vLLM 0.11.0 CUDA Library Mismatch on ARM64 with CUDA 13.x

### Your current environment

Collecting environment information...
==============================
        System Info
==============================
OS                           : Ubuntu 24.04.3 LTS (aarch64)
GCC version                  : (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version                : Could not collect
CMake version                : version 3.31.2
Libc version                 : glibc-2.39

==============================
       PyTorch Info
==============================
PyTorch version              : 2.8.0+cpu
Is debug build               : False
CUDA used to build PyTorch   : Could not collect
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.13.5 | packaged by Anaconda, Inc. | (main, Jun 12 2025, 16:17:47) [GCC 11.2.0] (64-bit runtime)
Python platform              : Linux-6.11.0-1016-nvidia-aarch64-with-glibc2.39

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : False
CUDA runtime version         : 13.0.88
CUDA_MODULE_LOADING set to   : N/A
GPU models and configuration : GPU 0: NVIDIA GB10
Nvidia driver version        : 580.95.05
cuDNN version                : Could not collect
HIP runtime version          : N/A
MIOpen runtime version       : N/A
Is XNNPACK available         : True

==============================
          CPU Info
==============================
Architecture:                         aarch64
CPU op-mode(s):                       64-bit
Byte Order:                           Little Endian
CPU(s):                               20
On-line CPU(s) list:                  0-19
Vendor ID:                            ARM
Model name:                           Cortex-X925
Model:                                1
Thread(s) per core:                   1
Core(s) per socket:                   10
Socket(s):                            1
Stepping:                             r0p1
CPU(s) scaling MHz:                   110%
CPU max MHz:                          4004.0000
CPU min MHz:                          1378.0000
BogoMIPS:                             2000.00
Flags:                                fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 sm3 sm4 asimddp sha512 sve asimdfhm dit uscat ilrcpc flagm sb paca pacg dcpodp sve2 sveaes svepmull svebitperm svesha3 svesm4 flagm2 frint svei8mm svebf16 i8mm bf16 dgh bti ecv afp wfxt
Model name:                           Cortex-A725
Model:                                1
Thread(s) per core:                   1
Core(s) per socket:                   10
Socket(s):                            1
Stepping:                             r0p1
CPU(s) scaling MHz:                   130%
CPU max MHz:                          2860.0000
CPU min MHz:                          338.0000
BogoMIPS:                             2000.00
Flags:                                fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 sm3 sm4 asimddp sha512 sve asimdfhm dit uscat ilrcpc flagm sb paca pacg dcpodp sve2 sveaes svepmull svebitperm svesha3 svesm4 flagm2 frint svei8mm svebf16 i8mm bf16 dgh bti ecv afp wfxt
L1d cache:                            1.3 MiB (20 instances)
L1i cache:                            1.3 MiB (20 instances)
L2 cache:                             25 MiB (20 instances)
L3 cache:                             24 MiB (2 instances)
NUMA node(s):                         1
NUMA node0 CPU(s):                    0-19
Vulnerability Gather data sampling:   Not affected
Vulnerability Itlb multihit:          Not affected
Vulnerability L1tf:                   Not affected
Vulnerability Mds:                    Not affected
Vulnerability Meltdown:               Not affected
Vulnerability Mmio stale data:        Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed:               Not affected
Vulnerability Spec rstack overflow:   Not affected
Vulnerability Spec store bypass:      Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:             Mitigation; __user pointer sanitization
Vulnerability Spectre v2:             Not affected
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

==============================
Versions of relevant libraries
==============================
[pip3] numpy==2.2.6
[pip3] nvidia-cublas==13.0.0.19
[pip3] nvidia-cuda-cupti==13.0.48
[pip3] nvidia-cuda-nvrtc==13.0.48
[pip3] nvidia-cuda-runtime==13.0.48
[pip3] nvidia-cudnn-cu13==9.13.0.50
[pip3] nvidia-cufft==12.0.0.15
[pip3] nvidia-cufile==1.15.0.42
[pip3] nvidia-curand==10.4.0.35
[pip3] nvidia-cusolver==12.0.3.29
[pip3] nvidia-cusparse==12.6.2.49
[pip3] nvidia-cusparselt-cu13==0.8.0
[pip3] nvidia-nccl-cu13==2.27.7
[pip3] nvidia-nvjitlink==13.0.39
[pip3] nvidia-nvshmem-cu13==3.3.24
[pip3] nvidia-nvtx==13.0.39
[pip3] pyzmq==27.0.2
[pip3] segmentation_models_pytorch==0.5.0
[pip3] torch==2.8.0
[pip3] torchaudio==2.8.0
[pip3] torchvision==0.23.0
[pip3] transformers==4.55.2
[pip3] triton==3.5.0
[conda] numpy                                2.2.6            pypi_0              pypi
[conda] nvidia-cublas                        13.0.0.19        pypi_0              pypi
[conda] nvidia-cuda-cupti                    13.0.48          pypi_0              pypi
[conda] nvidia-cuda-nvrtc                    13.0.48          pypi_0              pypi
[conda] nvidia-cuda-runtime                  13.0.48          pypi_0              pypi
[conda] nvidia-cudnn-cu13                    9.13.0.50        pypi_0              pypi
[conda] nvidia-cufft                         12.0.0.15        pypi_0              pypi
[conda] nvidia-cufile                        1.15.0.42        pypi_0              pypi
[conda] nvidia-curand                        10.4.0.35        pypi_0              pypi
[conda] nvidia-cusolver                      12.0.3.29        pypi_0              pypi
[conda] nvidia-cusparse                      12.6.2.49        pypi_0              pypi
[conda] nvidia-cusparselt-cu13               0.8.0            pypi_0              pypi
[conda] nvidia-nccl-cu13                     2.27.7           pypi_0              pypi
[conda] nvidia-nvjitlink                     13.0.39          pypi_0              pypi
[conda] nvidia-nvshmem-cu13                  3.3.24           pypi_0              pypi
[conda] nvidia-nvtx                          13.0.39          pypi_0              pypi
[conda] pyzmq                                27.0.2           pypi_0              pypi
[conda] segmentation-models-pytorch          0.5.0            pypi_0              pypi
[conda] torch                                2.8.0            pypi_0              pypi
[conda] torchaudio                           2.8.0            pypi_0              pypi
[conda] torchvision                          0.23.0           pypi_0              pypi
[conda] transformers                         4.55.2           pypi_0              pypi
[conda] triton                               3.5.0            pypi_0              pypi

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
vLLM Version                 : 0.11.0
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
        GPU0    NIC0    NIC1    NIC2    NIC3    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      NODE    NODE    NODE    NODE    0-19    0               N/A
NIC0    NODE     X      PIX     NODE    NODE
NIC1    NODE    PIX      X      NODE    NODE
NIC2    NODE    NODE    NODE     X      PIX
NIC3    NODE    NODE    NODE    PIX      X

Legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing at most a single PCIe bridge
  NV#  = Connection traversing a bonded set of # NVLinks

NIC Legend:

  NIC0: rocep1s0f0
  NIC1: rocep1s0f1
  NIC2: roceP2p1s0f0
  NIC3: roceP2p1s0f1

==============================
     Environment Variables
==============================
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1


### How you are installing vllm

# vLLM CUDA Library Compatibility Issue - ARM64 NVIDIA GB10

## Problem Summary
vLLM 0.11.0 fails to start on ARM64 NVIDIA GB10 GPU due to missing CUDA libraries. The system has CUDA 13.x installed but vLLM is compiled against CUDA 12.x, causing import errors.

## Environment Details

### System Information
- **Architecture**: ARM64 (aarch64) - Graviton-based
- **GPU**: NVIDIA GB10
- **CUDA Version**: 13.x (libcudart.so.13 available, not libcudart.so.12)
- **OS**: Linux

### Python & Package Versions
```
Python: 3.12.3
vLLM: 0.11.0
PyTorch: 2.8.0
torchvision: 0.23.0
torchaudio: 2.8.0
transformers: 4.57.1
CUDA Capability: 12.1
```

### Installed via
```bash
/usr/bin/python3.12 -m venv vllm_env
source vllm_env/bin/activate
pip install vllm
```

## Error Messages

### Error 1: libcudart.so.12 not found
```
ImportError: libcudart.so.12: cannot open shared object file: No such file or directory
```

Traceback:
```
File "/home/swatson/work/MachineLearning/vllm_env/lib/python3.12/site-packages/vllm/platforms/cuda.py", line 18, in <module>
    import vllm._C  # noqa
    ^^^^^^^^^^^^^^
ImportError: libcudart.so.12: cannot open shared object file: No such file or directory
```

### Error 2: libtorch_cuda.so not found
```
ImportError: libtorch_cuda.so: cannot open shared object file: No such file or directory
```

## Reproduction Steps

1. **Create clean environment with Python 3.12**:
```bash
/usr/bin/python3.12 -m venv vllm_env
source vllm_env/bin/activate
pip install --upgrade pip
```

2. **Install vLLM**:
```bash
pip install vllm==0.11.0
```

3. **Test import**:
```bash
python -c "import vllm; print('OK')"
```

4. **Try to run server**:
```bash
export LD_LIBRARY_PATH=/usr/local/cuda/targets/sbsa-linux/lib:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
python -m vllm.entrypoints.openai.api_server --model moonshotai/Kimi-Linear-48B-A3B-Instruct --port 8000
```

**Result**: `ImportError: libcudart.so.12: cannot open shared object file`

## Root Cause Analysis

1. **vLLM compiled for CUDA 12.x**: The vllm._C extension module is compiled against CUDA Toolkit 12.x
2. **System has CUDA 13.x**: The deployment environment has CUDA 13.x installed with libcudart.so.13
3. **CUDA Forward Compatibility**: CUDA 13.x is not backward compatible with CUDA 12.x libraries

### Available System Libraries
```bash
$ ldconfig -p | grep libcudart
libcudart.so.13 (libc6,AArch64) => /usr/local/cuda/targets/sbsa-linux/lib/libcudart.so.13
libcudart.so (libc6,AArch64) => /usr/local/cuda/targets/sbsa-linux/lib/libcudart.so
```

## Attempted Workarounds

### 1. LD_LIBRARY_PATH Modification
```bash
export LD_LIBRARY_PATH=/usr/local/cuda/targets/sbsa-linux/lib:$LD_LIBRARY_PATH
```
**Result**: ❌ Does not work (CUDA version mismatch, not search path issue)

### 2. Creating Library Symlinks
Cannot symlink libcudart.so.13 → libcudart.so.12 (version incompatibility)

### 3. Trying Different vLLM Versions
- vLLM 0.6.3: Requires torch==2.4.0 (unavailable, newer versions: 2.6.0+)
- vLLM 0.11.0: Requires CUDA 12.x (system has 13.x)
- vLLM 0.12.0+: Not available on PyPI (unclear release status)

## Expected Behavior

vLLM should either:

1. **Support CUDA 13.x** in wheel distributions
2. **Build wheels for multiple CUDA versions** (cu118, cu121, cu124)
3. **Detect system CUDA version** and download compatible wheels
4. **Provide clear error messages** indicating required CUDA version

## Possible Solutions

### Option A: Distribute CUDA 13.x wheels
Build and distribute vllm-0.11.0 wheels compiled for CUDA 13.x

### Option B: CUDA version detection
Auto-detect system CUDA and download matching wheel (like `torch` does)

### Option C: Documentation
Clearly document CUDA version requirements in README and installation guide

### Option D: Conda packages
Provide conda packages with proper CUDA dependency resolution

## Impact

- vLLM cannot run on ARM64 systems with CUDA 13.x
- Affects AWS Graviton instances and other ARM64 deployments
- Blocks users trying to run Kimi-Linear-48B or other large models on newer CUDA systems

## References

- **System CUDA**: /usr/local/cuda/ (version 13.x)
- **PyTorch CUDA Support**: Works fine with torch 2.8.0
- **GPU Query**: 
  ```
  $ nvidia-smi
  NVIDIA GB10, CUDA 12.1 capability
  ```

## Related Issues

This may be related to:
- Platform compatibility across CUDA versions
- ARM64 architecture support
- Pre-built wheel distribution strategy

## Requested Action

Please:
1. ✅ Acknowledge ARM64 + CUDA 13.x as supported configuration
2. ✅ Provide CUDA 13.x compatible wheels
3. ✅ Update documentation with CUDA version requirements
4. ✅ Consider automatic CUDA version detection (like PyTorch)

---

**Additional Notes:**
- This is blocking deployment of vLLM on modern ARM64 GPU instances
- PyTorch 2.8.0 works fine on the same system, showing CUDA 13.x support is possible
- Other frameworks (transformers, llama.cpp) handle this gracefully with multi-version wheels


### Before submitting a new issue...

- [x] Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.

Uh oh!

vLLM 0.11.0 CUDA Library Mismatch on ARM64 with CUDA 13.x #28669

Description

Your current environment

Collecting environment information...

============================== PyTorch Info

============================== Python Environment

============================== CUDA / GPU Info

============================== CPU Info

============================== Versions of relevant libraries

============================== vLLM Info

============================== Environment Variables

How you are installing vllm

vLLM CUDA Library Compatibility Issue - ARM64 NVIDIA GB10

Problem Summary

Environment Details

System Information

Python & Package Versions

Installed via

Error Messages

Error 1: libcudart.so.12 not found

Error 2: libtorch_cuda.so not found

Reproduction Steps

Root Cause Analysis

Available System Libraries

Attempted Workarounds

1. LD_LIBRARY_PATH Modification

2. Creating Library Symlinks

3. Trying Different vLLM Versions

Expected Behavior

Possible Solutions

Option A: Distribute CUDA 13.x wheels

Option B: CUDA version detection

Option C: Documentation

Option D: Conda packages

Impact

References

Related Issues

Requested Action

Before submitting a new issue...

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

==============================
PyTorch Info

==============================
Python Environment

==============================
CUDA / GPU Info

==============================
CPU Info

==============================
Versions of relevant libraries

==============================
vLLM Info

==============================
Environment Variables