vLLM Ascend KV Compression

English | 简体中文

An independently packaged TriAttention KV-cache compression plugin for the upstream-aligned vLLM-HUST and vLLM-Ascend-HUST stacks. Version 0.4 adapts the current hosts and Extension Manager without changing either host repository.

Status: experimental release candidate. Package lifecycle, Ascend kernel smoke, service correctness, and three cold 16K engineering-control pairs pass. The full V4.6 official-baseline, model-matched calibration, quality, and stability matrix remains a release gate. See Validation.

Ownership and maintenance

• School: Huazhong University of Science and Technology (HUST)
• Group: CGCL
• Advisor: Prof. Yao Wan (万瑶)
• Project lead: Sichen Liu (刘思辰), @Seas0
• Maintainers: Jiawan Zhang (张家万), @Jiawan23; Ruohao Wei (韦若皓), @kotoriqaq0; and @Seas0

The team agrees to maintain compatibility with vLLM-HUST and vLLM-Ascend-HUST and to publish this work through the vLLM-HUST Extension Manager. This is a CGCL-maintained plugin, not code built into either host.

Source and redistribution scope

The scoring method adapts TriAttention commit a4bc3c8f and the paper TriAttention: Efficient Long Reasoning with Trigonometric KV Compression. The Ascend paged-KV runtime was rewritten and does not copy the reference CUDA kernels.

Repository code and documentation are Apache-2.0. Model weights, datasets, raw logs, and calibration statistics are not included in the wheel or sdist. The development-only committed statistics have incomplete model/data provenance and must not be redistributed merely under this repository's license. See NOTICE, ownership and licensing, and calibration artifacts.

Compatibility

The supported topology is one Ascend NPU, the v1 scheduler and NPUModelRunner, one full-attention KV group, block size 128, and dense BF16/FP16 K/V. Unsupported combinations fail during startup. The manifest uses the stable vllm.general_plugins discovery entry point; current hosts do not expose a frozen native KV-lifecycle API, so the narrow host range and contract tests intentionally guard the internal integration seams.

Install, enable, disable, and uninstall

Install the host stack first, then the released wheel:

python -m pip install 'vllm-ascend-kvcompress-hust[manager]==0.4.0'

Copy examples/triattention.json, set stats_path to model- and revision-matched statistics, then run:

vllm-hust-ext extension validate org.vllm-hust.ascend-kvcompress
vllm-hust-ext extension configure org.vllm-hust.ascend-kvcompress \
  --file /absolute/path/triattention.json
vllm-hust-ext extension enable org.vllm-hust.ascend-kvcompress
vllm-hust-ext extension status org.vllm-hust.ascend-kvcompress

export VLLM_PLUGINS=ascend,ascend_kvcompress
vllm-hust-ext run -- vllm serve /path/to/model \
  --block-size 128 --no-enable-prefix-caching --no-async-scheduling

If VLLM_PLUGINS is unset, vLLM discovers all installed plugins. The package remains inert unless the manager or direct enable flag activates it. Stop all host processes before changing state:

vllm-hust-ext extension disable org.vllm-hust.ascend-kvcompress
vllm-hust-ext extension forget org.vllm-hust.ascend-kvcompress
python -m pip uninstall vllm-ascend-kvcompress-hust

These operations only change the plugin package and manager state. Detailed source/environment instructions are in the installation guide, and the isolated wheel workflow is in packaging and release.

For development without Extension Manager:

export VLLM_PLUGINS=ascend,ascend_kvcompress
export VLLM_ASCEND_KVCOMPRESS_ENABLED=1
export VLLM_ASCEND_KVCOMPRESS_CONFIG=/absolute/path/triattention.json
vllm serve /path/to/model --block-size 128 --no-enable-prefix-caching

Unset both VLLM_ASCEND_KVCOMPRESS_* variables and restart to disable it.

Runtime and long-context optimization

After explicit activation, the adapter validates and hooks the current scheduler, KV-cache manager, concrete Ascend block table, and NPUModelRunner. Compression is committed only after a synchronous model step: semantic RoPE positions stay unchanged, physical attention/slot indices use a per-request offset, and old blocks are released at the next scheduler barrier.

Version 0.4 uses direct paged-K scoring, persistent workspaces, fused NPU normalization/head/layer aggregation, dynamic JIT lengths, layer-stride sampling, device-resident offsets, and a tuned 4096-token physical budget. The final 910B2 kernel benchmark measured 1.82x paged-copy, 2.81x direct-score, and 1.34x aggregate speedups against generic references; offset update was 0.83x and is not claimed as an improvement.

In three cold engineering-control pairs on Qwen2.5-14B, each with four fixed 16,384+1,024 requests at 0.4 RPS/concurrency 4, all 24 requests completed with the expected output length. Compression retained 32/128 blocks per request (75% physical KV reduction). Median total throughput was 1431.3 versus 1129.5 tok/s (+26.7%); mean TPOT was 39.16 versus 52.43 ms (-25.3%); mean E2E was 44.37 versus 57.81 s (-23.2%). This is engineering evidence against a same-host compatibility control, not the required official V4.6 B0 claim.

Conflict matrix

Configuration and validation

The example uses a 4096-token budget, 1024-token recompute window, 512-token protected recent window, 8192-token scoring chunks, and every fourth scoring layer. Token counts must be positive multiples of block size 128.

• Current validation record
• V4.6-derived requirements
• Benchmark protocol
• Packaging and release
• Method extension API
• Calibration artifacts

VLLM_PLUGINS='' TORCH_DEVICE_BACKEND_AUTOLOAD=0 python -m pytest -q
python -m ruff check src tests scripts
python -m ruff format --check src tests scripts

An NPU release candidate must also pass numerical kernel smoke and the declared long-context quality/performance/HBM matrix on the exact host stack.