[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>
This commit is contained in:
Yongye Zhu
2026-10-12 23:24:43 -07:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 86b1c2ef15
commit a6009ea85b
23 changed files with 516 additions and 208 deletions
+9 -2
View File
@@ -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