A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[ci/build] Fix abi tag for aarch64 (#23329)
Signed-off-by: youkaichao <youkaichao@gmail.com>
youkaichao committed
Aug 21, 2025 at 23:32 UTC
e0b056e443135e98a525bcfd2f59d87f4a7bcc14
3 files changed
+32
-7
.buildkite/generate_index.py
+7
-2
@@ -22,11 +22,16 @@ filename = os.path.basename(args.wheel)
22
23
with open("index.html", "w") as f:
24
print(f"Generated index.html for {args.wheel}")
25
+ # sync the abi tag with .buildkite/scripts/upload-wheels.sh
26
if "x86_64" in filename:
27
x86_wheel = filename
27
- arm_wheel = filename.replace("x86_64", "aarch64")
28
+ arm_wheel = filename.replace("x86_64", "aarch64").replace(
29
+ "manylinux1", "manylinux2014"
30
+ )
31
elif "aarch64" in filename:
29
- x86_wheel = filename.replace("aarch64", "x86_64")
32
+ x86_wheel = filename.replace("aarch64", "x86_64").replace(
33
+ "manylinux2014", "manylinux1"
34
+ )
35
arm_wheel = filename
36
else:
37
raise ValueError(f"Unsupported wheel: {filename}")
.buildkite/scripts/upload-wheels.sh
+13
-2
@@ -14,8 +14,19 @@ fi
14
# Get the single wheel file
15
wheel="${wheel_files[0]}"
16
17
-# Rename 'linux' to 'manylinux1' in the wheel filename
18
-new_wheel="${wheel/linux/manylinux1}"
17
+# Detect architecture and rename 'linux' to appropriate manylinux version
18
+arch=$(uname -m)
19
+if [[ $arch == "x86_64" ]]; then
20
+ manylinux_version="manylinux1"
21
+elif [[ $arch == "aarch64" ]]; then
22
+ manylinux_version="manylinux2014"
23
+else
24
+ echo "Warning: Unknown architecture $arch, using manylinux1 as default"
25
+ manylinux_version="manylinux1"
26
+fi
27
+
28
+# Rename 'linux' to the appropriate manylinux version in the wheel filename
29
+new_wheel="${wheel/linux/$manylinux_version}"
30
mv -- "$wheel" "$new_wheel"
31
wheel="$new_wheel"
32
setup.py
+12
-3
@@ -643,16 +643,25 @@ if envs.VLLM_USE_PRECOMPILED:
643
if wheel_location is not None:
644
wheel_url = wheel_location
645
else:
646
+ import platform
647
+ arch = platform.machine()
648
+ if arch == "x86_64":
649
+ wheel_tag = "manylinux1_x86_64"
650
+ elif arch == "aarch64":
651
+ wheel_tag = "manylinux2014_aarch64"
652
+ else:
653
+ raise ValueError(f"Unsupported architecture: {arch}")
654
base_commit = precompiled_wheel_utils.get_base_commit_in_main_branch()
647
- wheel_url = f"https://wheels.vllm.ai/{base_commit}/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl"
655
+ wheel_url = f"https://wheels.vllm.ai/{base_commit}/vllm-1.0.0.dev-cp38-abi3-{wheel_tag}.whl"
656
+ nightly_wheel_url = f"https://wheels.vllm.ai/nightly/vllm-1.0.0.dev-cp38-abi3-{wheel_tag}.whl"
657
from urllib.request import urlopen
658
try:
659
with urlopen(wheel_url) as resp:
660
if resp.status != 200:
652
- wheel_url = "https://wheels.vllm.ai/nightly/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl"
661
+ wheel_url = nightly_wheel_url
662
except Exception as e:
663
print(f"[warn] Falling back to nightly wheel: {e}")
655
- wheel_url = "https://wheels.vllm.ai/nightly/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl"
664
+ wheel_url = nightly_wheel_url
665
666
patch = precompiled_wheel_utils.extract_precompiled_and_patch_package(
667
wheel_url)