Skip to content

Support gfx1250-strict #781

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
ipanfilo wants to merge 3 commits into
base: dev
Choose a base branch
from
Open

Support gfx1250-strict #781

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
d7fad61
Use gfx1250-strict AITER and CK. Add gfx1250-strict AITER V3 filter
ipanfilo Oct 27, 2026
30bd46b
QoLA update
ipanfilo Oct 27, 2026
c8f5ac0
Update QoLA and CK-JIT
ipanfilo Oct 27, 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 3rdparty/QoLA
Open in desktop
Submodule QoLA updated 6 files
+10 −5 patches/aiter/0008-mha-fwd-native-splitkv.patch
+10 −2 patches/aiter/0010-mha-fwd-with-sink.patch
+39 −0 patches/aiter/0012-gfx1250-strict.patch
+74 −0 patches/aiter/0013-ck-gpu-target-ids.patch
+1 −0 qola/build_tools/builder.py
+12 −1 qola/build_tools/config.py
2 changes: 1 addition & 1 deletion 3rdparty/ck_jit
Open in desktop
Submodule ck_jit updated 2 files
+3 −2 ck_jit_prebuild.py
+26 −1 ck_jit_utils.py
6 changes: 4 additions & 2 deletions transformer_engine/common/fused_attn_rocm/fused_attn_ck.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -499,7 +499,8 @@ void fused_attn_ck_fwd_impl(
const bool nvte_log_ck_config = getenv<bool>("NVTE_LOG_CK_CONFIG");

// ASM v3 does not support sink; force CK tile path
bool nvte_ck_uses_fwd_v3 = !has_sink && getenv<int>("NVTE_CK_USES_FWD_V3", 1);
bool nvte_ck_uses_fwd_v3 = !has_sink && getenv<int>("NVTE_CK_USES_FWD_V3", 1) &&
(cuda::sm_arch() != 125 || !cuda::sm_arch_name().starts_with("gfx1250-strict"));
int nvte_ck_how_v3_bf16_cvt = getenv<int>("NVTE_CK_HOW_V3_BF16_CVT", 1);
bool nvte_ck_zero_out_pad = getenv<int>("NVTE_CK_ZERO_OUT_PAD", 1);
NVTE_QKV_Format qkv_format = nvte_get_qkv_format(layout);
Expand Down Expand Up @@ -856,7 +857,8 @@ void fused_attn_ck_bwd_impl(
// ASM v3 bwd does not compute the sink gradient; force the CK tile path so
// has_sink users still get a correct (if slower) d_sink, mirroring the fwd
// ASM v3 sink guard above.
bool nvte_ck_uses_bwd_v3 = getenv<int>("NVTE_CK_USES_BWD_V3", 1) && !has_sink;
bool nvte_ck_uses_bwd_v3 = !has_sink && getenv<int>("NVTE_CK_USES_BWD_V3", 1) &&
(cuda::sm_arch() != 125 || !cuda::sm_arch_name().starts_with("gfx1250-strict"));
bool nvte_ck_is_v3_atomic_fp32 = getenv<int>("NVTE_CK_IS_V3_ATOMIC_FP32", 1);
int nvte_ck_how_v3_bf16_cvt = getenv<int>("NVTE_CK_HOW_V3_BF16_CVT", 1);

Expand Down
Loading