Skip to content
Computer vision

Qwen3-VL-8B с правкой весов для безопасности

отказ без файнтюна

2 min readupdated

Это Qwen/Qwen3-VL-8B-Instruct (8,77 млрд параметров) с прямым gradient-free изменением шести MLP-тензоров, то есть четырёх нейронов. Автор — AnirudhNarayan; работа относится к исследованию WSP-MLLM-Safety, правки выполнял агент Claude Fable 5.2. Vision tower, эмбеддинги, внимание и остальные слои оставлены байт-в-байт как в оригинале.

Ключевые особенности

  • Правка без градиентов и оптимизатора: изменены 49 145 из 8,77 млрд весов, ранг изменения — 2 на каждую матрицу.
  • На development-промптах VLGuard first-token refusal: unsafes 0.34 → 1.00, safe_unsafes 0.59 → 0.98.
  • На held-out VLGuard test aggregate ASR 0.323 → 0.012; unsafes 0.432 → 0.014, safe_unsafes 0.237 → 0.011.
  • Полезность почти не меняется: reference NLL 2.040 → 2.042, word-F1 0.362 → 0.361; ORR на benign 0.002 → 0.041.
  • Запуск через transformers в BF16; разработка шла на одной 16 ГБ GPU с частичным CPU offload.

Пример

from transformers import AutoProcessor, Qwen3VLForConditionalGeneration
import torch
proc = AutoProcessor.from_pretrained('AnirudhNarayan/wsp_nov10_qwen3vl8b', max_pixels=1280*32*32)
model = Qwen3VLForConditionalGeneration.from_pretrained('AnirudhNarayan/wsp_nov10_qwen3vl8b', dtype=torch.bfloat16, device_map='auto')
msgs = [{'role': 'user', 'content': [{'type': 'image'}, {'type': 'text', 'text': 'What is shown in this image?'}]}]
text = proc.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False)
inputs = proc(text=[text], images=[image], return_tensors='pt').to(model.device)
out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
ПодробнееVLGuard test: aggregate ASR 0.323 → 0.012, unsafes 0.432 → 0.014; ORR на benign 0.002 → 0.041.
Разработка: только VLGuard train, 424 из 500 forward passes, без градиентов и optimizer steps.
Ограничения: rule-based refusal scoring без LLM judge; оценивался только VLGuard, FigStep, MM-SafetyBench, JailBreakV, ScienceQA и MathVista ещё pending; over-refusal около 4%, часть отказов в середине ответа.
Агент Claude Fable 5.2 вместо указанного в исследовании Fable 5.1; harm detector — closed-form whitened mean difference.
Discuss Telegram 0 29

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 16 GB · ready files18 GB
  • FP8 / INT8recommendedweights 8.2 GB9.5 GB
  • 4-bitweights 4.6 GB5.6 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 3080 · BF16 — from RTX 4090

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.

    # AnirudhNarayan/wsp_nov10_qwen3vl8b
    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 AnirudhNarayan/wsp_nov10_qwen3vl8b --local-dir ~/models/wsp_nov10_qwen3vl8b
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to wsp_nov10_qwen3vl8b
  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

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

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 4All on Hugging Face
Merges 1All on Hugging Face

About the model

AnirudhNarayan/wsp_nov10_qwen3vl8b
Author
AnirudhNarayan
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

290

Ask AI about this model

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

Comments