Metadata-Version: 2.4
Name: vllm-hust-engine
Version: 2026.11.17.8
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
Dynamic: license-file

# vLLM-HUST Engine — frozen B1 r8

This package contains the complete **2026-11-17 B1 r8 runtime**: 9,235 checksum-pinned files, original runtime/native artifacts, MODs, provenance, and the fixed evaluation tools. It does not substitute the latest upstream vLLM or download private GitHub assets at installation time.

Frozen composition: CorrectedWorker; BetterScale; Microflow; KV materialization; M01 correction; packed host state; sealed serving warmup; guarded asynchronous progress H2D; E36/R54 capacity; eligible prefill-tail1–256 priority. MTP2, FULL graphs, host State4GiB/rank, KV20GiB/rank, prefill4096, TP2 Ascend910B2. Fusion experiments are excluded.

## Runtime requirements

The included native binaries require **Linux aarch64, CPython3.12, CANN9.1.0 and the fixed B1 Python environment**. The frozen dependency contract is checked before admitting an installation. In particular, the measured environment uses Torch2.13.0+cpu, torch-npu2.13.0rc1, triton-ascend3.6.0, Transformers5.14.1 and tokenizers0.22.2. CANN, accelerator drivers, model weights and the B0 baseline are external requirements. Installing this package does not replace system Torch/CANN or load a model.

```sh
python -m pip install vllm-hust-engine==2026.11.17.8
vllm-hust-engine verify
vllm-hust-engine install --python /absolute/path/to/fixed-b1-python --out ./frozen-r8
vllm-hust-engine evaluate --installation ./frozen-r8 -- \
  --model /absolute/path/to/Qwen3.5-35B-A3B \
  --baseline-deployment /absolute/path/to/original018 \
  --physical-devices 0,1 --out ./results/r8-full
```

`evaluate` uses the exact frozen one-click controller and automatically follows logs. Its default is all five fixed profiles: MMLU200, HLE stratified200 (64/10/113/13), Frontier160, SWE public trace32 tasks, Terminal public trace32 tasks. Add `--datasets mmlu` for the quick profile. Reference denominator verification is enabled by default. HLE/Frontier auxiliary judging uses the same model in separate unmeasured stages; SWE/Terminal are inference trace replay, not task-solving scores.

The r8 runtime was measured in C15 and the new B0/r8 campaign. This package's deployment relocation is checked on CPU without claiming a new NPU result or replacing the running campaign. Historical results and current campaign status remain in the original release.

[Original frozen Release and evidence](https://github.com/vLLM-HUST/vllm-hust-engine/releases/tag/full-answer-b1-20261117-r8) · [Engine documentation](https://vllm-hust.sage.org.ai/engine.html)

PyPI version `2026.11.17.8` maps to frozen engine r8. Its bundled evaluation tools are revision r4 (`bd5448c9f266cc0301c3d4cef2a10b9ab98f7205`), including SDK hash-cache release before startup memory admission. The 9,235 frozen runtime files are unchanged.
