[Build] Fix DeepGEMM CUDA 12.9 release builds (#57554)
Signed-off-by: khluu <khluu000@gmail.com> Co-authored-by: OpenAI Codex <noreply@openai.com> Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
Kevin H. Luu committed
Sep 20, 2026 at 20:14 UTC
eb87980585c10f7e5ee8f498fd3830898101d208
2 files changed
+7
-6
cmake/external_projects/deepgemm.cmake
+4
-3
@@ -29,10 +29,11 @@ if(DEEPGEMM_SRC_DIR)
29
else()
30
# Keep in sync with tools/install_deepgemm.sh
31
set(_DEEPGEMM_UPSTREAM_REPO "https://github.com/vllm-project/DeepGEMM.git")
32
- # Pinned to the tip of the fork's dev branch: upstream 2.8.0 plus the SM120
32
+ # Pinned to the fork's dev branch: upstream 2.8.0 plus the SM120
33
# port, the SM90 paged-MQA kv_block=32/next_n=4 port, configurable SwiGLU
34
- # alpha/beta, and SiTU for FP8/FP4 Mega MoE.
35
- set(_DEEPGEMM_UPSTREAM_TAG "a6bbb8000161c0dc3a85a0300a905f76898a7913")
34
+ # alpha/beta, and SiTU for FP8/FP4 Mega MoE. Also includes the CUDA 12.x
35
+ # layout header fix from vllm-project/DeepGEMM#12.
36
+ set(_DEEPGEMM_UPSTREAM_TAG "e1f418c2a4f20818221f6b0e578b4c2f634d4c3f")
37
38
set(_deepgemm_fc_root "${FETCHCONTENT_BASE_DIR}")
39
if(NOT _deepgemm_fc_root)
tools/install_deepgemm.sh
+3
-3
@@ -7,10 +7,10 @@ set -e
7
# Default values
8
# Keep DEEPGEMM_GIT_REF in sync with cmake/external_projects/deepgemm.cmake
9
DEEPGEMM_GIT_REPO="https://github.com/vllm-project/DeepGEMM.git"
10
-# NOTE: This targets the vLLM fork's dev branch tip, which carries the sm120
10
+# NOTE: This targets the vLLM fork's dev branch, which carries the sm120
11
# and sm90 paged-MQA ports plus the SwiGLU alpha/beta and SiTU Mega MoE
12
-# activations
13
-DEEPGEMM_GIT_REF="a6bbb8000161c0dc3a85a0300a905f76898a7913"
12
+# activations, plus the CUDA 12.x layout header fix from vllm-project/DeepGEMM#12.
13
+DEEPGEMM_GIT_REF="e1f418c2a4f20818221f6b0e578b4c2f634d4c3f"
14
WHEEL_DIR=""
15
16
# Parse command line arguments