lllyasviel/Fooocus

system memory slowly increase

Open

#1,646 建立於 2023年12月29日

在 GitHub 查看
 (7 留言) (1 反應) (0 負責人)Python (48,116 star) (7,843 fork)batch import
bughelp wanted

描述

Read Troubleshoot

[x] I admit that I have read the Troubleshoot before making this issue.

Describe the problem The memory used by the program seems to be slowly increasing

Full Console Log image image image This is a potential problem that cannot be expressed in log. You can see that the memory occupied by fooocus is slowly increasing. My startup params: python entry_with_update.py --listen 0.0.0.0 --all-in-fp16 --port 8000 --always-gpu

貢獻者指南

技術棧
pythonpytorch
領域
backendmachine learning
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。
3
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。
needs investigation
前置要求
PythonPyTorchmemory profiler
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。
30
研究方向
Use Python memory profilers (e.g., memory profiler, tracemalloc) to track memory usage over time. Focus on the image generation loop in the Fooocus codebase, particularly around tensor allocations and GPU memory management. Check for any global variables or caches that might accumulate. Review the use of torch.cuda.empty cache() and ensure tensors are properly deleted after use. The startup param all in fp16 suggests mixed precision, which could be related to memory issues if not handled correctly.