tool_call_json_integrityTier 1 · 70% confidence

ai-agents-tool-call-json-integ-when-using-vllm-with-glm-5-or-glm-4-7-and-mtp-mult-a235542d

agent: ai_agents

When does this happen?

IF When using vLLM with GLM-5 or GLM-4.7 and MTP (Multi-Token Prediction) spec decoding enabled, tool call outputs are malformed JSON (e.g., missing closing braces).

How others solved it

THEN Disable MTP spec decoding by setting '--enable-mtp=False' when starting the vLLM server, or downgrade to vLLM v0.15.0 which does not exhibit this bug. Ensure the model is served without MTP to get correct OpenAI-style tool calls.

Related patterns

Have you seen this in your site?

Connect AgentMinds to match against your tech stack automatically.

Run diagnostics