lm-sys/FastChat

Issues with VLLM Integration Speedup

Open

#2,362 建立於 2023年9月5日

在 GitHub 查看
 (4 留言) (0 反應) (0 負責人)Python (38,959 star) (4,736 fork)batch import
good first issue

描述

Hello,

I've been trying to work with the [vLLM integration] and I'm facing some performance discrepancies. According to the documentation, I should achieve a significant speedup, but in my tests, I'm seeing different results:

Directly running with FastChat: 16 t/s Using the VLLM integration: 25 t/s (only 1.5x speedup) VLLM offline inference: 90 t/s (expected 6x speedup) I'm running Vicuna33b on a gin H100 gpu. Has anyone experienced this before? Are there any additional configurations or tweaks I might be missing to get the desired speedup?

Thanks in advance for any guidance or advice!

貢獻者指南

技術棧
python
領域
performancebackend
議題類型
bug
難度面向新貢獻者的預計實作難度，1 表示很小改動，5 表示專家級工作。
4
預計時間有經驗貢獻者完成調查、實作、測試並準備 pull request 的粗略時間範圍。
half day
活動狀態議題目前的可參與程度：新鮮、活躍、陳舊、阻塞或等待維護者輸入。
stale
清晰度議題是否清楚說明預期改動、驗收標準和下一步。
mostly clear
前置要求
PythonFastChat basicsvLLM integration
新手友善度1-100 的估計分數，表示該議題對首次貢獻者的友善程度。
35
研究方向
Investigate the vLLM integration code in FastChat, specifically the configuration and how it differs from offline vLLM inference. Compare the settings used in FastChat (e.g., in `fastchat/serve/vllm worker.py`) with the offline example. Check the documentation at `docs/vllm integration.md` for any missing flags. Look at the environment setup (e.g., CUDA, model loading) that might cause the performance gap. Also, search for similar issues in the FastChat repo for hints.