Metadata-Version: 2.4
Name: vllm-hust-engine
Version: 2026.11.17.8.post2
Summary: vLLM-HUST Engine frozen B1 r8 runtime for Ascend 910B2
Author: vLLM-HUST
License-Expression: Apache-2.0
Project-URL: Homepage, https://vllm-hust.sage.org.ai/engine.html
Project-URL: Repository, https://github.com/vLLM-HUST/vllm-hust-engine
Project-URL: Frozen release, https://github.com/vLLM-HUST/vllm-hust-engine/releases/tag/full-answer-b1-20261117-r8
Classifier: Development Status :: 4 - Beta
Classifier: Operating System :: POSIX :: Linux
Classifier: Programming Language :: Python :: 3.12
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: <3.13,>=3.12
Description-Content-Type: text/markdown
License-File: LICENSE
License-File: NOTICE
Requires-Dist: vllm==0.28.1.post1.dev143+gf18cf803c.empty
Requires-Dist: vllm-ascend==0.25.1rc2.dev125+hust.20260903.4.g74f0c0a27
Requires-Dist: vllm-hust-runtime==2026.11.17.8+r8.3f7eacc1
Dynamic: license-file

# vLLM-HUST Engine r8 · packaging revision2

This is the same frozen r8 runtime and all9235 checksum-pinned files. post2 changes package installation, not engine kernels, MODs, evaluation inputs or measured results.

In the qualified CPython3.12 / Linux aarch64 / CANN9.1.0 environment:

```sh
python -m pip install --no-deps -r https://github.com/vLLM-HUST/vllm-hust-website/releases/download/engine-wheels-r8-v1/r8-pip-requirements-post2.txt
```

The fixed requirements file explicitly lists this Engine package, the same-name vllm and vllm-ascend wheels, and the runtime support wheel, with SHA256 verification for all four. One pip operation replaces older official distributions while retaining import vllm, import vllm_ascend and the vllm CLI. --no-deps retains the externally qualified vendor environment: upstream Triton-Ascend and vLLM dependency metadata contain conflicting NumPy requirements. No dependency metadata is removed or falsified, and no separate downloader or installation script is needed. The complete original SDK and fixed evaluation tools remain included. Native artifacts are not recompiled or retagged as manylinux.

Qualified dependencies: Torch2.13.0+cpu, Torch-NPU2.13.0rc1, Triton-Ascend3.6.0, Transformers5.14.1 and tokenizers0.22.2. CANN, drivers and model weights remain external. The qualified vendor environment is a prerequisite; this command installs the four pinned engine distributions only, not Torch, Triton-Ascend, CANN or model weights. Unsupported ABI causes pip to fail. post1 was an unpublished candidate rejected during full dependency resolution; use post2 with this complete four-package manifest.

The Qwen3.5-35B-A3B BF16 / TP2 / dual Ascend910B2 MOD recipe (CorrectedWorker, BetterScale, Microflow, KV materialization, M01 correction, packed-host, sealed warmup, guarded async progress, E36/R54 and short-prefill tail; MTP2/FULL, host4GiB/rank and KV20GiB/rank) uses the frozen r8 launcher. Generic vllm serve does not select this model-specific recipe automatically.

SDK verification and evaluation remain available through vllm-hust-engine verify / install / evaluate. Installation does not start a model or create a new performance result.

[Installation and MOD details](https://vllm-hust.sage.org.ai/engine.html#install-runtime) · [Frozen engine](https://github.com/vLLM-HUST/vllm-hust-engine/releases/tag/full-answer-b1-20261117-r8)
