Skip to content

fix(gemm_a8w8_blockscale): prevent scale OOB and support Triton 3.6 #4406

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
valarLip merged 4 commits into from
Jul 29, 2026
+25 −0
Merged

fix(gemm_a8w8_blockscale): prevent scale OOB and support Triton 3.6 #4406

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
965b17e
fix(gemm_a8w8_blockscale_bpreshuffle): honor strided x_scale
bingxche Jul 27, 2026
e6c7e9e
fix(triton): cap gfx950 preshuffle stages on 3.6
bingxche Jul 27, 2026
7583e52
style(triton): keep parsed version for compatibility checks
bingxche Jul 28, 2026
1dbf724
test(gemm_a8w8_blockscale): isolate strided scale coverage
bingxche Jul 28, 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: 2 additions & 0 deletions aiter/ops/gemm_op_a8w8.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -988,6 +988,7 @@ def gemm_a8w8_blockscale_bpreshuffle(
w_scale,
dtype=dtype,
config=_fallback_cfg,
is_x_scale_tranposed=x_scale.stride(0) != 1,
)
config = get_CKGEMM_config(
m, n, k, AITER_CONFIGS.AITER_CONFIG_GEMM_A8W8_BLOCKSCALE_BPRESHUFFLE_FILE
Expand Down Expand Up @@ -1015,6 +1016,7 @@ def gemm_a8w8_blockscale_bpreshuffle(
w_scale,
dtype=dtype,
backend=backend,
is_x_scale_tranposed=x_scale.stride(0) != 1,
)
if config is not None:
libtype = config["libtype"]
Expand Down
12 changes: 12 additions & 0 deletions aiter/ops/triton/gemm/basic/gemm_a8w8_blockscale.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,7 @@

import torch
import triton
from packaging.version import Version

from aiter.ops.triton._triton_kernels.common.splitk_reduce import (
_gemm_splitk_reduce_kernel,
Expand All @@ -25,6 +26,7 @@

_LOGGER = AiterTritonLogger()
_FORCE_GFX1250_EX = os.environ.get("AITER_FORCE_GFX1250_EX", "0") == "1"
_TRITON_VERSION = Version(triton.__version__)

_GLUON_SUPPORTED_ARCHS = ("gfx1250",)

Expand Down Expand Up @@ -305,6 +307,16 @@ def gemm_a8w8_blockscale_preshuffle(
if config is None:
config, _ = _get_config(M, N, K, True, backend=backend)

# Triton 3.6 fails TritonAMDGPUConvertToBufferOps for gfx950 preshuffle
# configs with three pipeline stages. Keep the tuned tile and split-K.
if (
backend == "triton"
and get_arch() == "gfx950"
and _TRITON_VERSION < Version("3.7.0")
and config.get("num_stages", 1) > 2
):
config["num_stages"] = 2

kernel_type_from_config = config.pop("kernel_type", None)
if kernel_type_from_config is not None:
kernel_type = kernel_type_from_config
Expand Down
11 changes: 11 additions & 0 deletions op_tests/test_gemm_a8w8_blockscale.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -116,6 +116,17 @@ def test_gemm(dtype, m, n, k, ck_preshuffle=True, use_flydsl=False):
b, avg_b = run_func(x, gemm_weight, gemm_x_scale, w_scale, dtype)

err_ck = checkAllclose(a, b, msg="ck", catastrophic_check=True)
if ck_preshuffle:
x_scale_strided = x_scale.transpose(0, 1).contiguous().transpose(0, 1)
b_strided = aiter.gemm_a8w8_blockscale_bpreshuffle(
x, gemm_weight, x_scale_strided, w_scale, dtype
)
checkAllclose(
a,
b_strided,
msg="ck strided x_scale",
catastrophic_check=True,
)
ret["ck us"] = avg_b
ret["ck TFLOPS"] = m * n * k * 2 / avg_b / 1e6
ret["ck TB/s"] = (x.nbytes + weight.nbytes) / avg_b / 1e6
Expand Down
Loading