Skip to content

feat(evolve-kernel): HIP C++ 优化模式 + Multi-GPU 并行 + WebUI 重写 - #25

Open
6eanut wants to merge 1 commit into
HuangPuStar:next-appfrom
6eanut:next-app
Open

feat(evolve-kernel): HIP C++ 优化模式 + Multi-GPU 并行 + WebUI 重写#25
6eanut wants to merge 1 commit into
HuangPuStar:next-appfrom
6eanut:next-app

Conversation

@6eanut

@6eanut 6eanut commented Aug 7, 2026

Copy link
Copy Markdown
Contributor

新增功能:

  • HIP C++ (from scratch) 优化模式:自适应提示词、hipcc 编译验证、.cpp 持久化
  • Multi-GPU 并行优化:4 GPU 独立并行,per-GPU shape target 拆分
  • hipprof 细粒度 GPU profiling (bandwidth/occupancy/l2 cache)
  • 自适应 HIP 提示词:不再硬编码 int8 GEMM,agent 自行分析算法
  • 自动 HIP 升级:near_optimal 或 headroom<10% 时提示词自动建议 HIP 重写

WebUI 改进:

  • 全新 Multi-GPU Dashboard:4 张 GPU 卡片,phase 着色,一键展开 per-GPU detail
  • Per-GPU Detail Panel:状态机 + Kernel Library + Shape Targets + Agents 滚动列表
  • Shape Target 卡片:TP badge、M≤16 / M=4096 重点标记、K×N 尺寸
  • Kernel Library 重写:彩色行、进度条、headroom badge、hipprof 标记、固定表头
  • 代码展示带行号 (CSS counter)、改进摘要改为条目列表

Bug 修复:

  • Per-GPU shape 正则:支持 (M, K) @ (K, N) 格式,避免误匹配 kernel 源码
  • launcher._parse_state_dir 正则修复
  • 向后兼容旧 kernel_library.json 字段

…ent upstream

Layers the fork's evolve_kernel work (HIP C++ mode, 4-GPU parallel
optimization, hipprof profiling, Multi-GPU WebUI rewrite) onto the current
upstream next-app base. Other evolve_kernel features already upstream are
preserved; the tree here is a functional superset (includes the worker_nodes
remote-perf wiring). torch import made optional so CPU-only CI can collect
tests, matching upstream convention.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant