vLLM cu129 aarch64 build — 21a22110

generated 2026-09-05 11:47 +07 · wheel directory · how to install · upstream commit

Build

wheelvllm-0.28.1rc1.dev364+cu129.20260904.g21a22110-cp38-abi3-manylinux_2_28_aarch64.whl
variantcu129 (CUDA 12.9.1, torch 2.13.0+cu129)
vLLM commit21a22110404d98dcab5f51cd90e9acb0acc15e47
build date20260904
build timenot recorded
ninja steps463
wheel size697.1 MB
peak memory78.7 GiB cgroup / 73.8 GiB anon
builder[email protected] (aarch64, 20x2 jobs, cpuset 0-19)
inspected objectvllm/_C_stable_libtorch.abi3.so
ELFELF 64-bit LSB shared object, ARM aarch64
cubins100, 100a, 103, 103a, 121, 121a, 80, 89, 90, 90a
max glibc2.17
max libstdc++3.4.21

GPU coverage

Compiled cubins, no JIT needed. Verified present: 100, 100a, 103, 103a, 121, 121a, 80, 89, 90, 90a
GPUrequested archcubin
GH200 / H200 (Grace Hopper)9.0asm_90a
GB200 (Blackwell)10.0asm_100a
GB300 (Blackwell Ultra)10.3asm_103a
GB10 (DGX Spark)12.1asm_121a

Validation

PASSED

Structural checks only — neither machine has an aarch64 GPU. Functional proof comes from validate-gb10.py, which rents an NVIDIA GB10 and runs a real generation.

Build log (tail)

BUILD_OK /home/build/vllm-nightly/dist/cu129/vllm-0.28.1rc1.dev364+g21a221104.d20260903.cu129-cp38-abi3-linux_aarch64.whl
Warning: Permanently added '[192.168.1.141]:30422' (ED25519) to the list of known hosts.

VALIDATION {
  "wheel": "vllm-0.28.1rc1.dev364+g21a221104.d20260903.cu129-cp38-abi3-linux_aarch64.whl",
  "checks": {
    "shared_objects": 13,
    "inspected": "vllm/_C_stable_libtorch.abi3.so",
    "file": "ELF 64-bit LSB shared object, ARM aarch64",
    "cubin_archs": [
      "100",
      "100a",
      "103",
      "103a",
      "121",
      "121a",
      "80",
      "89",
      "90",
      "90a"
    ],
    "max_glibc": "2.17",
    "max_glibcxx": "3.4.21"
  }
}