Skip to content

perf: reduce repeated Triton autotuning for L2Norm #1

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
Kamleecoder wants to merge 1 commit into
base: core_r0.16.0
Choose a base branch
from
Open

perf: reduce repeated Triton autotuning for L2Norm #1

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
20 changes: 6 additions & 14 deletions mindspeed/ops/triton/l2norm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -82,18 +82,17 @@ def l2norm_bwd_kernel1(
for num_warps in [1, 2, 4, 8, 16]
for BT in BT_LIST
],
key=['D', 'NB']
key=['D']
)
@triton.jit
@triton.jit(do_not_specialize=['T'])
def l2norm_fwd_kernel(
x,
y,
rstd,
eps,
T: tl.constexpr,
T,
D: tl.constexpr,
BD: tl.constexpr,
NB: tl.constexpr,
BT: tl.constexpr,
bt_size,
):
Expand All @@ -119,19 +118,18 @@ def l2norm_fwd_kernel(
for num_warps in [1, 2, 4, 8, 16]
for BT in BT_LIST
],
key=['D', 'NB']
key=['D']
)
@triton.jit
@triton.jit(do_not_specialize=['T'])
def l2norm_bwd_kernel(
y,
rstd,
dy,
dx,
eps,
T: tl.constexpr,
T,
D: tl.constexpr,
BD: tl.constexpr,
NB: tl.constexpr,
BT: tl.constexpr,
bt_size,
):
Expand Down Expand Up @@ -187,7 +185,6 @@ def l2norm_fwd(

rstd = torch.empty((T,), dtype=torch.float32, device=x.device)
if D <= 512:
NB = triton.cdiv(T, 2048)
bt_size = 32

def grid(meta):
Expand All @@ -202,7 +199,6 @@ def grid(meta):
T=T,
D=D,
BD=BD,
NB=NB,
bt_size=bt_size,
)
else:
Expand Down Expand Up @@ -235,9 +231,7 @@ def l2norm_bwd(
BD = min(MAX_FUSED_SIZE, triton.next_power_of_2(D))
if D > BD:
raise RuntimeError("This layer norm doesn't support feature dim >= 64KB.")

if D <= 512:
NB = triton.cdiv(T, 2048)
bt_size = 40
l2norm_bwd_kernel[(bt_size,)](
y=y,
Expand All @@ -248,7 +242,6 @@ def l2norm_bwd(
T=T,
D=D,
BD=BD,
NB=NB,
bt_size=bt_size,
)
else:
Expand All @@ -261,7 +254,6 @@ def l2norm_bwd(
D=D,
BD=BD,
)

return dx.view(y_shape_og)


Expand Down