KeyError: invalid tool call parser 'openai' when using --tool-call-parser openai

[wanghuacan]

Description:
When I try to launch vllm serve with the --tool-call-parser openai option, the server crashes with a KeyError.

Steps to Reproduce:
Run the following command:

vllm serve openai/gpt-oss-120b  \
  --tensor_parallel_size 8 \
  --max-model-len 131072 \
  --max-num-batched-tokens 10240 \
  --max-num-seqs 128 \
  --tool-call-parser openai \
  --enable-auto-tool-choice \
  --gpu-memory-utilization 0.85 \
  --no-enable-prefix-caching

Actual Behavior:
The server raises the following error:

(APIServer pid=966592)     raise KeyError(f"invalid tool call parser: {args.tool_call_parser} "
(APIServer pid=966592) KeyError: 'invalid tool call parser: openai (chose from { deepseek_v3, glm45, granite-20b-fc, granite, hermes, hunyuan_a13b, internlm, jamba, kimi_k2, llama4_pythonic, llama4_json, llama3_json, minimax, mistral, phi4_mini_json, pythonic, qwen3_coder, step3, xlam })'

Expected Behavior:
openai should be a valid option for --tool-call-parser, or the documentation should clarify which parsers are supported.

Environment:

• vLLM version: [0.10.1]
• GPU type and count: [H100 * 8]

Question:

• Is openai parser supported in the latest version?
• If not, which parser should be used for OpenAI-style tool calls?

[wizche]

From what I see the openai tool parser has been merged into the main branch, but it isn’t in a tagged release yet. Let’s wait for the next vLLM release.

[foloumi]

But even the specific pre-release for GPT OSS has this issue. Using
uv pip install --pre vllm==0.10.1+gptoss \ --extra-index-url https://wheels.vllm.ai/gpt-oss/ \ --extra-index-url https://download.pytorch.org/whl/nightly/cu128 \ --index-strategy unsafe-best-match
I get the same KeyError!

[foloumi]

From what I see the openai tool parser has been merged into the main branch, but it isn’t in a tagged release yet. Let’s wait for the next vLLM release.

Do we know when the next release will be?

[Cyb4Black]

fyi as long as main branch is compiling you can find nightly builds of vllm-openai image here:
https://hub.docker.com/repository/docker/hanseware/vllm-nightly

stumbled upon this issue myself yesterday and switched to this nightly build (for now). works for me on an h100 deployed via kubernetes.