vLLM for Baidu Kunlun

---

Latest News 🔥

• [2026/07] 🚧 v0.25.1 under development — Added Qwen3.5 / Qwen3.5-MoE, Gemma4 (text and multimodal), GLM MoE DSA, and DFlash speculative decoding

• [2026/02] ⚡ Performance optimizations — Fused MoE with small batches, optimized attention metadata building, Multi-LoRA inference achieves 80%+ of non-LoRA performance

• [2026/02] 🔧 DeepSeek-V3.2 MTP support — Added MTP (Multi-Token Prediction) for DeepSeek-V3.2, with RoPE and decoding stage kernel optimizations

• [2026/01] 🔢 New quantization methods — Support for compressed-tensors W4A16, AWQ MoE W4A16, and DeepSeek-V3.2 W8A8 quantization

• [2026/01] 🛠️ CI/CD overhaul — Added E2E tests, unit test CI, ruff format checks, and modular CI workflow refactoring

• [2025/12] 🎉 v0.11.0 released — Added Qwen3-Omni, Qwen3-Next, Seed-OSS support (Release Notes)

• [2025/12] 📦 v0.10.1.1 released — 5+ multimodal models, AWQ/GPTQ quantization for dense models, Piecewise Kunlun Graph, vLLM V1 engine, Flash-Infer Top-K/Top-P sampling with 10-100× speedup (Release Notes)

• [2025/12] 🌟 Initial release of vLLM Kunlun — Open sourced on Dec 8, 2025

---

Overview

vLLM Kunlun (vllm-kunlun) is a community-maintained hardware plugin designed to seamlessly run vLLM on the Kunlun XPU. It is the recommended approach for integrating the Kunlun backend within the vLLM community, adhering to the principles outlined in the RFC Hardware Pluggable.

This plugin provides a hardware-pluggable interface that decouples the integration of the Kunlun XPU with vLLM. By utilizing vLLM Kunlun, popular open-source models — including Transformer-like, Mixture-of-Expert (MoE), Embedding, and Multi-modal LLMs — can run effortlessly on the Kunlun XPU.

✨ Key Features

• Seamless Plugin Integration — Works as a standard vLLM platform plugin via Python entry points, no need to modify vLLM source code

• Broad Model Support — Supports 20+ mainstream LLMs including Qwen, Llama, DeepSeek, GLM, Gemma4, Kimi-K2, and multimodal models

• Quantization Support — W8A8 (INT8), AWQ, GPTQ, and compressed-tensors W4A16 for MoE and dense models

• LoRA Fine-Tuning — LoRA and Multi-LoRA adapter support for Qwen series models

• Piecewise Kunlun Graph — Hardware-accelerated graph optimization for high-performance inference

• FlashMLA Attention — Optimized multi-head latent attention for DeepSeek MLA architectures

• Speculative Decoding — MTP (Multi-Token Prediction) for DeepSeek-V3.2 and DFlash/EAGLE-style proposers

• Tensor Parallelism — Multi-device parallel inference with distributed execution support

• OpenAI-Compatible API — Serve models with the standard OpenAI API interface

---

Prerequisites

• Hardware: Kunlun3 P800

• OS: Ubuntu 20.04

• Software:

  • Python >= 3.10

  • PyTorch >= 2.5.1 (KL3-customized xpytorch build, see Installation)

  • vLLM (matching version, see requirements.txt)

  • transformers == 5.2.0 (Qwen3.5 requires transformers 5.x)

---

Supported Models

Generative Models

Model	Support	Quantization	LoRA	Kunlun Graph
Qwen2	✅	✅	✅	✅
Qwen2.5	✅	✅	✅	✅
Qwen3	✅	✅	✅	✅
Qwen3-Moe	✅	✅	✅	✅
Qwen3-Next	✅	✅	✅	✅
Qwen3.5	✅	✅		✅
Qwen3.5-Moe	✅	✅		✅
MiMo-V2-Flash	✅	✅		✅
Llama2	✅	✅	✅	✅
Llama3	✅	✅	✅	✅
Llama3.1	✅	✅		✅
gpt-oss	✅	✅
Gemma4	✅			✅
GLM4.5	✅	✅		✅
GLM4.5Air	✅	✅		✅
GLM5	✅	✅		✅
InternLM2	✅			✅
Seed-OSS	✅			✅
DeepSeek-R1	✅	✅		✅
DeepSeek-V3	✅	✅		✅
DeepSeek-V3.2	✅	✅		✅
Kimi-K2	✅	✅		✅

Multimodal Language Models

Model	Support	Quantization	LoRA	Kunlun Graph
Qwen2-VL	✅	✅		✅
Qwen2.5-VL	✅	✅		✅
Qwen3-VL	✅	✅		✅
Qwen3-VL-MoE	✅	✅		✅
Gemma4	✅			✅
InternVL-2.5	✅			✅
InternVL-3.5	✅			✅
InternS1	✅			✅

---

Performance Visualization 🚀

High-performance computing at work: How different models perform on the Kunlun3 P800.

Current environment: 16-way concurrency, input/output size 2048.

---

Quick Start

Start an OpenAI-Compatible API Server

python -m vllm.entrypoints.openai.api_server \
    --host 0.0.0.0 \
    --port 8356 \
    --model <your-model-path> \
    --gpu-memory-utilization 0.9 \
    --trust-remote-code \
    --max-model-len 32768 \
    --tensor-parallel-size 1 \
    --dtype float16 \
    --max_num_seqs 128 \
    --max_num_batched_tokens 32768 \
    --block-size 128 \
    --no-enable-prefix-caching \
    --no-enable-chunked-prefill \
    --distributed-executor-backend mp \
    --served-model-name <your-model-name>

Send a Request

curl http://localhost:8356/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "<your-model-name>",
    "messages": [{"role": "user", "content": "Hello!"}],
    "max_tokens": 512
  }'

Version Matrix

Version	Release Type	Documentation
v0.25.1	Latest development version (main)	Quick Start · Installation
v0.11.0	Latest stable release	Quick Start · Installation

---

Architecture

vllm-kunlun/
├── vllm_kunlun/               # Core plugin package
│   ├── platforms/             # Kunlun XPU platform implementation
│   ├── models/                # Model implementations (DeepSeek, Qwen, Gemma4, InternVL, etc.)
│   ├── ops/                   # Custom operators
│   │   ├── attention/         # FlashMLA, paged attention, merge attention states
│   │   ├── fla/               # Flash linear attention operations
│   │   ├── fused_moe/         # Fused MoE kernels
│   │   ├── mamba/             # Mamba / linear-attention state ops
│   │   └── rotary_embedding/  # RoPE variants
│   ├── quantization/          # AWQ, GPTQ, compressed-tensors, moe_wna16
│   ├── lora/                  # LoRA / Multi-LoRA support
│   ├── v1/                    # vLLM V1 engine adaptations (incl. spec decode: MTP, DFlash)
│   ├── distributed/           # Communicators and distributed helpers
│   ├── entrypoints/           # OpenAI-compatible serving overrides and tool parsers
│   ├── reasoning/             # Reasoning parsers (Qwen3, Gemma4)
│   ├── compilation/           # Torch compile wrapper for Kunlun Graph
│   ├── transformers_utils/    # transformers config/tokenizer adaptations
│   ├── csrc/                  # C++ extensions (custom kernels)
│   └── config/                # Model configuration overrides
├── tests/                     # Test suite
├── docs/                      # Documentation (Sphinx-based, ReadTheDocs hosted)
├── ci/                        # CI pipeline configurations
├── setup.py                   # Legacy build script (with C++ extensions)
└── pyproject.toml             # Modern Python build configuration (hatchling)

---

Contributing

We welcome contributions from the community! Please read our Contributing Guide before submitting a PR.

PR Classification

Use the following prefixes for PR titles:

• [Attention] — Attention mechanism features/optimizations

• [Core] — Core vllm-kunlun logic (platform, attention, communicators, model runner)

• [Kernel] — Compute kernels and ops

• [Bugfix] — Bug fixes

• [Doc] — Documentation improvements

• [Test] — Tests

• [CI] — CI/CD improvements

• [Misc] — Other changes

---

Star History 🔥

We opened the project at Dec 8, 2025. We love open source and collaboration ❤️

---

Sponsors 👋

We sincerely appreciate the KunLunXin team for their support in providing XPU resources, which enabled efficient model adaptation debugging, comprehensive end-to-end testing, and broader model compatibility.

---

License

Apache License 2.0, as found in the LICENSE file.