vLLM-HUST Engine R9
R9 包含完整冻结 SDK、原生库、MOD、逐文件校验清单和固定评测工具。它在 R8 执行核心上增加显式的 Intern-S2 35B 适配和状态页复制修复;Qwen 保留原 Worker 与原词表准入路径。
支持环境为 CPython 3.12、Linux aarch64、双 Ascend 910B2、CANN 9.1.0,以及固定的 Torch 2.13.0+cpu、Torch-NPU 2.13.0rc1、Triton-Ascend 3.6.0、Transformers 5.14.1。驱动、供应商 Python 环境和模型权重需预先准备。供应商依赖元数据中的 NumPy 要求仍有矛盾,因此安装清单使用 --no-deps 保留已验证环境,不声称自动安装全部依赖。
在独立环境安装固定的四个包,保留旧环境以便回退:
python -m pip install --no-cache-dir --force-reinstall --no-deps -r https://github.com/vLLM-HUST/vllm-hust-website/releases/download/engine-wheels-r9-v1/r9-pip-requirements-post1.txt
vllm-hust-engine verify
vllm-hust-engine install --out /data/hust-r9-install
vllm-hust-engine serve --installation /data/hust-r9-install -- \
--model /models/Intern-S2-35B --out /data/r9-serve-001 \
--cache /tmp/r9-compiler-001 --physical-devices 0,7 --devices 0,1 --port 8000
模型选择以 config.json 的架构、层数、词表、MoE 与注意力参数为准,不根据模型目录名称猜测。书生使用 hust_intern_s2.worker.Worker;Qwen 使用原来的 betterscale.qwen35_worker.Worker。两者均采用 BF16、TP2、MTP2、FULL、E36/R54、每 rank 4 GiB 主机状态和 20 GiB FA KV 预算。默认服务仅绑定本机地址。
普通 vllm serve 不会自动启用这个冻结配方。使用上述 Engine 服务入口,或明确使用已校验的部署和工具。安装和校验命令不会启动模型。运行记录保存在 --out;热路径记录写入独立本地临时目录,退出后归档,具体路径写入 runtime-storage.json。
evaluate --installation ... -- --engine-profile r9 ... 保留固定五项 full / MMLU200 quick 语义及原始 Qwen 对照契约。书生服务和项目联调的模型间测试独立记录;没有同模型 B0、参考分母和冻结任务结果时,不宣称书生完成该项目验收。R9 发行与切换默认版本分别处理,当前默认仍按评测入口契约执行。
性能和质量按实际记录披露;完整模型吞吐、首 token 延迟、尾延迟、任务得分及项目最终目标分别判断。回退时使用保留的 R8 环境及部署,不改写旧 Release 或旧记录。
Metadata
Release files for vllm-hust-engine 2026.11.19.9.post1
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| vllm_hust_engine-2026.11.19.9.post1.tar.gz | 117.7 MB | Without provenance | Details for vllm_hust_engine-2026.11.19.9.post1.tar.gz |
Release files / vllm_hust_engine-2026.11.19.9.post1.tar.gz
| Download URL | vllm_hust_engine-2026.11.19.9.post1.tar.gz |
|---|---|
| Size | 117.7 MB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
efbc0cd0173849ab5e10df26afa34bde4ca1a4b4a59ba04bd36b7c57d263f0de
|
|
BLAKE2b-256 checksum How to use checksums |
87686d5c56c82298482a6cba4434cf7497f5b511a713b04a3ad159649b0f8818
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.2.0 CPython/3.12.10
|
Security for vllm-hust-engine 2026.11.19.9.post1
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.