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-160/sources/flagscale/ 也可打开整站首页。 构建自 |
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. 👍 |
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. 👍 |
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. 👍 |
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. 👍 |
CLA Signature Passlicy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
f02aa24 to
9bad828
Compare
CLA Signature Passlicy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
把 flagos-ai/FlagScale 的昇腾 Quick Start 迁到本仓共享引擎：文档改为 #test 方言、模型走 Hugging Face、vLLM 改官方三步 wheel，避免源码编译与大陆代理路径。
vllm-plugin-FL v0.3.0-rc1.post1 会调用 vLLM 0.24 才有的 logical_device_id_to_visible_device_id；0.23 上该方法为 None， 双卡 inference worker 启动即 TypeError。
vLLM 0.24 的 CudaCommunicator 会 import flashinfer.comm；该包在 import 时断言 libcudart，昇腾双卡 worker 启动即崩。
triton-ascend 3.2.2 不接受 FlagGems 5.3.4 的 chained or，双卡加载模型时 pow_scalar 编译失败。用官方 VLLM_FL_FLAGOS_BLACKLIST_APPEND 改走 torch_npu。
社区 vLLM 0.24 用 AttentionBackendEnum 查找后端，不认识插件返回的 ASCEND_FL，加载 Qwen2 时会直接失败。
triton-ascend 编不了 FlagGems 5.3.4 的昇腾 copy/pow kernel，加载权重时会在 _copy_kernel 上以 coreDim=0 退出。
vLLM 0.24 的 model runner 会读 supports_update_block_table，插件 tag 里的 Ascend builder 没有这个属性，生成第一批 token 时会 AttributeError。
去掉 FlagGems 与 vLLM FL 插件及本地改源，改用 vLLM 0.23.0 + vllm-ascend 0.23.0，避免再为 FL 接口打补丁。
9bad828 to
48840d8
Compare
CLA Signature Passlicy666, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
2 participants
Summary
flagos-ai/FlagScale昇腾 Quick Start 迁到sources/flagscale/。装 vLLM 改为统一三步 pip（vllm/vllm-ascend0.23.0，最后强制重装triton-ascend），不再源码编译 vLLM / vllm-ascend；模型改为 Hugging FaceQwen/Qwen2.5-0.5B；双卡离线推理的预期输出含device_type npu、backend=hccl与output.outputs[0].text=。flagscale-quick-start.yml（linux-aarch64-a2-2，不传container_options/--volume）+tests/flagscale子类；首页推理区增加卡片和 toctree。相对源仓的刻意差异
--volume和VLLM_TARGET_DEVICE=empty源码编译。香港 A2 直连官方 URL，模型走 Hub 默认缓存。vllm-plugin-FL官方表是0.20.2或0.24.0。本文跟已跑通的文档引擎配了对，钉vLLM 0.23.0官方 wheel，插件用接近 0.24 的v0.3.0-rc1.post1。若 NPU 上插件 API 对不上，再单独改插件 tag，不要退回源码编 vLLM。Test plan
#test+ 3 条#test-setup）、可见正文泄漏检查、actionlint、测试文件py_compileworkflow_dispatch跑冷/热缓存；fork PR 的doc_url会 404，属预期platform_name PlatformFL/device_type npu/dist_backend hccl；推理块出现backend=hccl与output.outputs[0].text=/root/.cache/huggingface/hub，可见块不再完整下载