Sparse-vLLM

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

★ 65 Open GitHub ↗

A sparse-first inference engine (sparsevllm). It also contains DeltaKV compressor training + evaluation tooling (deltakv).

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<p align="center">English | <a href="README_zh.md">简体中文</a></p>

A sparse-first inference engine for long-context LLM serving.

## Project Overview

Sparse-vLLM is an inference framework built with sparsity as the first design principle. Instead of layering sparse methods on top of a conventional KV cache, it rethinks cache layout, controller flow, and kernels so that multiple sparse mechanisms can plug in cleanly.

> **Note:** DeltaKV compressor training code is maintained separately in
> [CURRENTF/DeltaKV](https://github.com/CURRENTF/DeltaKV). This repository only
> keeps the native DeltaKV inference implementation under `src/sparsevllm/`;
> it does not include DeltaKV training code or an HF reference implementation.

## Key Runtime Principles

- Public commands and `LLM(...)` kwargs should use `sparse_method`; Sparse-vLLM
  normalizes it internally to `vllm_sparse_method`.
- Sparse method runtime state belongs in
  `src/sparsevllm/engine/cache_manager/`; `attention.py` should stay generic.
- Prefill scheduling is method-specific and registry-owned. The source of
  truth is `src/sparsevllm/method_registry.py`, not benchmark scripts.
- Sparse-vLLM currently uses two prefill policies: `all_chunked` and the
  special `long_bs1full_short_batch` policy.
- `long_bs1full_short_batch` is only for methods that are registered to need a
  complete long-prefill pass before their sparse/cache transformation. Long
  requests run as full prefill with batch size 1; short requests still use
  chunked batching.
- Benchmark reports should record the sparse method, prefill policy, prefill
  chunk size, prompt length, batch size, and any DeltaKV checkpoint.

## Core Sparse Methods

Sparse-vLLM supports physical eviction, logical masking, query-aware selection,
and hybrid KV compression. The main method families are `streamingllm`,
`snapkv`, `h2o`, `pyramidkv`, `omnikv`, `quest`, and `deltakv`.

| Method | Type | Short Description |
| --- | --- | --- |
| `vanilla` | Dense baseline | Runs full attention and keeps the standard KV cache behavior for correctness and performance baselines. |
| `streamingllm` / `attention-sink` | Physical eviction | Keeps fixed sink tokens plus a recent window, then physically evicts older tokens outside that policy. |
| `snapkv`, `pyramidkv` | Physical eviction | Selects important historical tokens during prefill/finalization and stores only the retained KV tokens. |
| `h2o` | Physical eviction | Accumulates normalized token importance in one score vector per layer/sequence, retaining heavy hitters plus a recent suffix after every prefill chunk and decode step. Prefill uses normalized attention mass; the optimized decode path max-reduces raw QK logits across query heads before token-wise normalization. The total intermediate/final budgets are `h2o_prefill_budget` and `h2o_decode_budget`; `h2o_recent_ratio` splits each budget and `h2o_prefill_score_window` controls chunk scoring. |
| `omnikv` | Logical masking | Keeps tokens in storage but masks the attention read view so sparse layers attend only selected context. |
| `quest` | Query-aware selection | Uses decode-time query-aware page selection while keeping prefill dense. |
| `deltakv` / `deltakv-*` | Hybrid compression | Keeps a small full-precision pool and stores older context through DeltaKV compression or related ablations. |

Read the method overview and integration rules in
[Core Sparse Methods](docs/en/features/sparse-methods.md).

## Supported Models

| Model | Supported |
| --- | :---: |
| Qwen2.5 | ✅ |
| Qwen3 | ✅ |
| Qwen3MoE | ✅ |
| Qwen3.5 / Qwen3.6 | ✅ |
| Llama 3 / 3.1 | ✅ |
| MiniMax M2.7 | ✅ |

See [Supported Models](docs/en/features/supported-models.md) for the precision,
parallelism, and sparse-method compatibility matrices.

## Documentation

| Topic | Link |
| --- | --- |
| Quick setup and minimal usage | [Getting Started](docs/en/getting_started/README.md) |
| Model, precision, and parallelism support | [Supported Models](docs/en/features/supported-models.md) |
| Sparse method taxonomy and extension rules | [Core Sparse Methods](docs/en/features/README.md) |
| Runtime architecture | [Architecture](docs/en/design/README.md) |
| Runtime parameter semantics | [Runtime Parameter Semantics](docs/en/configuration/runtime-parameter-semantics.md) |
| Benchmark commands | [Benchmarks](docs/en/benchmarking/README.md) |
| DeltaKV inference | [DeltaKV](docs/en/features/deltakv.md) |
| Reproducibility checklist | [Reproducibility](docs/en/getting_started/reproducibility.md) |

The full documentation index is maintained in [docs/en/README.md](docs/en/README.md).

## Quick Start

Sparse-vLLM requires Python 3.10 or newer. Install the package from the
repository root using the runtime versions pinned in `pyproject.toml`.

### Conda

```bash
conda create -n svllm python=3.10 -y
conda activate svllm

pip install torch==2.11.0 torchvision==0.26.0 triton==3.6.0 \
  --index-url https://download.pytorch.org/whl/cu130

# FlashInfer publishes the CUDA-specific JIT cache on a separate index.
pip install "flashinfer-jit-cache>=0.6.15" \
  --index-url https://flashinfer.ai/whl/cu130

pip install -e .

# Optional
MAX_JOBS=8 pip install flash-attn --no-build-isolation
pip install flashinfer-cubin --index-url https://flashinfer.ai/whl
```

PyTorch wheels include their CUDA runtime, while compiled extensions such as
`flash-attn` use the CUDA toolchain active in the environment.

### uv

```bash
uv venv --python 3.10
source .venv/bin/activate

uv pip install torch==2.11.0 torchvision==0.26.0 triton==3.6.0 \
  --index-url https://download.pytorch.org/whl/cu130
uv pip install "flashinfer-jit-cache>=0.6.15" \
  --index-url https://flashinfer.ai/whl/cu130

uv pip install -e .

# Optional
MAX_JOBS=8 uv pip install flash-attn --no-build-isolation
uv pip install flashinfer-cubin --index-url https://flashinfer.ai/whl
```

The explicit indexes select the CUDA 13.0 builds of PyTorch and the FlashInfer
JIT cache.

Qwen3.5/Qwen3.6 FP8 mixed-attention inference additionally requires the
optional Python dependencies:

```bash
# uv
uv pip install -e ".[qwen35]"

# Conda/pip
pip install -e ".[qwen35]"
```

Vanilla, OmniKV, or QuEST prefix-cache offload without Qwen3.5/Qwen3.6 uses
the smaller CUDA-specific extra:

```bash
pip install -e ".[prefix-offload]"
```

Sparse-vLLM currently supports Qwen3.5/Qwen3.6 checkpoints only in the
block-scaled FP8 format.

The Qwen3.5/Qwen3.6 prefill causal Conv1D and decode Conv1D/GDN packing paths
use repository-local Triton kernels; `sglang-kernel` and a local CUDA extension
build are not required.

For the full dependency list and a minimal `LLM(...)` example, see
[Getting Started](docs/en/getting_started/README.md).

## Benchmarks

Use `scripts/benchmarks/bench_sparse_vllm.py` for throughput measurements and
the `benchmark/` entrypoints for LongBench, MathBench, SCBench, NIAH, and
multimodal evaluations.

See [Benchmarks](docs/en/benchmarking/README.md) for command examples and backend notes.

## Contributing Sparse Methods

New sparse methods should keep method-specific runtime state in
`src/sparsevllm/engine/cache_manager/` and keep
`src/sparsevllm/layers/attention.py` generic.

## Acknowledgements

This project is inspired by and/or references ideas and implementation techniques from:

- `LightLLM` (`ModelTC/LightLLM`)
- `SGLang` (`sgl-project/sglang`)
- `ShadowKV` (`ByteDance-Seed/ShadowKV`)
- `nano-vllm` (`GeeeekExplorer/nano-vllm`)

## License

[Apache License 2.0](LICENSE)

## Citation
```text
@article{hao2026deltakv,
  title={DeltaKV: Residual-Based KV Cache Compression via Long-Range Similarity},
  author={Hao, Jitai and Huang, Qiang and Wang, Yaowei and Zhang, Min and Yu, Jun},
  journal={arXiv preprint arXiv:2602.08005},
  year={2026}
}

@inproceedings{hao2025omnikv,
  title={Omnikv: Dynamic context selection for efficient long-context llms},
  author={Hao, Jitai and Zhu, Yuke and Wang, Tian and Yu, Jun and Xin, Xin and Zheng, Bo and Ren, Zhaochun and Guo, Sheng},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025}
}
```

Sparse-vLLM

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README