# SOME DESCRIPTIVE TITLE. # Copyright (C) 2025, vllm-ascend team # This file is distributed under the same license as the vllm-ascend # package. # FIRST AUTHOR , 2026. # #, fuzzy msgid "" msgstr "" "Project-Id-Version: vllm-ascend\n" "Report-Msgid-Bugs-To: \n" "POT-Creation-Date: 2026-08-05 06:27+0000\n" "PO-Revision-Date: YEAR-MO-DA HO:MI+ZONE\n" "Last-Translator: FULL NAME \n" "Language-Team: zh_CN \n" "Language: zh_CN\n" "MIME-Version: 1.0\n" "Content-Type: text/plain; charset=utf-8\n" "Content-Transfer-Encoding: 8bit\n" "Plural-Forms: nplurals=1; plural=0;\n" "Generated-By: Babel 2.18.0\n" #: ../../source/faqs.md:1 msgid "FAQs" msgstr "常见问题解答" #: ../../source/faqs.md:3 msgid "Version Specific FAQs" msgstr "版本特定常见问题解答" #: ../../source/faqs.md:5 msgid "" "[[v0.23.0rc1] FAQ & Feedback](https://github.com/vllm-project/vllm-" "ascend/issues/12238)" msgstr "" "[[v0.23.0rc1] 常见问题与反馈](https://github.com/vllm-project/vllm-" "ascend/issues/12238)" #: ../../source/faqs.md:6 msgid "" "[[v0.22.1rc1] FAQ & Feedback](https://github.com/vllm-project/vllm-" "ascend/issues/10593)" msgstr "" "[[v0.22.1rc1] 常见问题与反馈](https://github.com/vllm-project/vllm-" "ascend/issues/10593)" #: ../../source/faqs.md:7 msgid "" "[[v0.21.0rc1] FAQ & Feedback](https://github.com/vllm-project/vllm-" "ascend/issues/9970)" msgstr "" "[[v0.21.0rc1] 常见问题与反馈](https://github.com/vllm-project/vllm-" "ascend/issues/9970)" #: ../../source/faqs.md:8 msgid "" "[[v0.20.2rc1] FAQ & Feedback](https://github.com/vllm-project/vllm-" "ascend/issues/9586)" msgstr "" "[[v0.20.2rc1] 常见问题与反馈](https://github.com/vllm-project/vllm-" "ascend/issues/9586)" #: ../../source/faqs.md:9 msgid "" "[[v0.19.1rc1] FAQ & Feedback](https://github.com/vllm-project/vllm-" "ascend/issues/8819)" msgstr "" "[[v0.19.1rc1] 常见问题与反馈](https://github.com/vllm-project/vllm-" "ascend/issues/8819)" #: ../../source/faqs.md:10 msgid "" "[[v0.18.0] FAQ & Feedback](https://github.com/vllm-project/vllm-" "ascend/issues/8238)" msgstr "" "[[v0.18.0] 常见问题与反馈](https://github.com/vllm-project/vllm-ascend/issues/8238)" #: ../../source/faqs.md:12 msgid "General FAQs" msgstr "通用常见问题解答" #: ../../source/faqs.md:14 msgid "1. What devices are currently supported?" msgstr "1.当前支持哪些设备？" #: ../../source/faqs.md:16 msgid "" "Currently, **ONLY** Atlas A2 series (Ascend-cann-kernels-910b), Atlas A3 " "series (Atlas-A3-cann-kernels) and Atlas 300I (Ascend-cann-kernels-310p) " "series are supported:" msgstr "" "目前**仅**支持 Atlas A2 系列（Ascend-cann-kernels-910b）、Atlas A3 系列（Atlas-A3-cann-" "kernels）和 Atlas 300I 系列（Ascend-cann-kernels-310p）：" #: ../../source/faqs.md:18 msgid "" "Atlas A2 Training series (Atlas 800T A2, Atlas 900 A2 PoD, Atlas 200T A2 " "Box16, Atlas 300T A2)" msgstr "" "Atlas A2 训练系列（Atlas 800T A2、Atlas 900 A2 PoD、Atlas 200T A2 Box16、Atlas 300T " "A2）" #: ../../source/faqs.md:19 msgid "Atlas 800I A2 Inference series (Atlas 800I A2)" msgstr "Atlas 800I A2 推理系列（Atlas 800I A2）" #: ../../source/faqs.md:20 msgid "" "Atlas A3 Training series (Atlas 800T A3, Atlas 900 A3 SuperPoD, Atlas 9000 " "A3 SuperPoD)" msgstr "" "Atlas A3 训练系列（Atlas 800T A3、Atlas 900 A3 SuperPoD、Atlas 9000 A3 SuperPoD）" #: ../../source/faqs.md:21 msgid "Atlas 800I A3 Inference series (Atlas 800I A3)" msgstr "Atlas 800I A3 推理系列（Atlas 800I A3）" #: ../../source/faqs.md:22 msgid "[Experimental] Atlas 300I Inference series (Atlas 300I Duo)." msgstr "[实验性] Atlas 300I 推理系列（Atlas 300I Duo）。" #: ../../source/faqs.md:23 msgid "" "[Experimental] Currently for 310I Duo the stable version is vllm-ascend " "v0.10.0rc1." msgstr "[实验性] 目前 310I Duo 的稳定版本为 vllm-ascend v0.10.0rc1。" #: ../../source/faqs.md:25 msgid "Below series are NOT supported yet:" msgstr "以下系列尚不支持：" #: ../../source/faqs.md:27 msgid "Atlas 200I A2 (Ascend-cann-kernels-310b) unplanned yet" msgstr "Atlas 200I A2（Ascend-cann-kernels-310b）尚未规划" #: ../../source/faqs.md:28 msgid "Ascend 910, Ascend 910 Pro B (Ascend-cann-kernels-910) unplanned yet" msgstr "Ascend 910、Ascend 910 Pro B（Ascend-cann-kernels-910）尚未规划" #: ../../source/faqs.md:30 msgid "" "From a technical view, vllm-ascend supports devices if TorchNPU is " "supported. Otherwise, we have to implement it by using custom ops. We also " "welcome you to join us to improve together." msgstr "" "从技术角度来看，如果 TorchNPU 受支持，vllm-ascend " "即可支持相应设备。否则，我们需要通过自定义算子来实现。我们也欢迎您加入我们，共同改进。" #: ../../source/faqs.md:32 msgid "2. How to get our docker containers?" msgstr "2.如何获取我们的 Docker 容器？" #: ../../source/faqs.md:34 msgid "" "You can get our containers at `Quay.io`, e.g., [vllm-" "ascend](https://quay.io/repository/ascend/vllm-ascend?tab=tags) and " "[cann](https://quay.io/repository/ascend/cann?tab=tags)." msgstr "" "您可以在 `Quay.io` 获取我们的容器，例如 [vllm-" "ascend](https://quay.io/repository/ascend/vllm-ascend?tab=tags) 和 " "[cann](https://quay.io/repository/ascend/cann?tab=tags)。" #: ../../source/faqs.md:36 msgid "" "If you are in China, you can use `daocloud` or some other mirror sites to " "accelerate your downloading:" msgstr "如果您在中国，可以使用 `daocloud` 或其他镜像站点来加速下载：" #: ../../source/faqs.md:46 msgid "Load Docker Images for offline environment" msgstr "为离线环境加载 Docker 镜像" #: ../../source/faqs.md:48 msgid "" "If you want to use container image for offline environments (no internet " "connection), you need to download container image in an environment with " "internet access:" msgstr "如果您想在离线环境（无网络连接）中使用容器镜像，需要在有网络的环境中下载容器镜像：" #: ../../source/faqs.md:50 msgid "**Exporting Docker images:**" msgstr "**导出 Docker 镜像：**" #: ../../source/faqs.md:62 msgid "**Importing Docker images in environment without internet access:**" msgstr "**在无网络环境中导入 Docker 镜像：**" #: ../../source/faqs.md:74 msgid "3. What models does vllm-ascend support?" msgstr "3.vllm-ascend支持哪些模型？" #: ../../source/faqs.md:76 msgid "" "Find more details " "[here](https://docs.vllm.ai/projects/ascend/en/latest/user_guide/support_matrix/supported_models.html)." msgstr "" "更多详情请见[此处](https://docs.vllm.ai/projects/ascend/en/latest/user_guide/support_matrix/supported_models.html)。" #: ../../source/faqs.md:78 msgid "4. How to get in touch with our community?" msgstr "4.如何与我们的社区取得联系？" #: ../../source/faqs.md:80 msgid "" "There are many channels that you can communicate with our community " "developers / users:" msgstr "您可以通过多种渠道与我们的社区开发者/用户交流：" #: ../../source/faqs.md:82 msgid "" "Submit a GitHub [issue](https://github.com/vllm-project/vllm-" "ascend/issues?page=1)." msgstr "" "提交 GitHub [issue](https://github.com/vllm-project/vllm-" "ascend/issues?page=1)。" #: ../../source/faqs.md:83 msgid "" "Join our [weekly " "meeting](https://docs.google.com/document/d/1hCSzRTMZhIB8vRq1_qOOjx4c9uYUxvdQvDsMV2JcSrw/edit?tab=t.0#heading=h.911qu8j8h35z)" " and share your ideas." msgstr "" "参加我们的[周会](https://docs.google.com/document/d/1hCSzRTMZhIB8vRq1_qOOjx4c9uYUxvdQvDsMV2JcSrw/edit?tab=t.0#heading=h.911qu8j8h35z)并分享您的想法。" #: ../../source/faqs.md:84 msgid "" "Join our [WeChat](https://github.com/vllm-project/vllm-" "ascend/issues/227) group and ask your questions." msgstr "" "加入我们的[微信](https://github.com/vllm-project/vllm-" "ascend/issues/227)群组并提问。" #: ../../source/faqs.md:85 msgid "" "Join our ascend channel in [vLLM " "forums](https://discuss.vllm.ai/c/hardware-support/vllm-ascend-" "support/6) and publish your topics." msgstr "" "加入 [vLLM 论坛](https://discuss.vllm.ai/c/hardware-support/vllm-ascend-" "support/6) 中的 ascend 频道并发布您的主题。" #: ../../source/faqs.md:87 msgid "5. What features does vllm-ascend V1 support?" msgstr "5.vllm-ascend V1支持哪些功能？" #: ../../source/faqs.md:89 msgid "" "Find more details " "[here](https://docs.vllm.ai/projects/ascend/en/latest/user_guide/support_matrix/supported_features.html)." msgstr "" "更多详情请见[此处](https://docs.vllm.ai/projects/ascend/en/latest/user_guide/support_matrix/supported_features.html)。" #: ../../source/faqs.md:91 msgid "" "6. How to solve the problem of \"Failed to infer device type\" or " "\"libatb.so: cannot open shared object file\"?" msgstr "" "6.如何解决“Failed to infer device type”或“libatb.so: cannot open shared object " "file”的问题？" #: ../../source/faqs.md:93 msgid "" "Basically, the reason is that the NPU environment is not configured " "correctly. You can:" msgstr "基本上，原因是 NPU 环境配置不正确。您可以：" #: ../../source/faqs.md:95 msgid "" "try `source /usr/local/Ascend/nnal/atb/set_env.sh` to enable NNAL package." msgstr "尝试执行 `source /usr/local/Ascend/nnal/atb/set_env.sh` 以启用 NNAL 包。" #: ../../source/faqs.md:96 msgid "" "try `source /usr/local/Ascend/ascend-toolkit/set_env.sh` to enable CANN " "package." msgstr "尝试执行 `source /usr/local/Ascend/ascend-toolkit/set_env.sh` 以启用 CANN 包。" #: ../../source/faqs.md:97 msgid "try `npu-smi info` to check whether the NPU is working." msgstr "尝试执行 `npu-smi info` 检查 NPU 是否正常工作。" #: ../../source/faqs.md:99 msgid "" "If the above steps are not working, you can try the following code in Python" " to check whether there are any errors:" msgstr "如果上述步骤无效，您可以尝试在 Python 中运行以下代码检查是否有错误：" #: ../../source/faqs.md:107 msgid "" "If all above steps are not working, feel free to submit a GitHub issue." msgstr "如果以上所有步骤均无效，请随时提交 GitHub issue。" #: ../../source/faqs.md:109 msgid "7. How does vllm-ascend work with vLLM?" msgstr "7.vllm-ascend如何与vLLM协同工作？" #: ../../source/faqs.md:111 msgid "" "`vllm-ascend` is a hardware plugin for vLLM. Stable releases usually align " "with the same vLLM version, while RC releases may use the corresponding vLLM" " final release version. For example, `vllm-ascend` `v0.18.0rc1` matches vLLM" " `v0.18.0`. For the main branch, we ensure that `vllm-ascend` and `vllm` are" " compatible at every commit." msgstr "" "`vllm-ascend` 是 vLLM 的硬件插件。稳定版本通常与相同版本的 vLLM 对齐，而 RC 版本可能使用对应的 vLLM " "最终发布版本。例如，`vllm-ascend` `v0.18.0rc1` 匹配 vLLM `v0.18.0`。对于 main 分支，我们确保 " "`vllm-ascend` 和 `vllm` 在每个提交上均保持兼容。" #: ../../source/faqs.md:113 msgid "" "8. Does vllm-ascend support Prefill-Decode (PD) Disaggregation feature?" msgstr "8.vllm-ascend 是否支持 Prefill-Decode (PD) 分离特性？" #: ../../source/faqs.md:115 msgid "" "Yes, vllm-ascend supports Prefill-Decode Disaggregation feature with " "Mooncake backend. See the [official " "tutorial](https://docs.vllm.ai/projects/ascend/en/latest/tutorials/features/pd_disaggregation_mooncake_multi_node.html)" " for example." msgstr "" "是的，vllm-ascend 支持基于 Mooncake 后端的 Prefill-Decode " "分离特性。示例请参见[官方教程](https://docs.vllm.ai/projects/ascend/en/latest/tutorials/features/pd_disaggregation_mooncake_multi_node.html)。" #: ../../source/faqs.md:117 msgid "9. Does vllm-ascend support quantization method?" msgstr "9.vllm-ascend 是否支持量化方法？" #: ../../source/faqs.md:119 msgid "" "Currently, w8a8, w4a8, and w4a4 quantization methods are already supported " "by vllm-ascend." msgstr "目前，vllm-ascend 已支持 w8a8、w4a8 和 w4a4 量化方法。" #: ../../source/faqs.md:121 msgid "10. How is vllm-ascend tested?" msgstr "10.vllm-ascend 如何测试？" #: ../../source/faqs.md:123 msgid "" "vllm-ascend is tested in three aspects: functions, performance, and " "accuracy." msgstr "vllm-ascend 从三个方面进行测试：功能、性能和精度。" #: ../../source/faqs.md:125 msgid "" "**Functional test**: We added CI, including part of vllm's native unit tests" " and vllm-ascend's own unit tests. In vllm-ascend's tests, we test basic " "functionalities, popular model availability, and [supported " "features](https://docs.vllm.ai/projects/ascend/en/latest/user_guide/support_matrix/supported_features.html)" " through E2E test." msgstr "" "**功能测试**：我们增加了CI，包括部分vllm原生单元测试和vllm-ascend自有单元测试。在vllm-" "ascend的测试中，我们通过E2E测试验证基本功能、主流模型可用性以及[支持的特性](https://docs.vllm.ai/projects/ascend/en/latest/user_guide/support_matrix/supported_features.html)。" #: ../../source/faqs.md:127 msgid "" "**Performance test**: We provide [benchmark](https://github.com/vllm-" "project/vllm-ascend/tree/main/benchmarks) tools for E2E performance " "benchmark, which can be easily re-run locally. We will publish a perf " "website to show the performance test results for each pull request." msgstr "" "**性能测试**：我们提供了[基准测试](https://github.com/vllm-project/vllm-" "ascend/tree/main/benchmarks)工具用于E2E性能基准测试，可轻松在本地重新运行。我们将发布一个性能网站来展示每个拉取请求的性能测试结果。" #: ../../source/faqs.md:129 msgid "" "**Accuracy test**: We are working on adding accuracy test to the CI as well." msgstr "**精度测试**：我们也在努力将精度测试加入CI。" #: ../../source/faqs.md:131 msgid "" "**Nightly test**: we'll run full test every night to make sure the code is " "working." msgstr "**夜间测试**：我们将每晚运行完整测试以确保代码正常运行。" #: ../../source/faqs.md:133 msgid "" "For each release, we'll publish the performance test and accuracy test " "report in the future." msgstr "对于每个版本，我们未来将发布性能测试和精度测试报告。" #: ../../source/faqs.md:135 msgid "11. How to fix the error \"InvalidVersion\" when using vllm-ascend?" msgstr "11.如何修复使用vllm-ascend时出现的\"InvalidVersion\"错误？" #: ../../source/faqs.md:137 msgid "" "The problem is usually caused by the installation of a development or " "editable version of the vLLM package. In this case, we provide the " "environment variable `VLLM_VERSION` to let users specify the version of vLLM" " package to use. Please set the environment variable `VLLM_VERSION` to the " "version of the vLLM package you have installed. The format of `VLLM_VERSION`" " should be `X.Y.Z`." msgstr "" "该问题通常由安装了开发版或可编辑版的vLLM包导致。此时，我们提供环境变量`VLLM_VERSION`让用户指定要使用的vLLM包版本。请将环境变量`VLLM_VERSION`设置为您已安装的vLLM包版本。`VLLM_VERSION`的格式应为`X.Y.Z`。" #: ../../source/faqs.md:139 msgid "12. How to handle the out-of-memory issue?" msgstr "12.如何处理内存不足问题？" #: ../../source/faqs.md:141 msgid "" "OOM errors typically occur when the model exceeds the memory capacity of a " "single NPU. For general guidance, you can refer to [vLLM OOM troubleshooting" " documentation](https://docs.vllm.ai/en/latest/usage/troubleshooting/#out-" "of-memory)." msgstr "" "OOM错误通常发生在模型超过单个NPU内存容量时。通用指导可参考[vLLM " "OOM故障排查文档](https://docs.vllm.ai/en/latest/usage/troubleshooting/#out-of-" "memory)。" #: ../../source/faqs.md:143 msgid "" "In scenarios where NPUs have limited high bandwidth memory (on-chip memory) " "capacity, dynamic memory allocation/deallocation during inference can " "exacerbate memory fragmentation, leading to OOM. To address this:" msgstr "在NPU高带宽内存（片上内存）容量有限的场景下，推理过程中的动态内存分配/释放会加剧内存碎片化，导致OOM。为解决此问题：" #: ../../source/faqs.md:145 msgid "" "**Limit `--max-model-len`**: It can save the on-chip memory usage for KV " "cache initialization step." msgstr "**限制`--max-model-len`**：可节省KV缓存初始化步骤的片上内存使用。" #: ../../source/faqs.md:147 msgid "" "**Adjust `--gpu-memory-utilization`**: If unspecified, the default value is " "`0.9`. You can decrease this value to reserve more memory to reduce " "fragmentation risks. See details in: [vLLM - Inference and Serving - Engine " "Arguments](https://docs.vllm.ai/en/latest/cli/serve/#-gpu-memory-" "utilization)." msgstr "" "**调整`--gpu-memory-" "utilization`**：若未指定，默认值为`0.9`。可降低此值以预留更多内存，减少碎片化风险。详见：[vLLM - 推理与服务 - " "引擎参数](https://docs.vllm.ai/en/latest/cli/serve/#-gpu-memory-utilization)。" #: ../../source/faqs.md:149 msgid "" "**Configure `PYTORCH_NPU_ALLOC_CONF`**: Set this environment variable to " "optimize NPU memory management. For example, you can use `export " "PYTORCH_NPU_ALLOC_CONF=expandable_segments:True` to enable virtual memory " "feature to mitigate memory fragmentation caused by frequent dynamic memory " "size adjustments during runtime. See details in " "[PYTORCH_NPU_ALLOC_CONF](https://www.hiascend.com/document/detail/zh/Pytorch/700/comref/Envvariables/Envir_012.html)." msgstr "" "**配置`PYTORCH_NPU_ALLOC_CONF`**：设置此环境变量以优化NPU内存管理。例如，可使用`export " "PYTORCH_NPU_ALLOC_CONF=expandable_segments:True`启用虚拟内存功能，缓解运行时频繁动态内存大小调整导致的内存碎片化。详见[PYTORCH_NPU_ALLOC_CONF](https://www.hiascend.com/document/detail/zh/Pytorch/700/comref/Envvariables/Envir_012.html)。" #: ../../source/faqs.md:151 msgid "13. Failed to enable NPU graph mode when running DeepSeek" msgstr "13.运行DeepSeek时启用NPU图模式失败" #: ../../source/faqs.md:153 msgid "" "Enabling NPU graph mode for DeepSeek may trigger an error. This is because " "when both MLA (Multi-Head Latent Attention) and NPU graph mode are active, " "the number of queries per KV head must be 32, 64, or 128. However, " "DeepSeek-V2-Lite has only 16 attention heads, which results in 16 queries " "per KV—a value outside the supported range. Support for NPU graph mode on " "DeepSeek-V2-Lite will be added in a future update." msgstr "" "为DeepSeek启用NPU图模式可能触发错误。这是因为当MLA（多头潜在注意力）和NPU图模式同时激活时，每个KV头的查询数必须为32、64或128。然而，DeepSeek-V2-Lite仅有16个注意力头，导致每个KV头16个查询——该值不在支持范围内。对DeepSeek-V2-Lite的NPU图模式支持将在未来更新中添加。" #: ../../source/faqs.md:155 #, python-brace-format msgid "" "And if you're using DeepSeek-V3 or DeepSeek-R1, please make sure after the " "tensor parallel split, `num_heads`/`num_kv_heads` is {32, 64, 128}." msgstr "" "如果您使用的是DeepSeek-V3或DeepSeek-R1，请确保张量并行拆分后，`num_heads`/`num_kv_heads`为{32, " "64, 128}。" #: ../../source/faqs.md:162 msgid "" "14. Failed to reinstall vllm-ascend from source after uninstalling vllm-" "ascend" msgstr "14.卸载vllm-ascend后从源码重新安装失败" #: ../../source/faqs.md:164 msgid "" "You may encounter the problem of C/C++ compilation failure when reinstalling" " vllm-ascend from source using pip. If the installation fails, use `python " "setup.py install` (recommended) to install, or use `python setup.py clean` " "to clear the cache." msgstr "" "使用pip从源码重新安装vllm-ascend时可能遇到C/C++编译失败的问题。如果安装失败，请使用`python setup.py " "install`（推荐）进行安装，或使用`python setup.py clean`清除缓存。" #: ../../source/faqs.md:166 msgid "15. How to generate deterministic results when using vllm-ascend?" msgstr "15.使用vllm-ascend时如何生成确定性结果？" #: ../../source/faqs.md:168 msgid "There are several factors that affect output determinism:" msgstr "影响输出确定性的因素有以下几个：" #: ../../source/faqs.md:170 msgid "" "Sampler method: using **greedy sampling** by setting `temperature=0` in " "`SamplingParams`, e.g.:" msgstr "采样方法：通过在`SamplingParams`中设置`temperature=0`使用**贪心采样**，例如：" #: ../../source/faqs.md:195 msgid "Set the following environment parameters:" msgstr "设置以下环境参数：" #: ../../source/faqs.md:204 msgid "" "16. How to fix the error \"ImportError: Please install vllm[audio] for audio" " support\" for the multi-modal models?" msgstr "" "16.如何修复多模态模型的错误\"ImportError: Please install vllm[audio] for audio " "support\"？" #: ../../source/faqs.md:206 msgid "" "Some multi-modal models requires the `librosa` package to be installed, you " "need to install the `qwen-omni-utils` package to ensure all dependencies are" " met, for Qwen-omni, run `pip install qwen-omni-utils`. This package will " "install `librosa` and its related dependencies, resolving the `ImportError: " "No module named 'librosa'` issue and ensuring that the audio processing " "functionality works correctly." msgstr "" "部分多模态模型需要安装`librosa`包，您需要安装`qwen-omni-utils`包以确保满足所有依赖。对于Qwen-omni，运行`pip " "install qwen-omni-utils`。该包将安装`librosa`及其相关依赖，解决`ImportError: No module " "named 'librosa'`问题，确保音频处理功能正常工作。" #: ../../source/faqs.md:209 msgid "" "17. How to troubleshoot and resolve size capture failures resulting from " "stream resource exhaustion, and what are the underlying causes?" msgstr "17.如何排查和解决流资源耗尽导致的大小捕获失败问题？根本原因是什么？" #: ../../source/faqs.md:217 msgid "" "When vLLM Ascend recognizes this capture-time stream-resource signature in " "the error text, it re-raises the error with targeted guidance for ACL graph " "sizing and mitigation." msgstr "当vLLM Ascend在错误文本中识别出这种捕获时流资源特征时，会重新抛出错误，并附带针对ACL图大小调整和缓解措施的有针对性指导。" #: ../../source/faqs.md:219 msgid "Recommended mitigation strategies:" msgstr "推荐的缓解策略：" #: ../../source/faqs.md:221 msgid "" "Upgrade to a newer HDK/CANN stack if one is available for your environment. " "Recent releases improve ACL graph capacity, so older workarounds may no " "longer be necessary." msgstr "如果您的环境有可用的更新版HDK/CANN栈，请升级。最新版本提升了ACL图容量，因此旧有的变通方法可能不再需要。" #: ../../source/faqs.md:222 #, python-brace-format msgid "" "Manually reduce the configured graph sizes, for example: " "'{\"cudagraph_capture_sizes\":[size1, size2, size3, ...]}', or lower " "`max_cudagraph_capture_size`." msgstr "" "手动减小配置的图大小，例如：'{\"cudagraph_capture_sizes\":[size1, size2, size3, " "...]}'，或降低`max_cudagraph_capture_size`。" #: ../../source/faqs.md:223 msgid "" "If your workload is mostly uniform decode, try ACLGraph's `FULL` or " "`FULL_DECODE_ONLY` mode instead of the `PIECEWISE`." msgstr "如果您的负载主要是均匀解码，请尝试ACLGraph的`FULL`或`FULL_DECODE_ONLY`模式，而不是`PIECEWISE`。" #: ../../source/faqs.md:224 msgid "" "If you use `PIECEWISE` or `FULL_AND_PIECEWISE` and still hit this failure " "after upgrading, set `cudagraph_capture_sizes` manually according to your " "real workload and reduce the configured coverage." msgstr "" "如果您使用 `PIECEWISE` 或 `FULL_AND_PIECEWISE`，升级后仍然遇到此失败，请根据实际工作负载手动设置 " "`cudagraph_capture_sizes`，并减少配置的覆盖范围。" #: ../../source/faqs.md:225 msgid "" "If you are debugging a startup failure, temporarily disable graph mode " "(`cudagraph_mode=\"NONE\"` / `enforce_eager=True`) to confirm the issue is " "capture-related." msgstr "" "如果您正在调试启动失败，请临时禁用图模式（`cudagraph_mode=\"NONE\"` / " "`enforce_eager=True`）以确认问题与捕获相关。" #: ../../source/faqs.md:227 msgid "" "Root cause analysis: ACL graph capture can still fail when the runtime " "resources required by the selected graph sizes exceed what the current " "software/hardware stack can provide. This is most visible in `PIECEWISE` " "scenarios because the number of captured graphs scales with model depth and " "capture-size coverage. vLLM Ascend no longer auto-shrinks the PIECEWISE " "capture-size set locally, so the practical mitigations are to upgrade the " "HDK/CANN stack or reduce the configured graph sizes explicitly. The runtime " "guidance is intentionally narrow: it is only added when capture fails with " "the confirmed stream-resource signature above." msgstr "" "根因分析：当所选图大小所需的运行时资源超出当前软件/硬件栈所能提供的范围时，ACL 图捕获仍可能失败。这在 `PIECEWISE` " "场景中最为明显，因为捕获的图数量随模型深度和捕获大小覆盖范围而扩展。vLLM Ascend 不再自动缩小本地的 PIECEWISE " "捕获大小集合，因此实际的缓解措施是升级 HDK/CANN " "栈或显式减少配置的图大小。运行时指导信息有意保持狭窄：仅当捕获失败且具有上述确认的流资源特征时才会添加。" #: ../../source/faqs.md:230 msgid "18. How to install custom version of torch_npu?" msgstr "18.如何安装自定义版本的 torch_npu？" #: ../../source/faqs.md:232 msgid "" "TorchNPU will be overridden when installing vllm-ascend. If you need to " "install a specific version of TorchNPU, you can manually install the " "specified version of TorchNPU after vllm-ascend is installed." msgstr "" "安装 vllm-ascend 时，TorchNPU 将被覆盖。如果您需要安装特定版本的 TorchNPU，可以在安装 vllm-ascend " "后手动安装指定版本的 TorchNPU。" #: ../../source/faqs.md:234 msgid "" "19. On certain systems (e.g., Kylin OS), `docker pull` may fail with an " "`invalid tar header` error" msgstr "19.在某些系统（例如 Kylin OS）上，`docker pull` 可能会失败并出现 `invalid tar header` 错误" #: ../../source/faqs.md:236 msgid "" "On certain operating systems, such as Kylin OS, you may encounter an " "`invalid tar header` error during the `docker pull` process:" msgstr "" "在某些操作系统（例如 Kylin OS）上，您可能在 `docker pull` 过程中遇到 `invalid tar header` 错误：" #: ../../source/faqs.md:242 msgid "" "This is often due to system compatibility issues. You can resolve this by " "using an offline loading method with a second machine." msgstr "这通常是由于系统兼容性问题导致的。您可以通过使用第二台机器进行离线加载的方法来解决。" #: ../../source/faqs.md:244 msgid "" "On a separate host machine (e.g., a standard Ubuntu server), pull the image " "for the target ARM64 architecture and package it into a `.tar` file." msgstr "在单独的主机（例如标准 Ubuntu 服务器）上，拉取目标 ARM64 架构的镜像，并将其打包为 `.tar` 文件。" #: ../../source/faqs.md:257 msgid "Transfer the image archive" msgstr "传输镜像归档文件" #: ../../source/faqs.md:259 msgid "" "Copy the `vllm_ascend_.tar` file (where `` is the image tag you " "used) to your target machine" msgstr "将 `vllm_ascend_.tar` 文件（其中 `` 是您使用的镜像标签）复制到目标机器" #: ../../source/faqs.md:261 msgid "" "20. Why am I getting an error when executing the script to start a Docker " "container? The error message is: \"operation not permitted\"" msgstr "20.为什么执行启动 Docker 容器的脚本时出现错误？错误信息为：“operation not permitted”" #: ../../source/faqs.md:263 msgid "" "When using `--shm-size`, you may need to add the `--privileged=true` flag to" " your `docker run` command to grant the container necessary permissions. " "Please be aware that using `--privileged=true` grants the container " "extensive privileges on the host system, which can be a security risk. Only " "use this option if you understand the implications and trust the container's" " source." msgstr "" "使用 `--shm-size` 时，您可能需要在 `docker run` 命令中添加 `--privileged=true` " "标志，以授予容器必要的权限。请注意，使用 `--privileged=true` " "会授予容器对主机系统的广泛权限，这可能带来安全风险。仅在您了解其影响并信任容器来源的情况下使用此选项。" #: ../../source/faqs.md:265 msgid "" "21. How to set `SOC_VERSION` when building from source on a CPU-only " "machine?" msgstr "21.如何在仅 CPU 的机器上从源码构建时设置 `SOC_VERSION`？" #: ../../source/faqs.md:267 msgid "" "When building from source (e.g. `pip install -e .`), the build may try to " "infer the target chip via `npu-smi`. If `npu-smi` is not available (common " "in CPU-only build environments), you must set `SOC_VERSION` manually before " "installation." msgstr "" "从源码构建时（例如 `pip install -e .`），构建过程可能尝试通过 `npu-smi` 推断目标芯片。如果 `npu-smi` " "不可用（在仅 CPU 的构建环境中很常见），您必须在安装前手动设置 `SOC_VERSION`。" #: ../../source/faqs.md:269 msgid "" "You can use the defaults from `Dockerfile*` as a reference. For example:" msgstr "您可以使用 `Dockerfile*` 中的默认值作为参考。例如：" #: ../../source/faqs.md:285 msgid "22. Why does TPOT increase drastically as concurrency grows?" msgstr "22.为什么TPOT会随着并发量的增长而急剧增加？" #: ../../source/faqs.md:287 msgid "" "When testing a vLLM server, one may find that TPOT increases as concurrency " "increases (for example, TPOT increases by 0.5 ~ 1ms when concurrency " "increases by 4). This phenomenon is normal in most cases. However, sometimes" " TPOT may increase dramatically (10 to 100ms for example) as concurrency " "grows. This is possibly caused by " "[**PREEMPTION**](https://docs.vllm.ai/en/latest/configuration/optimization/#preemption)" " in vLLM. Generally, when your server hits KV cache limits, vLLM tries to " "free KV cache of requests to ensure sufficient space for other requests, " "which is called preemption in vLLM. When a request is preempted, the default" " behavior is to recompute the KV cache of this request again in the future, " "which is why the performance might drop significantly. There are several " "ways to verify this:" msgstr "" "测试 vLLM 服务器时，可能会发现 TPOT 随着并发增加而增加（例如，并发增加 4 时 TPOT 增加 0.5 ~ " "1ms）。这种现象在大多数情况下是正常的。然而，有时 TPOT 可能会随着并发增长而急剧增加（例如 10 到 100ms）。这可能是由 vLLM 中的 " "[**抢占**](https://docs.vllm.ai/en/latest/configuration/optimization/#preemption)" " 引起的。通常，当您的服务器达到 KV 缓存限制时，vLLM 会尝试释放请求的 KV 缓存，以确保为其他请求提供足够空间，这在 vLLM " "中称为抢占。当请求被抢占时，默认行为是将来重新计算该请求的 KV 缓存，这就是性能可能显著下降的原因。有几种方法可以验证这一点：" #: ../../source/faqs.md:290 msgid "" "vLLM usually logs stats on your server. You might see metrics like `GPU KV " "cache usage: 99.0%,`. When reaching 100%, it triggers preemption." msgstr "" "vLLM 通常会在您的服务器上记录统计信息。您可能会看到类似 `GPU KV cache usage: 99.0%,` 的指标。当达到 100% " "时，会触发抢占。" #: ../../source/faqs.md:291 msgid "" "When launching a vLLM server, you will see logs like `GPU KV cache size: " "66340 tokens` and `Maximum concurrency for 16,384 tokens per request: 4.05`." " These are estimated KV cache capacity for a single DP group. You can adjust" " the overall request traffic according to this." msgstr "" "启动 vLLM 服务器时，您会看到类似 `GPU KV cache size: 66340 tokens` 和 `Maximum concurrency" " for 16,384 tokens per request: 4.05` 的日志。这些是单个 DP 组的估计 KV " "缓存容量。您可以根据此调整整体请求流量。" #: ../../source/faqs.md:293 msgid "" "Preemption cannot be avoided completely since KV cache usage always has a " "limit. But there are methods to reduce the chances of preemption. As is " "suggested in " "[**PREEMPTION**](https://docs.vllm.ai/en/latest/configuration/optimization/#preemption)," " the core strategy is to increase available KV cache. For example, one can " "increase `--gpu-memory-utilization` or decrease `--max-num-seqs` && `--max-" "num-batched-tokens`." msgstr "" "由于 KV 缓存使用始终有限制，抢占无法完全避免。但有一些方法可以减少抢占的可能性。正如 " "[**抢占**](https://docs.vllm.ai/en/latest/configuration/optimization/#preemption)" " 中所建议的，核心策略是增加可用的 KV 缓存。例如，可以增加 `--gpu-memory-utilization` 或减少 `--max-num-" "seqs` 和 `--max-num-batched-tokens`。" #: ../../source/faqs.md:295 msgid "23. How do I choose between single-node and multi-node deployment?" msgstr "23.如何在单节点和多节点部署之间选择？" #: ../../source/faqs.md:297 msgid "" "Single-node deployment is recommended when the model fits within the memory " "of a single node's NPUs. For models like Qwen3-32B (BF16), which requires 4 " "× 64GB cards, multi-NPU within a single node (TP) is sufficient. Multi-node " "deployment is only needed when the total NPU count exceeds a single node's " "capacity." msgstr "" "当模型适合单节点NPU内存时，建议采用单节点部署。对于像Qwen3-32B（BF16）这样需要4×64GB卡的模型，单节点内的多NPU（TP）就足够了。只有当NPU总数超过单节点容量时，才需要多节点部署。" #: ../../source/faqs.md:299 msgid "24. What quantization method should I use?" msgstr "24.我应该使用哪种量化方法？" #: ../../source/faqs.md:301 msgid "" "**BF16**: Best accuracy, highest memory footprint. Use for accuracy-critical" " applications or when memory is sufficient." msgstr "**BF16**：最佳精度，最高内存占用。用于对精度要求高的应用或内存充足时。" #: ../../source/faqs.md:302 msgid "" "**W8A8**: Good balance of accuracy and memory reduction. Use for large " "models (e.g., 32B) on memory-constrained hardware." msgstr "**W8A8**：精度和内存减少的良好平衡。用于内存受限硬件上的大型模型（例如 32B）。" #: ../../source/faqs.md:303 msgid "" "**W4A8/W4A4**: Maximum memory reduction. Suitable for deploying larger " "models on smaller hardware configurations, with some accuracy trade-off." msgstr "**W4A8/W4A4**：最大程度减少内存占用。适用于在较小硬件配置上部署更大模型，但会牺牲一定精度。" #: ../../source/faqs.md:305 msgid "25. When should I enable FlashComm_v1?" msgstr "25.何时应启用 FlashComm_v1？" #: ../../source/faqs.md:307 msgid "" "Enable FlashComm_v1 (`VLLM_ASCEND_ENABLE_FLASHCOMM1=1`) when using Tensor " "Parallelism (TP ≥ 2) with high concurrency. It is threshold-protected and " "will not activate in low-concurrency scenarios where it could degrade " "performance." msgstr "" "在使用张量并行（TP ≥ 2）且高并发时启用 " "FlashComm_v1（`VLLM_ASCEND_ENABLE_FLASHCOMM1=1`）。该功能受阈值保护，在可能降低性能的低并发场景下不会激活。" #: ../../source/faqs.md:309 msgid "26. What is the difference between FIA and PA operators for attention?" msgstr "26.FIA 和 PA 注意力算子有何区别？" #: ../../source/faqs.md:311 msgid "" "FIA (Flash Attention) is the default attention operator in vLLM-Ascend. In " "some batch-size settings (particularly medium concurrency), FIA may exhibit " "suboptimal performance. The PA (Page Attention) operator can be manually " "enabled via `pa_shape_list` in `--additional-config`. When the runtime batch" " size matches a value in `pa_shape_list`, the framework switches to PA. This" " is a temporary tuning knob — future FIA optimizations will make this " "parameter obsolete." msgstr "" "FIA（Flash Attention）是 vLLM-Ascend 中的默认注意力算子。在某些批次大小设置（尤其是中等并发）下，FIA " "可能表现欠佳。可通过 `--additional-config` 中的 `pa_shape_list` 手动启用 PA（Page " "Attention）算子。当运行时批次大小与 `pa_shape_list` 中的某个值匹配时，框架会切换到 PA。这是一个临时调优参数——未来 FIA" " 的优化将使该参数失效。"