[Kernel][Build] Bump DeepGEMM to the 26/09/30 release and adapt the sparse indexer (#61149)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

tools/build_deepgemm_C.py

@@ -35,8 +35,9 @@ env = os.environ.copy()
 env.pop("DG_SKIP_CUDA_BUILD", None)
 
 print(f"[build_deepgemm_C] building in {src} with {sys.executable}", flush=True)
+# CMake decides when to rebuild; setuptools would only check python_api.cpp.
 subprocess.check_call(
-    [sys.executable, "setup.py", "build_ext", "--inplace"],
+    [sys.executable, "setup.py", "build_ext", "--inplace", "--force"],
     cwd=src,
     env=env,
 )

tools/check_wheel_deepgemm.py

@@ -4,10 +4,12 @@
 """Assert the vendored DeepGEMM package has the TORCH_LIBRARY binding layout.
 
 Expects ``deep_gemm/_C.py`` plus a single ``_C_extension*.so`` (abi3) under
-``vllm.third_party.deep_gemm``. Run after vLLM is installed, e.g. the H100
-deepgemm kernel tests in .buildkite/test_areas/kernels.yaml.
+``vllm.third_party.deep_gemm``, and that the package imports: vLLM treats an
+import failure as DeepGEMM being absent. Run after vLLM is installed, e.g. the
+H100 deepgemm kernel tests in .buildkite/test_areas/kernels.yaml.
 """
 
+import importlib
 import importlib.util
 import sys
 from pathlib import Path
@@ -38,6 +40,11 @@ def main() -> int:
     if missing:
         print(f"missing: {missing}", file=sys.stderr)
         return 1
+    try:
+        importlib.import_module("vllm.third_party.deep_gemm")
+    except Exception as e:
+        print(f"vllm.third_party.deep_gemm failed to import: {e!r}", file=sys.stderr)
+        return 1
     return 0
 
 

tools/install_deepgemm.sh

@@ -7,10 +7,10 @@ set -e
 # Default values
 # Keep DEEPGEMM_GIT_REF in sync with cmake/external_projects/deepgemm.cmake
 DEEPGEMM_GIT_REPO="https://github.com/vllm-project/DeepGEMM.git"
-# NOTE: This targets the vLLM fork's dev branch, which carries the sm120
-# and sm90 paged-MQA ports plus the SwiGLU alpha/beta and SiTU Mega MoE
-# activations, plus the CUDA 12.x layout header fix from vllm-project/DeepGEMM#12.
-DEEPGEMM_GIT_REF="1e1842a833699298f7afc02eefb2ed168fff6938"
+# 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"
 WHEEL_DIR=""
 
 # Parse command line arguments