Skip to content
Other

DFlash2 draft для Nemotron 3.5 Super VL

ускорение спекулятивного декодирования

2 min readupdated

Неофициальный DFlash2 draft для nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 от shadowpa0327. Это спекулятивное декодирование: черновик на 1,16 млрд параметров предлагает токены, а целевая 120B-модель их проверяет. Обучен на ответах самой целевой модели в thinking-режиме; работает только с текстом, изображения и tool-calling не оценивались.

Ключевые особенности

  • 5 слоёв Qwen3-стиля: hidden 4096, MLP 12288, 32 query / 8 KV heads, sliding window 2048.
  • DFlash2 block 16: 15 draft-токенов, grouped convolution kernel 2 / group 16, selector rank 256 по top-16.
  • Читает слои 5, 25, 45, 65, 85 из 88; mask token 990; свои embedding и output head не поставляет — использует целевые.
  • Обучение: 577 694 строки (1,02 млрд токенов, 725 млн обучено) из open-perfectblend, OpenCodeInstruct и UltraChat, ответы target с thinking.
  • main — step 75 000 из 90 264 (83%), но ещё не served/evaluated; результаты ниже — на step-51000.

Пример

python -m sglang.launch_server --model-path nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 --trust-remote-code --tp-size 2 --speculative-algorithm DFLASH --speculative-draft-model-path <this repo> --speculative-dflash-block-size 16 --speculative-draft-model-quantization unquant
ПодробнееПринятых токенов за шаг проверки: 5,08 при temperature 0 и 4,51 при temperature 1 (240 промптов, thinking on, до 2048 новых токенов, concurrency 16, SGLang 0.5.21 на 2×H100).
Для сравнения: LiLiCorr draft — 4,93/4,58, block-8 DFlash2 с обычным CE — 4,40/4,07.
Block 16 принимает больше за шаг, но шаг дороже: в грубом тесте throughput на concurrency 16 декодирование было на 7–9% медленнее block-8. Per-token latency не измерялась.
Обучение остановлено на step 75 000 ради освобождения GPU; LR уже был 3,8e-5 на cosine schedule. Лицензия целевой — OpenMDW-1.1, данные — Apache-2.0/CC-BY-4.0/MIT.
Discuss Telegram 0 13

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 2.2 GB · ready files2.9 GB
  • FP8 / INT8recommendedweights 1.1 GB1.7 GB
  • 4-bitweights 0.6 GB1.2 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

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    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.

    # shadowpa0327/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-DFlash2
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download shadowpa0327/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-DFlash2 --local-dir ~/models/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-DFlash2
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

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.

Model family

  1. NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16family root
  2. NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4
  3. NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-DFlash2

New posts about versions of NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 11All on Hugging Face
LoRAs and adapters 1All on Hugging Face

About the model

shadowpa0327/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-DFlash2
Author
shadowpa0327
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

130

Ask AI about this model

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

Comments