[Build] Upgrade runtime image to Ubuntu 24.04, pick up rdma-core > 44 (#51058)
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com> Co-authored-by: OpenAI Codex <codex@openai.com> Co-authored-by: Michael Goin <mgoin64@gmail.com>
Tyler Michael Smith committed
Aug 8, 2026 at 21:17 UTC
cb1a52aee3b31c692a6629ebd7ff5503b3cd13f7
3 files changed
+7
-5
docker/Dockerfile
+2
-2
@@ -24,7 +24,7 @@
24
25
ARG CUDA_VERSION=13.0.3
26
ARG PYTHON_VERSION=3.12
27
-ARG UBUNTU_VERSION=22.04
27
+ARG UBUNTU_VERSION=24.04
28
# DeepEPv2 requires NCCL >= 2.30.4 (GIN backend).
29
# This version is used for CUDA 12+ builds to keep the DeepEP build and
30
# runtime NCCL versions compatible.
@@ -765,7 +765,7 @@ RUN --mount=type=cache,target=/opt/uv/cache \
765
# Install gdrcopy (saves ~6s per build)
766
# TODO (huydhn): There is no prebuilt gdrcopy package on 12.9 at the moment
767
ARG GDRCOPY_CUDA_VERSION=12.8
768
-ARG GDRCOPY_OS_VERSION=Ubuntu22_04
768
+ARG GDRCOPY_OS_VERSION=Ubuntu24_04
769
ARG TARGETPLATFORM
770
COPY tools/install_gdrcopy.sh /tmp/install_gdrcopy.sh
771
RUN set -eux; \
docker/versions.json
+3
-3
@@ -8,7 +8,7 @@
8
"default": "3.12"
9
},
10
"UBUNTU_VERSION": {
11
- "default": "22.04"
11
+ "default": "24.04"
12
},
13
"NCCL_VERSION": {
14
"default": "2.30.7"
@@ -17,7 +17,7 @@
17
"default": "pytorch/manylinux2_28-builder:cuda13.0"
18
},
19
"FINAL_BASE_IMAGE": {
20
- "default": "nvidia/cuda:13.0.3-base-ubuntu22.04"
20
+ "default": "nvidia/cuda:13.0.3-base-ubuntu24.04"
21
},
22
"GET_PIP_URL": {
23
"default": "https://bootstrap.pypa.io/get-pip.py"
@@ -74,7 +74,7 @@
74
"default": "12.8"
75
},
76
"GDRCOPY_OS_VERSION": {
77
- "default": "Ubuntu22_04"
77
+ "default": "Ubuntu24_04"
78
},
79
"BITSANDBYTES_VERSION_X86": {
80
"default": "0.46.1"
docs/getting_started/installation/gpu.cuda.inc.md
+2
@@ -321,6 +321,8 @@ You can add any other [engine-args](https://docs.vllm.ai/en/latest/configuration
321
322
vLLM's Docker image comes with [CUDA compatibility libraries](https://docs.nvidia.com/deploy/cuda-compatibility/index.html) pre-installed. This allows you to run vLLM on systems with NVIDIA drivers that are older than the CUDA Toolkit version used in the image, but only supports select professional and datacenter NVIDIA GPUs.
323
324
+For CUDA 13 images, the minimum host kernel is Linux 4.15 when running normally because CUDA 13 requires an R580 or newer driver. Compatibility mode supports R535 and R570 host drivers; R535 lowers the minimum host kernel to Linux 3.10, while R570 still requires Linux 4.15. Upgrading the container userland to Ubuntu 24.04 does not raise these driver-defined kernel requirements. See NVIDIA's [forward compatibility matrix](https://docs.nvidia.com/deploy/cuda-compatibility/forward-compatibility.html#use-the-right-cuda-forward-compatibility-package), [R535 requirements](https://download.nvidia.com/XFree86/Linux-x86_64/535.104.05/README/minimumrequirements.html), and [R580 requirements](https://download.nvidia.com/XFree86/Linux-x86_64/580.76.05/README/minimumrequirements.html).
325
+
326
To enable this feature, set the `VLLM_ENABLE_CUDA_COMPATIBILITY` environment variable to `1` or `true` when running the container:
327
328
```bash