Default-off Ascend attention-path optimization MOD family for vLLM-HUST. The current package is a qualification gate: it is inert by default and rejects all recovered legacy selectors until a current-Host implementation exists.
Shared-prefix attention, FIA/FA execution, RoPE and attention-path optimization experiments.
This repository is the canonical integration and evidence home for the family. Every optimization is opt-in and disabled by default until its own compatibility and performance qualification is complete.
| Legacy selector | Disposition | Reason |
|---|---|---|
core-boundary |
Retired | Current Core and Ascend implement their own evolved decode/prefill boundary logic |
persistent-matmul-empty |
Retired | The original formal environment never consumed this path |
linear-swiglu-graph |
Relocated | This is an operator/fusion concern and must not be duplicated in the attention family |
pyramidkv |
Port required | It depends on the removed native_model_runner_v1.py lifecycle and old provider build |
An empty VLLM_HUST_ATTENTION_OPTIMIZATIONS value is a no-op. Supplying any
legacy selector fails closed with its disposition. This prevents an old wheel
or deployment preset from silently patching a changed Host.
The audit is pinned to the Core commit verified by the Ascend Host,
c4df37dfcc7f72ca41cd43801a468526ecf99a2b,
and Ascend 7602797593170da4c531f1d8882a7b82dce8a1bc:
python tools/check_host_contract.py /path/to/vllm-hust /path/to/vllm-ascend-hust
pytest -qThe family remains retest_pending. No legacy performance result is eligible
for the website until a new mechanism passes current-Host NPU qualification.
- Maintainer: ShuhaoZhangTony
- Advisors: none recorded
- Organization: vLLM-HUST
Measured single-MOD, composition, and compatibility results are published through vllm-hust-benchmark and presented with the corresponding MOD in the MOD catalog.