Skip to content

fix(gemm): add EVEN_MN heuristic to restore vectorized store in gemm_… #2482

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

Merged
Liang-jianhao97 merged 2 commits into from
Mar 27, 2026
Merged

fix(gemm): add EVEN_MN heuristic to restore vectorized store in gemm_… #2482

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
18faa7f
fix(gemm): add EVEN_MN heuristic to restore vectorized store in gemm_…
Mar 26, 2026
912f5f8
fix(gemm): add EVEN_MN heuristic to batched_gemm_a8w8 kernel
Mar 26, 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
19 changes: 15 additions & 4 deletions aiter/ops/triton/_triton_kernels/gemm/basic/gemm_a16w16.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,7 @@
"NUM_KSPLIT",
"SPLITK_BLOCK_SIZE",
"EVEN_K",
"EVEN_MN",
"cache_modifier",
"activation",
"use_activation",
Expand Down Expand Up @@ -45,6 +46,8 @@
{
"EVEN_K": lambda args: (args["K"] % (args["SPLITK_BLOCK_SIZE"]) == 0)
and (args["SPLITK_BLOCK_SIZE"] % args["BLOCK_SIZE_K"] == 0),
"EVEN_MN": lambda args: (args["M"] % args["BLOCK_SIZE_M"] == 0)
and (args["N"] % args["BLOCK_SIZE_N"] == 0),
}
)
@triton.jit(
Expand Down Expand Up @@ -74,6 +77,7 @@ def _gemm_a16_w16_kernel(
NUM_KSPLIT: tl.constexpr,
SPLITK_BLOCK_SIZE: tl.constexpr,
EVEN_K: tl.constexpr,
EVEN_MN: tl.constexpr,
cache_modifier: tl.constexpr,
activation: tl.constexpr,
use_activation: tl.constexpr,
Expand Down Expand Up @@ -117,8 +121,12 @@ def _gemm_a16_w16_kernel(
# Create pointers for first block of A and B input matrices
offs_k = tl.arange(0, BLOCK_SIZE_K)
offs_k_split = split_k_start + offs_k
offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
if EVEN_MN:
offs_am = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
offs_bn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
else:
offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N

a_ptrs = a_ptr + (
offs_am[:, None] * stride_am + offs_k_split[None, :] * stride_ak
Expand Down Expand Up @@ -177,8 +185,11 @@ def _gemm_a16_w16_kernel(
+ stride_cn * offs_cn[None, :]
+ pid_k * stride_ck
)
c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
tl.store(c_ptrs, c, mask=c_mask)
if EVEN_MN:
tl.store(c_ptrs, c)
else:
c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
tl.store(c_ptrs, c, mask=c_mask)


@triton.jit(repr=_gemm_a16w16_reduce_repr)
Expand Down
25 changes: 19 additions & 6 deletions ...s/gemm/batched/batched_gemm_a8w8_a_per_token_group_prequant_w_per_batched_tensor_quant.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,7 @@
"BLOCK_SIZE_K",
"GROUP_SIZE_M",
"EVEN_K",
"EVEN_MN",
"cache_modifier",
],
)
Expand All @@ -24,6 +25,8 @@
@triton.heuristics(
{
"EVEN_K": lambda args: args["K"] % args["BLOCK_SIZE_K"] == 0,
"EVEN_MN": lambda args: (args["M"] % args["BLOCK_SIZE_M"] == 0)
and (args["N"] % args["BLOCK_SIZE_N"] == 0),
}
)
@triton.jit(
Expand Down Expand Up @@ -64,6 +67,7 @@ def _batched_gemm_a8w8_a_per_token_group_prequant_w_per_batched_tensor_quant_ker
BLOCK_SIZE_K: tl.constexpr,
GROUP_SIZE_M: tl.constexpr,
EVEN_K: tl.constexpr,
EVEN_MN: tl.constexpr,
cache_modifier: tl.constexpr,
):
"""
Expand Down Expand Up @@ -136,8 +140,12 @@ def _batched_gemm_a8w8_a_per_token_group_prequant_w_per_batched_tensor_quant_ker
tl.assume(batch_id >= 0)

offs_k = tl.arange(0, BLOCK_SIZE_K)
offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
if EVEN_MN:
offs_am = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
offs_bn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
else:
offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
a_ptrs = a_ptr + (
batch_id * stride_ab
+ offs_am[:, None] * stride_am
Expand Down Expand Up @@ -175,7 +183,10 @@ def _batched_gemm_a8w8_a_per_token_group_prequant_w_per_batched_tensor_quant_ker
accumulator *= b_scale

if HAS_BIAS:
offs_bias = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
if EVEN_MN:
offs_bias = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
else:
offs_bias = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
bias = tl.load(bias_ptr + batch_id * stride_biasb + offs_bias)
accumulator = accumulator.to(bias_ptr.type.element_ty) + bias[None, :]

Expand All @@ -189,9 +200,11 @@ def _batched_gemm_a8w8_a_per_token_group_prequant_w_per_batched_tensor_quant_ker
+ stride_cm * offs_cm[:, None]
+ stride_cn * offs_cn[None, :]
)
c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)

tl.store(c_ptrs, c, mask=c_mask)
if EVEN_MN:
tl.store(c_ptrs, c)
else:
c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
tl.store(c_ptrs, c, mask=c_mask)


def _get_config(
Expand Down
Loading