Update DeepGEMM commit (#61744)

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>

.buildkite/check-torch-abi.py

@@ -17,7 +17,6 @@ ALLOWED_UNSTABLE_LIBRARIES: tuple[str, ...] = (
     # _C_stable_libtorch, but these two modules still bind libtorch directly.
     "_C.abi3.so",
     "_rocm_C.abi3.so",
-    "third_party/deep_gemm/_C*.so",
 )
 
 

cmake/external_projects/deepgemm.cmake

@@ -32,8 +32,9 @@ else()
   # Pinned to the fork's dev branch: upstream's 26/09/30 release (2.8.1)
   # plus the SM120 port, the SM90 paged-MQA kv_block=32/next_n=4 port,
   # configurable SwiGLU alpha/beta, SiTU for FP8/FP4 Mega MoE, and NVFP4
-  # Mega MoE. TORCH_LIBRARY build with a CPython 3.10 stable-ABI floor.
-  set(_DEEPGEMM_UPSTREAM_TAG "6ce1c5ec6627a6bf11dbcc698b0b1e19c9248432")
+  # Mega MoE, plus swiglustep for BF16/FP8/NVFP4 Mega MoE.
+  # Uses the PyTorch 2.11 stable ABI and CPython 3.10 limited API.
+  set(_DEEPGEMM_UPSTREAM_TAG "a7a927731b109fdfec770df7c2a5be5d7e2216b5")
 
   set(_deepgemm_fc_root "${FETCHCONTENT_BASE_DIR}")
   if(NOT _deepgemm_fc_root)

tools/install_deepgemm.sh

@@ -9,8 +9,9 @@ set -e
 DEEPGEMM_GIT_REPO="https://github.com/vllm-project/DeepGEMM.git"
 # NOTE: This targets the vLLM fork's dev branch: upstream's 26/09/30 release
 # plus the sm120 and sm90 paged-MQA ports, the SwiGLU alpha/beta and SiTU
-# Mega MoE activations, and NVFP4 Mega MoE.
-DEEPGEMM_GIT_REF="6ce1c5ec6627a6bf11dbcc698b0b1e19c9248432"
+# Mega MoE activations, NVFP4 Mega MoE, and BF16/FP8/NVFP4 swiglustep.
+# Uses the PyTorch 2.11 stable ABI and CPython 3.10 limited API.
+DEEPGEMM_GIT_REF="a7a927731b109fdfec770df7c2a5be5d7e2216b5"
 WHEEL_DIR=""
 
 # Parse command line arguments