias=bias.to(torch.float32),
eps=eps,
)
if concat_ux:
y = torch.cat([u, x, y], dim=1)
if concat_u and concat_x:
y = torch.cat([u, x, y], dim=1)
elif concat_u:
y = torch.cat([u, y], dim=1)
elif concat_x:
y = torch.cat([x, y], dim=1)
y = F.dropout(
y,
p=dropout_ratio,
Expand All @@ -76,7 +83,8 @@ def pytorch_hstu_compute_output(
dropout_ratio: float,
training: bool,
silu_u: bool = False,
concat_ux: bool = False,
concat_u: bool = False,
concat_x: bool = False,
group_norm: bool = False,
num_heads: int = 1,
linear_dim: int = -1,
Expand All @@ -91,7 +99,8 @@ def pytorch_hstu_compute_output(
dropout_ratio=dropout_ratio,
training=training,
silu_u=silu_u,
concat_ux=concat_ux,
concat_u=concat_u,
concat_x=concat_x,
group_norm=group_norm,
num_heads=num_heads,
linear_dim=linear_dim,
Expand Down
28 changes: 20 additions & 8 deletions generative_recommenders/ops/tests/hstu_compute_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,7 +37,8 @@ class HSTUComputeTest(unittest.TestCase):
N=st.integers(min_value=1000, max_value=1000),
D=st.integers(min_value=128, max_value=128),
L=st.integers(min_value=512, max_value=512),
concat_ux=st.booleans(),
concat_u=st.booleans(),
concat_x=st.booleans(),
group_norm=st.booleans(),
num_heads=st.sampled_from([4]),
training=st.just(False),
Expand Down Expand Up @@ -69,7 +70,8 @@ def test_compute_output(self, *args, **kwargs) -> None:
N=st.just(1500000),
D=st.just(512),
L=st.just(512),
concat_ux=st.sampled_from([True]),
concat_u=st.sampled_from([True]),
concat_x=st.sampled_from([True]),
group_norm=st.sampled_from([False]),
num_heads=st.sampled_from([4]),
training=st.just(False),
Expand Down Expand Up @@ -97,7 +99,8 @@ def _test_compute_output(
N: int,
D: int,
L: int,
concat_ux: bool,
concat_u: bool,
concat_x: bool,
group_norm: bool,
num_heads: int,
training: bool,
Expand Down Expand Up @@ -145,10 +148,17 @@ def _test_compute_output(
.requires_grad_()
)
norm_eps = 1e-6
# When group_norm=True, only concat_ux = concat_u and concat_x is supported
if group_norm:
L_mult = 3 if (concat_u and concat_x) else 1
else:
L_mult = 1
if concat_u:
L_mult += 1
if concat_x:
L_mult += 1
output_weight = (
torch.empty(
(L * 3 if concat_ux else L, D), dtype=dtype, device=torch.device("cuda")
)
torch.empty((L * L_mult, D), dtype=dtype, device=torch.device("cuda"))
.uniform_(-0.1, 0.1)
.requires_grad_()
)
Expand All @@ -168,7 +178,8 @@ def _test_compute_output(
norm_eps=norm_eps,
dropout_ratio=dropout_ratio,
output_weight=output_weight,
concat_ux=concat_ux,
concat_u=concat_u,
concat_x=concat_x,
group_norm=group_norm,
num_heads=num_heads,
linear_dim=L // num_heads,
Expand Down Expand Up @@ -204,7 +215,8 @@ def _test_compute_output(
norm_eps=norm_eps,
dropout_ratio=dropout_ratio,
output_weight=output_weight,
concat_ux=concat_ux,
concat_u=concat_u,
concat_x=concat_x,
group_norm=group_norm,
num_heads=num_heads,
linear_dim=L // num_heads,
Expand Down
Loading