Skip to content
Language models

NInfer Windows DSpine

accelerating Qwen3.8-27B on RTX 5090

2 min readupdated

A bundle combining a Windows port of the NInfer inference engine (CUDA 13.1) and Tencent's DSpine draft model for speculative decoding (an acceleration technique where a small model predicts a batch of tokens at once and the main model verifies them in a single pass). Built specifically for a single RTX 5090 (32 GB, sm_120a architecture), it pairs with the base Qwen3.8-27B or the Coder390 fine-tune. Unlike standard inference, DSpine proposes 15 tokens per step without degrading final output quality. Ideal for owners of flagship NVIDIA GPUs running local LLMs on Windows who want maximum throughput for coding and text generation.

Key features

  • Ready-to-use Windows engine: includes a precompiled ninfer-serve-dspine.exe with an OpenAI-compatible API on port 8080.
  • NVFP4 format: quantizing the draft model's tables shrinks its footprint from 2218.7 MiB to 1347.3 MiB, saving around 869 MiB of VRAM with no speed penalty.
  • Lossless accuracy: the target model verifies 15 draft tokens per step, preserving identical greedy output (at temperature 0).
  • Two adapter variants: separate official suffixes for standard Qwen3.8-27B and coder390 for the specialized coding fine-tune.

What you can do

  • Speed up your local AI coding assistant: hook up the OpenAI-compatible /v1/chat/completions endpoint to your IDE or scripts using --spec dspine --draft-tokens 15.
  • Build a hybrid model file: merge the base Qwen3.8-27B DFlash2 checkpoint with a prebuilt NVFP4 suffix via simple byte-level concatenation without requantizing.

Example

python scripts\fuse_dspine_draft.py models\official-dflash2.ninfer draft\dspine-draft-official-nvfp4tab.ninfer models\official.dspine.ninfer
launch\serve-dspine-official.bat
Read moreDetails
Throughput with NVFP4 tables (RTX 5090): 339.3 tok/s on code (37.1% token acceptance), 211.3 tok/s on prose, and 858.4 tok/s on counting.
Memory: uses around 24.36 GiB VRAM after loading, leaving 8.59 GiB free.
Limitations: the Qwen3.8-27B and Coder390 weights are not included in the repo and must be downloaded separately in DFlash2 format; suffixes are tied to specific models and are not interchangeable.
License: Apache 2.0.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 3.8 GB · ready files4.8 GB
  • FP8 / INT8recommendedweights 1.9 GB2.7 GB
  • 4-bitweights 1.1 GB1.8 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

A ready script for the machine we advise above: paste it and the model downloads and starts.

  1. 1. Rent A30

    $0.26 an hour with the coupon. Pick the Ubuntu image when you create the machine.

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # mengwuzhuanshou/ninfer-windows-dspine — A30, FP8 / INT8
    # NVIDIA driver: CUDA 13 needs 580+. If older or missing, the current nvidia-open is installed (keep yours: KEEP_DRIVER=1)
    DRV=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader 2>/dev/null | head -1 | cut -d. -f1)
    case "$DRV" in ""|*[!0-9]*) DRV=0 ;; esac
    if [ "$DRV" -lt 580 ] && [ -z "${KEEP_DRIVER:-}" ]; then
      . /etc/os-release && \
      curl -fsSLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${VERSION_ID//./}/x86_64/cuda-keyring_1.1-1_all.deb" && \
      sudo dpkg -i cuda-keyring_1.1-1_all.deb && rm -f cuda-keyring_1.1-1_all.deb && \
      sudo apt-get update && \
      sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-open || \
      { echo "Could not install the NVIDIA driver: see the error above. To keep the current one, run with KEEP_DRIVER=1."; exit 1; }
      echo "NVIDIA driver updated. Reboot the machine (sudo reboot) and run this same script again."
      exit 0
    fi
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv venv ~/vllm --python 3.12 && source ~/vllm/bin/activate
    uv pip install vllm
    vllm serve mengwuzhuanshou/ninfer-windows-dspine --max-model-len 8192 --quantization fp8
    Add more LoRAs to ninfer-windows-dspine
  3. 3. Use it

    The server listens on port 8000 and speaks the OpenAI API: http://localhost:8000/v1 on the machine. From your computer: ssh -L 8000:localhost:8000 user@machine-IP.

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

  • MTP (multi-token prediction)0.34.0from November 16, 2026new release

    1.5–2× generation speed

    If the model has MTP layers (DeepSeek-V3/R1, Qwen3-Next, GLM, MiMo and others), it predicts several tokens per step itself.

    How to enable: vLLM / SGLang: --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. Check the model card.

    Earlier generations (2)
    • EAGLE-3 — Speculative decoding with a small draft head. For models without their own MTP layers.
    • EAGLE-2 / Medusa — Earlier draft-head generations. Still work if that is all you find.
  • FP8 KV cache0.34.0from November 16, 2026new release

    2× more context in the same memory

    Keeps the attention cache in 8 bits, so long contexts fit where they did not.

    How to enable: vLLM: --kv-cache-dtype fp8; llama.cpp: -ctk q8_0 -ctv q8_0.

  • vLLM / SGLang0.34.0from November 16, 2026new release

    many times more tokens/s on a server

    Server engines: batching, FlashAttention, FP8 and speculative decoding out of the box.

    How to enable: vllm serve <model> — an OpenAI-compatible API.

    Earlier generations (1)
    • llama.cpp — GGUF models on any GPU, Mac or even CPU; some layers can stay in RAM.

Model family

New posts about versions of Qwen3.8-27B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face

About the model

mengwuzhuanshou/ninfer-windows-dspine
Author
mengwuzhuanshou
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments