vLLM cu130 aarch64 build — df09c767
Build
| wheel | vllm-0.28.1rc1.dev287+cu130.20260902.gdf09c767-cp38-abi3-manylinux_2_28_aarch64.whl |
|---|
| variant | cu130 (CUDA 13.0.0, torch 2.13.0+cu130) |
|---|
| vLLM commit | df09c767355fd1313001fbf14ad363bba5bf0ffd |
|---|
| build date | 20260902 |
|---|
| build time | not recorded |
|---|
| ninja steps | 466 |
|---|
| wheel size | 508.0 MB |
|---|
| peak memory | ? GiB cgroup / ? GiB anon |
|---|
| builder | [email protected] (aarch64, 20x2 jobs, cpuset 0-19) |
|---|
| inspected object | vllm/_C_stable_libtorch.abi3.so |
|---|
| ELF | ELF 64-bit LSB shared object, ARM aarch64 |
|---|
| cubins | 100, 120, 80, 89, 90, 90a |
|---|
| max glibc | 2.17 |
|---|
| max libstdc++ | 3.4.21 |
|---|
GPU coverage
Compiled cubins, no JIT needed. Verified present:
100, 120, 80, 89, 90, 90a
| GPU | requested arch | cubin |
|---|
| GH200 / H200 (Grace Hopper) | 9.0a | sm_90a |
| GB200 + GB300 (Blackwell family) | 10.0f | sm_100 |
| GB10 (DGX Spark) | 12.0f | sm_120 |
Validation
PASSED
- wheel contains
vllm/_C*.so
file reports ELF 64-bit LSB shared object, ARM aarch64cuobjdump --list-elf shows a cubin for every GPU above- needs at most glibc
2.17 /
GLIBCXX 3.4.21
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)
real 127m52.265s
user 1284m25.975s
sys 50m5.475s
BUILD_OK /home/build/vllm-nightly/dist/cu130/vllm-0.28.1rc1.dev287+gdf09c7673.d20260902-cp312-cp312-linux_aarch64.whl
Warning: Permanently added '[192.168.1.141]:30422' (ED25519) to the list of known hosts.
VALIDATION {
"wheel": "vllm-0.28.1rc1.dev287+gdf09c7673.d20260902-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",
"120",
"80",
"89",
"90",
"90a"
],
"max_glibc": "2.17",
"max_glibcxx": "3.4.21"
}
}