dsh-context-budget
DeepSeek Harness插件:将本地模型的上下文保持在GPU能良好处理的大小(实测预填充速度、硬上限、早期压缩)
DeepSeek Harness plugin: keep a local model's context at a size your GPU handles well (measured prefill speed, hard ceiling, early compaction)
怎么安装
dsh plugin add dsh-context-budget 简介
dsh-context-budget A DeepSeek Harness plugin that keeps a local model's context at a size your GPU handles well. dsh compacts a conversation when it reaches a fixed fraction of the model's declared context window. On local hardware the declared window is not very practical. For example, a 27B model that advertises 256K may read prompts at 300 tokens per second when the context is small and at 70 tokens per second past 100K. So a turn could take a few seconds to start early in the conversation, but later in the same session take half an hour. Also, the GPU may hang before the window is full. Th…
推荐参考
信息
- 协议
- MIT
- 语言
- JavaScript
- GitHub 星标
- 0
- 月下载
- 547
- 最近更新
- 2026-08-30
- 创建于
- 2026-08-29
基础安全检查
- 检查结果
- 无
- 收录来源
- curated:0xsline/awesome-deepseek-harness
- 主题标签
- deepseek-harness, dsh-plugin
同类推荐
hindsight
vectorize-io/hindsight/tree/main/hindsight-integrations/coding-agents
Hindsight:会学习的 Agent 长期记忆系统,自动召回/保存、知识页、深度反思与按仓库隔离的记忆银行。
firecrawl-mcp-server
firecrawl/firecrawl-mcp-server
🔥 官方Firecrawl MCP服务器 - 为Cursor、Claude及任何其他LLM客户端添加强大的网页抓取和搜索功能。
dsh-mnemon
omdsh-dev/dsh-mnemon
由 Mnemon 驱动的 DeepSeek Harness(DSH)跨 Agent、本地优先的持久记忆插件。它可在支持 Mnemon 的 Agent 之间共享长期记忆,并提供运行时记忆、可检索项目档案、语义召回、知识图谱和 Sidebar UI。
get-shit-done
gsd-build/get-shit-done
由 TÂCHES 开发的轻量且强大的元提示、上下文工程和规格驱动开发系统,适用于 Claude Code。