GPT-OSS & vLLM Troubleshooter FAQs

Question 1

Which vLLM flags are required for gpt-oss tool calling?

Accepted Answer

The vLLM server must be started with --tool-call-parser openai and --enable-auto-tool-choice to correctly handle tool calling for gpt-oss models.

Question 2

Does this skill help with client-side llama-stack issues?

Accepted Answer

Yes, by helping you identify when errors reported by clients like llama-stack are actually originating from the vLLM server, preventing wasted time searching the client codebase for fixes.

Question 3

Can I use this for both 20b and 120b gpt-oss models?

Accepted Answer

Yes, the skill covers diagnostic patterns and configuration requirements applicable to both the 20b and 120b variants of the GPT-OSS model family.

Question 4

How do I fix Unexpected token errors with GPT-OSS models?

Accepted Answer

The most common fix involves updating your model's generation_config.json files from the latest HuggingFace repository and ensuring your vLLM server is running a recent version.

Question 5

What causes openai_harmony.HarmonyError in vLLM?

Accepted Answer

This error originates from the vLLM server's serving layer, often indicating a mismatch between expected Harmony format control tokens and the tokens actually generated by the model.

GPT-OSS & vLLM Troubleshooter

Key Features

Use Cases

GPT-OSS & vLLM Troubleshooter

Key Features

Use Cases