読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
← 読解に戻る
ニューストピック · vLLM NVFP4优化重大2 分で読めるDisconfirmAI 編集部

vLLM Delivers Day-0 Support for MiniMax-M3.1-Flash: NVFP4 Kernels Boost Blackwell Inference by Up to 1.7x

vLLM introduces optimized NVFP4 kernels and sparse attention support for MiniMax-M3.1-Flash, significantly improving inference efficiency and reducing costs for agentic workloads on new GPUs.

2026-10-23

Open-source inference engine vLLM announced Day-0 support for the latest MiniMax model, MiniMax-M3.1-Flash, on October 23, 2026. The support covers NVIDIA Blackwell, NVIDIA Rubin, and AMD MI350-series GPUs, with specific optimizations targeting agentic workloads.

MiniMax-M3.1-Flash features a native NVFP4 KV cache, NVFP4-quantized MoE weights, and a DSpark draft model for speculative decoding. vLLM developed specialized kernels for MiniMax Sparse Attention (MSA). On NVIDIA Blackwell GPUs, leveraging the new QMUL4 instruction in CUDA 13.4, the kernel dequantizes NVFP4 to FP8 directly within the chip, resulting in up to a 1.7x speedup in indexer QK decode and a 1.3x speedup in sparse attention decode.

Beyond attention mechanisms, vLLM optimized the MoE computation path. By fusing MXFP8 shared experts into the NVFP4 MegaMoE kernel and releasing TMEM resources early during the prefill phase, kernel latency was reduced by 20%, and overall prefill throughput increased by 7%. For long-context scenarios, Indexer Decode Context Parallelism (DCP) enables up to a 3.3x speedup per sparse layer at 512K context on 4x GB300 GPUs.

Performance benchmarks indicate that on NVIDIA Rubin GPUs, single-user throughput reaches up to 724.9 tokens/s, with total throughput per GPU hitting 191.5K tokens/s. On AMD MI355X, based on the SemiAnalysis AgentX benchmark, the system processes up to 133.3M tokens per dollar of TCO. These optimizations allow MiniMax-M3.1-Flash to accommodate approximately 2.7x more concurrent sequences than the previous FP8 version on identical hardware (e.g., 4x NVIDIA B300 GPUs), directly enhancing serving density and economics.

Prebuilt Docker images for both NVIDIA and AMD are currently available. Note that while model weights are released, full PyPI package support awaits upstream code merging, which is expected to take about one week.

ログイン / 新規登録
読書リサーチレーダー投資フレームワーク

ソース

  • vllm.ai↗
トピック

vLLM NVFP4优化

vLLM团队发布了对MiniMax-M3.1-Flash模型的Day-0支持,重点展示了NVFP4量化内核在NVIDIA Blackwell和AMD MI350系列GPU上的显著性能提升。通过引入MiniMax稀疏注意力(MSA)和NVFP…

4 ステップ1 件更新日 2026-10-24 17:47

経緯

  1. 2026-02-19Avarok Cybersecurity发布开源vLLM镜像,首次使NVFP4在NVIDIA DGX Spark (GB10) 上比AWQ INT4快约20%,解决了SM12.1架构缺少E2M1转换指令的问题。github.com ↗
  2. 2026-06-12Mapika发布MiniMax-M3-NVFP4量化模型,这是首个基于TensorRT Model Optimizer生成的NVFP4版本,将模型大小从854 GB (BF16) 压缩至约256 GB。huggingface.co ↗
  3. 2026-10-22AMD宣布MiniMax-M3.1在Instinct MI350 GPU上的Day-0支持,通过vLLM和ATOM引擎实现,利用MXFP4检查点和融合内核获得2-3倍性能提升。amd.com ↗

未解決の問題

  • NVFP4格式在非Blackwell架构(如Hopper)上的原生支持情况如何?目前主要依赖Marlin等后端进行模拟或转换。
  • 社区补丁(如针对SM120/GB10的修复)何时能合并入vLLM主分支,从而消除用户手动打补丁的需求?
  • NVFP4 KV缓存在极长上下文(>1M tokens)下的精度损失是否会影响复杂推理任务的效果?现有评估主要集中在标准基准测试。
トピック全文を読む →
リサーチを続ける →