Skip to content

[BUG]: Error 803: CUDA driver/display driver mismatch when running dynamo serve with vLLM #1101

Description

@lauragonzalezmoran

Describe the Bug

When attempting to launch dynamo serve using the vLLM backend on a machine with 2×H100 GPUs, CUDA 12.8, and NVIDIA driver version 570.133.20, the system fails at runtime with the following error:

RuntimeError: Unexpected error from cudaGetDeviceCount(). Did you run some cuda functions before calling NumCudaDevices() that might have already set an error? Error 803: system has unsupported display driver / cuda driver combination

The error occurs during GPU initialization by the VllmWorker component, even though nvidia-smi confirms that CUDA and the H100 GPUs are visible both on the host and inside the container. As a result, the model engine fails to start and inference cannot proceed.

Additionally, the Planner component later fails due to a missing dynamo.json state file under /root/.dynamo/state.

This seems to be a runtime compatibility issue specific to this driver/CUDA combination or Docker setup.

Steps to Reproduce

1. Create a compatible Python environment

The python3 -m venv method installs Python 3.13.2 by default on Ubuntu 24.04, which is not supported by ai-dynamo-runtime. To avoid dependency resolution issues, I used Conda to create a Python 3.12 environment, which is supported by ai-dynamo-runtime:

conda create --name dynamo-test2 python=3.12
conda activate dynamo-test2

2. Install the ai-dynamo[all] package

pip install "ai-dynamo[all]"

Running a model

3. Fix shared library loading issue

When attempting to run a model with dynamo run (more specifically : dynamo run out=vllm deepseek-ai/DeepSeek-R1-Distill-Llama-8B) I initially hit a libpython3.12.so.1.0 error. To resolve it, I updated LD_LIBRARY_PATH:
export LD_LIBRARY_PATH="$CONDA_PREFIX/lib:$LD_LIBRARY_PATH"
Then I confirmed model loading works with:
dynamo run out=vllm deepseek-ai/DeepSeek-R1-Distill-Llama-8B. I was able to chat with the model.

Serving a model

I then wanted to serve the model specifically using --framework vllm so I followed the official instructions in https://github.com/ai-dynamo/dynamo/blob/main/examples/llm/README.md

4. Start infrastructure services

Note: You must clone the repo first (git clone https://github.com/ai-dynamo/dynamo.git) to access the Docker configs.

cd dynamo
docker compose -f deploy/metrics/docker-compose.yml up -d

5. Launch the containerized vLLM runtime

According to the docs, the next step is to run the container:
./container/run.sh -it --framework vllm

Immediately upon container launch, I see this error in the logs:
ERROR: The NVIDIA Driver is present, but CUDA failed to initialize. GPU functionality will not be available. [[ System has unsupported display driver / cuda driver combination (error 803) ]]
Despite the error, I'm dropped into a shell at /workspace.

6. Attempt to serve the model

Navigate to the LLM example and run:
cd $DYNAMO_HOME/examples/llm dynamo serve graphs.agg:Frontend -f ./configs/agg.yaml

This fails with the following critical error during cudaGetDeviceCount(): RuntimeError: Unexpected error from cudaGetDeviceCount(). Did you run some cuda functions before calling NumCudaDevices() that might have already set an error? Error 803: system has unsupported display driver / cuda driver combination.

I will attach the full output when running dynamo serve graphs.agg:Frontend -f ./configs/agg.yaml to the in Actual Behavior Section in case it is useful:

7. Confirm GPUs are visible

I ran nvidia-smi inside the container immediately after the failure:
nvidia-smi
✅ Both H100 GPUs are detected and idle.
✅ CUDA 12.8 and Driver 570.133.20 appear to be installed correctly.

Expected Behavior

Expected Behavior

I expected the model to be successfully served using dynamo serve graphs.agg:Frontend -f ./configs/agg.yaml, with vLLM initializing the GPU backend properly and no CUDA-related errors. Specifically:

  • The VllmWorker component should initialize without failure.
  • CUDA should be recognized and initialized correctly inside the container.
  • Inference requests should be accepted at the exposed endpoint (localhost:8000/chat/completions).

Actual Behavior

Actual Behavior

When attempting to serve the model using dynamo serve graphs.agg:Frontend -f ./configs/agg.yaml, the following occurred:

  • The container successfully started, and all components (Frontend, Planner, Processor, VllmWorker) began initialization.

  • However, during VllmWorker initialization, the following error occurred:

    RuntimeError: Unexpected error from cudaGetDeviceCount(): Error 803: system has unsupported display driver / cuda driver combination

  • This caused the engine process to fail, and consequently:

    • The VllmWorker crashed.
    • The planner later also failed due to a missing state file:
      FileNotFoundError: State file not found: /root/.dynamo/state/dynamo.json
  • Notably, nvidia-smi inside the container does detect the H100 GPUs correctly and shows no running processes, so CUDA visibility itself appears to be working — but initialization fails.

    ERROR: The NVIDIA Driver is present, but CUDA failed to initialize.
    [[ System has unsupported display driver / cuda driver combination (error 803) ]]
    

This blocks any ability to serve inference workloads using vLLM through Dynamo.

I have attached a file with the entire output of dynamo serve graphs.agg:Frontend -f ./configs/agg.yaml here.

root@_MY_MACHINE___workspace_examples_llm# dynamo serve graphs.txt

Environment

From nvidia-smi

  • GPU Model: 2 × NVIDIA H100 NVL

  • GPU Memory: 95,830 MiB each (0 MiB in use at error time)

  • Driver Version: 570.133.20

  • CUDA Version (reported by nvidia-smi): 12.8

From dynamo env:

System Information:
OS: Ubuntu 24.04
Glibc Version: 2.39
GCC Version: 13.3.0
Cmake Version: 3.28.3
Rust Version: 1.86.0
Docker Version: N/A

CPU Information:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 52 bits physical, 57 bits virtual
Byte Order: Little Endian
CPU(s): 128
On-line CPU(s) list: 0-127
Vendor ID: AuthenticAMD
Model name: AMD EPYC 9334 32-Core Processor
CPU family: 25
Model: 17
Thread(s) per core: 2
Core(s) per socket: 32
Socket(s): 2
Stepping: 1
Frequency boost: enabled
CPU(s) scaling MHz: 60%
CPU max MHz: 3910.2529
CPU min MHz: 1500.0000
BogoMIPS: 5391.93
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp ibrs_enhanced vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor smca fsrm flush_l1d
Virtualization: AMD-V
L1d cache: 2 MiB (64 instances)
L1i cache: 2 MiB (64 instances)
L2 cache: 64 MiB (64 instances)
L3 cache: 256 MiB (8 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-31,64-95
NUMA node1 CPU(s): 32-63,96-127
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: Mitigation; safe RET
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected

Python Version: 3.12.3
Python Platform: Linux-5.15.0-134-generic-x86_64-with-glibc2.39

Python Packages:
ai-dynamo: 0.2.1
ai-dynamo-runtime: 0.2.1
ai-dynamo-vllm: 0.8.4.post1
genai-perf: 0.0.12
nixl: 0.2.1
numpy: 1.26.4
nvidia-cublas-cu12: 12.4.5.8
nvidia-cuda-cupti-cu12: 12.4.127
nvidia-cuda-nvrtc-cu12: 12.4.127
nvidia-cuda-runtime-cu12: 12.4.127
nvidia-cudnn-cu12: 9.1.0.70
nvidia-cufft-cu12: 11.2.1.3
nvidia-curand-cu12: 10.3.5.147
nvidia-cusolver-cu12: 11.6.1.9
nvidia-cusparse-cu12: 12.3.1.170
nvidia-ml-py: 12.575.51
nvidia-nccl-cu12: 2.21.5
nvidia-nvjitlink-cu12: 12.4.127
nvidia-nvtx-cu12: 12.4.127
pyzmq: 26.4.0
tensorrt_llm: Not installed
torch: 2.6.0
torchaudio: 2.6.0
transformers: 4.51.3
tritonclient: Not installed

Additional Context

No response

Screenshots

No response

Metadata

Metadata

Assignees

Labels

bugSomething isn't working

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions