Skip to content

[Bugfix] rmsnorm: annotate known_block_size on large-M small-N path #639

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
jhinpan wants to merge 3 commits into
base: main
Choose a base branch
from
Open

[Bugfix] rmsnorm: annotate known_block_size on large-M small-N path #639

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
5dbacfa
[Bugfix] rmsnorm: annotate known_block_size on large-M small-N path
jhinpan Jun 2, 2026
3cf86a0
Merge branch 'main' into fix/rmsnorm-large-m-small-n-block-size
jhinpan Jun 4, 2026
afff1f9
test(rmsnorm): cover the large-M small-N path (M>8192, N<=2048)
jhinpan Jun 4, 2026
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
2 changes: 1 addition & 1 deletion kernels/rmsnorm_kernel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -304,7 +304,7 @@ def _build_rmsnorm_large_m_small_n_module(M: int, N: int, dtype_str: str):
BLOCK_THREADS_SPECIAL = BLOCK_M * THREADS_PER_ROW
elem_bits = 32 if dtype_str == "f32" else 16

@flyc.kernel
@flyc.kernel(known_block_size=[BLOCK_THREADS_SPECIAL, 1, 1])
def rmsnorm_large_m_small_n_kernel(
Input: fx.Tensor,
Gamma: fx.Tensor,
Expand Down
5 changes: 5 additions & 0 deletions tests/kernels/test_rmsnorm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -167,6 +167,11 @@ def test_all():
# (16, 512, "bf16"), # BF16
# (1024, 8192, "bf16"), # BF16
(32768, 8192, "bf16"),
# Covers the large-M small-N path in build_rmsnorm_module
# (M > 8192 and N <= 2048): it launches BLOCK_M * THREADS_PER_ROW
# = 512..1024 threads/block, which requires known_block_size.
# N=512 is a real DeepSeek-R1 shape and hits the 1024-thread case.
(16384, 512, "bf16"),
]

do_compare = os.environ.get("ROCDSL_COMPARE_AITER", "0") == "1"
Expand Down