Skip to content

Reject sparse MMA K tiles with unhandled tails #3367

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

Reject sparse MMA K tiles with unhandled tails #3367

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
48 changes: 48 additions & 0 deletions testing/python/issue/test_tilelang_issue_2605.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,48 @@
"""Regression for per-tile K tails in the sparse MMA fallback."""

import pytest

import tilelang
import tilelang.language as T
import tilelang.testing
from tilelang.cuda.intrinsics.sparse_layout import get_e_factor


def _make_sparse_gemm(K: int, block_K: int):
M = N = 128
meta_dtype = "int16"
e_factor = get_e_factor(T.int8, T.dtype(meta_dtype))

@T.prim_func
def main(
A: T.Tensor((M, K // 2), "int8"),
E: T.Tensor((M, K // e_factor), meta_dtype),
B: T.Tensor((N, K), "int8"),
C: T.Tensor((M, N), "int32"),
):
with T.Kernel(1, 1, threads=128):
A_shared = T.alloc_shared((M, block_K // 2), "int8")
E_shared = T.alloc_shared((M, block_K // e_factor), meta_dtype)
B_shared = T.alloc_shared((N, block_K), "int8")
C_local = T.alloc_fragment((M, N), "int32")
T.clear(C_local)
for k in T.serial(T.ceildiv(K, block_K)):
T.copy(A[0, k * block_K // 2], A_shared)
T.copy(E[0, k * block_K // e_factor], E_shared)
T.copy(B[0, k * block_K], B_shared)
T.gemm_sp(A_shared, E_shared, B_shared, C_local, transpose_B=True)
T.copy(C_local, C[0, 0])

return main


@tilelang.testing.requires_cuda_compute_version_eq(9, 0)
def test_gemm_sp_rejects_per_tile_k_tail():
# The total K is a multiple of the 64-element int8 MMA atom, but each
# 96-element tile would independently drop its final 32 elements.
with pytest.raises(ValueError, match="K tile size 96.*divisible.*64"):
tilelang.compile(
_make_sparse_gemm(K=192, block_K=96),
>
pass_configs={tilelang.PassConfigKey.TL_DISABLE_WARP_SPECIALIZED: True},
)
2 changes: 2 additions & 0 deletions tilelang/cuda/op/gemm_sp/gemm_sp_mma.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -96,6 +96,8 @@ def lower(self, layout_map: dict, target: Target, thread_bounds: Range, thread_i
C_local = self.C
clear_accum = self.clear_accum
assert micro_size_k <= self.K, f"K dimension {self.K} should be >= micro size k {micro_size_k}"
if self.K % micro_size_k != 0:
raise ValueError(f"gemm_sp K tile size {self.K} must be divisible by the sparse MMA K atom {micro_size_k}")
if self.is_gemm_ss():

@T.prim_func
Expand Down
Loading