What if vLLM hits the context limit and DSH compaction also fails?
DeepSeekAgent EditorialPublished 11/10/2026Updated 11/10/2026
Short chats work with self-hosted vLLM, while long chats and compaction fail. Which budget should I check?
Original source
1 answer
These troubleshooting steps are based on the original report. The issue still needs to be reproduced, and a fix has yet to be confirmed.
Compare vLLM's configured context limit with the model's input and output budgets in Harness. System instructions, tools, and history can share the budget. If compaction itself is too large, preserve the session and start a new one with a concise handoff.