[feature] 迁移 axolotl Quick Start 并接入看护 - #149
Conversation
CLA Signature Passlicy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
|
📖 文档预览已随本 PR 关闭而清理。 |
| export PYTHONNOUSERSITE=1 | ||
| ``` | ||
|
|
||
| `PYTHONNOUSERSITE=1` 令 Python 忽略用户目录中的包，避免旧版 CANN 相关包干扰依赖解析。 |
There was a problem hiding this comment.
这个环境变量是必须的么？ 按道理说我们认为当前环境没有干扰的包
There was a problem hiding this comment.
已删。入门页按干净环境写，可见路径和测试里都不再设 PYTHONNOUSERSITE。
| 输出结果如下： | ||
|
|
||
| ```shell #test-result id="check-py" fuzzy="xxx" | ||
| Python 3.12.xxx |
There was a problem hiding this comment.
已改。软件表不再写死 3.12，只要求落在官方配套表范围内并满足 axolotl 下限；可见的 check-py 整段已删。
|
|
||
| ## 3. 安装 PyTorch NPU 栈 | ||
|
|
||
| 从华为 PyPI 额外索引安装与 CANN 9.1.0 匹配的 `torch==2.11.0` / `torch_npu==2.11.0`，并一并装上 `numpy`、`pyyaml`。版本行出现 `2.11.0+cpu` 属正常，`npu_available` 须为 `True`。 |
There was a problem hiding this comment.
建议就安装torch_npu和cann的优选版本，可以给一个链接，这样可以避免文档中写死版本号，避免要经常更新
There was a problem hiding this comment.
已改。可见命令只装官方当前推荐的 torch_npu，主索引走 CPU 轮子源，不再写死版本。配套链接见软件表。
|
|
||
| ```shell #test id="install-axolotl" load="axolotl_ver>>ver" | ||
| python -m pip install --no-build-isolation --no-deps "axolotl==<ver>" | ||
| python -m pip install \ |
There was a problem hiding this comment.
这是为什么要手动安装依赖包呢？ requirement.txt不行么
There was a problem hiding this comment.
已改。上游声明在 pyproject.toml。先装好 NPU 栈，再 pip install --no-build-isolation axolotl，不再手抄 pin。
CLA Signature Passlicy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
CLA Signature Passlicy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
33f1917 to
e2126f4
Compare
CLA Signature Passlicy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
把 axolotl 从 workflows 源仓迁到本仓共享引擎：可见下载改 Hugging Face， 薄触发器不传 --volume，着陆页只 include 以免侧栏出现「快速开始」夹层。
可见路径不再写死 torch/Python，也不再手抄依赖。seed: 42 下第一步 loss 钉 12.49，后几步只比对字段。
可见块只写工作目录下载；看护用隐藏 #test-setup 命中 /root/.cache/huggingface。
e2126f4 to
f0a4f69
Compare
CLA Signature Passlicy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
3 participants
把
axolotl的昇腾 Quick Start 从cosdt-ci-test/workflows迁进本仓，接到已有的quick-start-template.yml。Summary
torch/torch_npu，不再写死版本；用--no-build-isolation装 axolotl，不再手抄依赖。PYTHONNOUSERSITE和「必须先是 Python 3.12」门槛。seed: 42。看护钉第一步loss为12.49，后几步只比对loss/train_loss字段。#test-setup命中 runner 的/root/.cache/huggingface，页面上看不到缓存命令。Test plan
sources/axolotl/quick_start.md报 ok/root/.cacheaxolotl-quick-startjob 绿：npu_available True、"device": "npu:0"、第一步{'loss': '12.49'、Training completed!download-model应在数秒内结束，模型来自/root/.cache/huggingface/hub