I run in Kubernetes with the following setup vllm serve openai/gpt-oss-120b --trust-remote-code --tool-call-parser openai –enable-auto-tool-choice --max-model-len 60000 --gpu-memory-utilization 0.95 –tensor-parallel-size=2. Everything works when I set –tensor-parallel-size=1 or rollback to v0.10.2 …

Your setup is correct for running vLLM with tensor parallelism, but issues with --tensor-parallel-size >1 are often due to Kubernetes resource allocation, NCCL, or container configuration. Ensure that your pod requests and limits specify at least 2 GPUs, and that /dev/shm is mounted with sufficient …

@RunLLM But why it works on v0.10.2 then?

If it works on v0.10.2 but not newer versions, it’s likely due to changes in how vLLM handles tensor parallelism, NCCL, or distributed initialization. Recent updates may have introduced stricter checks, new defaults, or compatibility issues with Kubernetes GPU allocation or NCCL setup, which were mo…

I have the same issue here. Switching back to 0.10.2 did fix the the problem as OP mentioned. The issue happens with both the 0.11.0 container as well as the 0.11.0 pip installable package. Please fix the issue

RuntimeError: CUDA driver error: invalid device ordinal after the update to v0.11.0

General

TarunAitha October 13, 2025, 8:45am 5

I too observed same thing. Can we expect fix in the next VLLM release ?

1 Like

Topic		Replies	Views
Help for error when run vllm with tensor parallel General	7	1741	May 21, 2025
Added second 5090 and turne on tensor parallel 2 General	9	360	September 18, 2025
vLLM does not work with 2x 5090 in tp 2 General	8	739	September 18, 2025
NCCL error across 2 machines 2x4GPUs need advice General	2	394	November 12, 2025
求救各位大佬看看是什么问题。cuda12.9，pytorch2.8，vllm0.11.0 General	3	186	November 14, 2025

RuntimeError: CUDA driver error: invalid device ordinal after the update to v0.11.0

Related topics