Conversation
CLA Signature Passlicy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
|
📖 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 |
|
已更新:一步训练冻结 |
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. 👍 |
e7df44b to
4e10e40
Compare
CLA Signature Passlicy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
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。
4e10e40 to
30d5b18
Compare
CLA Signature Passlicy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
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。container_options/--volume,镜像钉cann:9.1.0-910b-ubuntu22.04-py3.12。.py+ 隐藏#test-setupcat+ 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)
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。config.json与model.safetensors存在),不再走 ModelScope。launch+Booster(TorchDDPPlugin)一步训练。Test plan
actionlint薄触发器py_compile测试子类#test+ 1 个隐藏#test-setup),#test-result4 条cosdt-ci-test//root/.cache/modelscope/gh-proxypython - <<;train_one_step.py围栏与隐藏cat字节一致OK,637s;install-torchnpu_available True;install-colossalaiaccel_name npu/accel_device npu:0;trainboosted_param_device npu:0OK,95s(其中train34s)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.com与PIP_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。