910B vllm-ascend服务化调用超时“Timed out while polling task status for task”

[18861870889]

🔎 Search before asking | 提交之前请先搜索

• I have searched the MinerU Readme and found no similar bug report.
• I have searched the MinerU Issues and found no similar bug report.
• I have searched the MinerU Discussions and found no similar bug report.

🤖 Consult the online AI assistant for assistance | 在线 AI 助手咨询

• I have consulted the online AI assistant but was unable to obtain a solution to the issue.

Description of the bug | 错误描述

设备：910B，HDK 25.2.0
镜像版本：vllm-ascend 0.18.0rc1：https://quay.io/ascend/vllm-ascend可下载
dockerfile: https://github.com/opendatalab/MinerU/blob/master/docker/china/npu.Dockerfile
backend：vllm 服务化
调用方法：参考mineru -p <input_path> -o <output_path> -b hybrid-http-client -u http://127.0.0.1:30000
错误日志：
INFO: Started server process [12446]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://127.0.0.1:39727 (Press CTRL+C to quit)
2026-04-15 08:59:01.568 | INFO | mineru.cli.client:run_planned_task:771 - Submitting batch 1/1 | 1 document, 14 pages in this batch | 14 pages total | task#1 [qwen3_embedding_technical_report]
2026-04-15 08:59:05.926 | INFO | mineru.backend.vlm.vlm_analyze:get_model:248 - get http-client predictor cost: 0.13s
2026-04-15 08:59:06.087 | INFO | mineru.backend.hybrid.hybrid_analyze:aio_doc_analyze:702 - Hybrid processing-window run. page_count=14, window_size=64, total_windows=1
2026-04-15 08:59:06.993 | INFO | mineru.backend.hybrid.hybrid_analyze:get_batch_ratio:511 - hybrid batch ratio (auto, vram=61GB): 16
onnxruntime cpuid_info warning: Unknown CPU vendor. cpuinfo_vendor value: 15
2026-04-15 08:59:18.547 | INFO | mineru.backend.hybrid.hybrid_analyze:aio_doc_analyze:724 - Hybrid processing window 1/1: pages 1-14/14 (14 pages)
Two Step Extraction: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 14/14 [00:16<00:00, 1.18s/it]
Layout Predict: 0%| | 0/14 [00:00<?, ?it/s]2026-04-15 09:00:35.646 | WARNING | mineru.cli.api_client:wait_for_task_result:638 - Timed out while polling task status for task#1 [qwen3_embedding_technical_report] (task_id=cdfc458d-1281-4d8c-8201-c6fe5d3c797b). This can happen during cold start; retrying until the task deadline.
2026-04-15 09:01:36.696 | WARNING | mineru.cli.api_client:wait_for_task_result:638 - Timed out while polling task status for task#1 [qwen3_embedding_technical_report] (task_id=cdfc458d-1281-4d8c-8201-c6fe5d3c797b). This can happen during cold start; retrying until the task deadline.
2026-04-15 09:02:37.702 | WARNING | mineru.cli.api_client:wait_for_task_result:638 - Timed out while polling task status for task#1 [qwen3_embedding_technical_report] (task_id=cdfc458d-1281-4d8c-8201-c6fe5d3c797b). This can happen during cold start; retrying until the task deadline.

How to reproduce the bug | 如何复现

按照错误描述中的内容进行复现即可

Operating System Mode | 操作系统类型

Linux

Operating System Version| 操作系统版本

Linux syn-140 5.10.0-153.56.0.134.oe2203sp2.aarch64 #1 SMP Wed May 29 14:08:32 CST 2024 aarch64 aarch64 aarch64 GNU/Linux

Python version | Python 版本

3.11

Software version | 软件版本 (mineru --version)

>=2.5

Backend name | 解析后端

vlm

Device mode | 设备模式

npu

[dosubot]

Hi @18861870889! I'm Dosu and I’m helping the MinerU team.

这个问题已有多人反馈（包括今天的 #4792）。根据日志，你的处理流程卡在了 Layout Predict 0% 的位置，此时每隔 60 秒的 HTTP 状态轮询请求超时，触发了这个 warning [1]。

核心问题：vllm 后端不支持 Ascend NPU 硬件 [2]。hybrid-http-client 依赖 vllm 服务进行推理，而 vllm（包括 vllm-ascend）在 MinerU 的 NPU 场景下存在兼容性问题，会导致推理请求挂起或超时。

建议的解决方案：

1. 切换后端为 pipeline 或 vlm-transformers，这两个后端支持 NPU：

   mineru -p <input_path> -o <output_path> -b pipeline --device npu

2. 检查 PyTorch 版本：确认安装的不是 CPU-only 版本（版本号不带 +cpu 后缀），且 torch_npu 和 CANN 驱动版本匹配 [3]。

3. NPU 首次推理需要 warmup：在 910B 上 DocAnalysis 模型初始化可能需要很长时间（曾有报告需要约 1700 秒），这是正常现象。加速效果在第二轮推理后才会体现 [3]。

4. 增加启动超时环境变量：

   export MINERU_LOCAL_API_STARTUP_TIMEOUT_SECONDS=600
   export MINERU_PDF_RENDER_TIMEOUT=600

另外，任务轮询总超时为 3600 秒（1 小时），这个值是硬编码的，无法通过 CLI 配置 [4]。但根本问题不是超时时间不够，而是后端架构不匹配。

维护者 @myhloli 在类似问题中也提到了需要确认模型源配置和推理引擎的选择 [5]。建议先尝试切换到 pipeline 后端看看能否正常工作。

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}