Skip to content

[feature] 迁移 ColossalAI Quick Start 并接入看护 - #164

Open
licy666 wants to merge 3 commits into
Ascend:mainfrom
licy666:feat/colossalai-quick-start
Open

licy666 wants to merge 3 commits into
Ascend:mainfrom
licy666:feat/colossalai-quick-start

Conversation

@licy666

@licy666 licy666 commented Sep 9, 2026

Copy link
Copy Markdown
Contributor

Summary

  • hpcaitech/ColossalAI 的昇腾 Quick Start 从 cosdt-ci-test/workflows 迁到 Ascend/docs:用户文档、测试、薄触发器、首页卡片和 toctree 同一 PR。
  • 可见命令直连 Hugging Face 与 GitHub 官方地址;模型走默认 Hub 缓存。薄触发器不传 container_options / --volume,镜像钉 cann:9.1.0-910b-ubuntu22.04-py3.12
  • 训练脚本拆成无标签 .py + 隐藏 #test-setup cat + 可见 python train_one_step.py。工作负载 #test-result 含设备锚点 boosted_param_device npu:0

选型(不进用户正文)

  • 安装仍用 colossalai==0.5.0 --no-deps:上游元数据钉 torch<=2.5.1,裸 pip install colossalai 会换掉 CANN 9.1 配套的 torch==2.9.0 / torch_npu==2.9.0.post2
  • 模型用 Qwen/Qwen2.5-0.5B(已 HEAD 确认 config.jsonmodel.safetensors 存在),不再走 ModelScope。
  • 不克隆上游源码:PyPI 轮子即可覆盖 launch + Booster(TorchDDPPlugin) 一步训练。

Test plan

  • actionlint 薄触发器
  • py_compile 测试子类
  • 解析块数 5(4 个 #test + 1 个隐藏 #test-setup),#test-result 4 条
  • 可见正文无 cosdt-ci-test / /root/.cache / modelscope / gh-proxy
  • 可见区无 python - <<train_one_step.py 围栏与隐藏 cat 字节一致
  • coder A2 冷缓存完整 unittest:OK,637s;install-torch npu_available Trueinstall-colossalai accel_name npu / accel_device npu:0train boosted_param_device npu:0
  • coder A2 热缓存完整 unittest:OK,95s(其中 train 34s)
  • 合入默认分支后 workflow_dispatch 香港 A2;合入前 dispatch 会对新 workflow 404

冷缓存在大陆 coder 上为打通 Hugging Face 出口,使用了进程级 HF_ENDPOINT=https://hf-mirror.comPIP_INDEX_URL(华为云 PyPI)。这两项不进用户文档。香港 A2 由引擎集群 PyPI 缓存 + 直连 huggingface.co

@ascend-robot

Copy link
Copy Markdown

CLA Signature Pass

licy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍

github-actions Bot pushed a commit that referenced this pull request Sep 9, 2026
@github-actions

github-actions Bot commented Sep 9, 2026

Copy link
Copy Markdown
Contributor

📖 文档预览https://ascend.github.io/docs/pr-preview/pr-164/sources/colossalai/

也可打开整站首页

构建自 4e10e40f658d1215ee67e238fb0faca57e1e7ef2
每次推送后自动更新,PR 关闭时自动清理。

@licy666

licy666 commented Sep 9, 2026

Copy link
Copy Markdown
Contributor Author

已更新:一步训练冻结 SEED=42 / PYTHONHASHSEED=42#test-result 钉死 coder 上 ModelScope 权重连跑两遍得到的 loss 5.531600transformers 钉在 4.51.3。用户文档仍走 Hugging Face 的 Qwen/Qwen2.5-0.5B

@ascend-robot

Copy link
Copy Markdown

CLA Signature Pass

licy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍

github-actions Bot pushed a commit that referenced this pull request Sep 9, 2026
@ascend-robot

Copy link
Copy Markdown

CLA Signature Pass

licy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍

github-actions Bot pushed a commit that referenced this pull request Sep 10, 2026
@licy666
licy666 force-pushed the feat/colossalai-quick-start branch from e7df44b to 4e10e40 Compare September 21, 2026 02:25
@ascend-robot

Copy link
Copy Markdown

CLA Signature Pass

licy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍

github-actions Bot pushed a commit that referenced this pull request Sep 21, 2026
香港 A2 直连 Hugging Face 与官方 GitHub,文档不再走 ModelScope、组织代理和容器 bind-mount。
SEED=42 与 PYTHONHASHSEED 固定随机性,预期输出钉死 coder 实测的 loss 5.531600。
单独一行 `...` 要求锚点前还有 stdout;香港 runner 把 ColossalAI 日志打到 stderr,导致 loss 已对上仍判 mismatch。
@ascend-robot

Copy link
Copy Markdown

CLA Signature Pass

licy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants