A Reddit user is requesting a production-ready VLLM configuration for the Qwen 2.5-27B model after experiencing significant performance degradation with FP8 safetensors.
The user reports that using FP8 weights results in quality benchmark drops of nearly 20% compared to llama.cpp variants, particularly under high concurrency loads.
They are asking the community to share their working VLLM configurations to resolve these issues.